Subscribe now

Prologue: Afternoon Tea on the Day of the Spring Equinox

March 20, 2026, Beijing, the Vernal Equinox.

On this day, day and night were nearly equal in length. Sunlight slanted down through the shadows of trees lining both sides of Chang’an Street, falling gently onto Tim Cook’s silver hair. As he stepped out of the black MPV and entered the conference room of the China Council for the Promotion of International Trade, dressed in a dark-blue suit, he looked like a traveler who had already journeyed far. Walking slowly, he spoke at a measured pace: “It’s very important,” he replied calmly, addressing the question about “the significance of the Chinese market to Apple.”

This is his last official visit to China as Apple’s CEO.

Today, September 1, he hands over Apple’s top executive position to John Ternus, Senior Vice President of Hardware Engineering, and takes on the role of Executive Chairman of the Board himself. An era is about to turn a new page. And Cook has chosen to reserve his most important of his final public appearances for China.

That day back in March, he met with Ren Hongbin, Chairman of the China Council for the Promotion of International Trade, over a traditional Chinese afternoon tea. There were no tense negotiations, no high-profile declarations under the glare of cameras—rather, it felt more like a conversation among old friends. Cook said that the importance of the Chinese market to Apple goes without saying. He also quoted an ancient Chinese proverb: “Many hands make light work.”

This statement sounds mild, even somewhat simple, yet it almost encapsulates his three-decade-long relationship with China. Cook is not one to excel at creating dramatic moments. Unlike Jobs, who often ignited the entire room with a single phrase—“one more thing”—Cook rarely resorts to grand rhetoric to define the business world. His language is more like a balance sheet: precise, calm, and quantifiable.

Yet it is precisely within this restraint that he has accomplished something profoundly complex: over the past three decades, he has tightly woven an American tech company into the Chinese market, Chinese manufacturing, Chinese developers, the Chinese supply chain, and the transformations occurring in Chinese society. From his first visit to China representing Apple in the late 1990s to his final trip to the country as CEO during the spring equinox of 2026, Cook has visited China roughly twenty times. He’s seen the lights burning late into the night at factories in Shenzhen, and he’s also witnessed the applause at press conferences in Beijing. He’s discussed tariffs, supply chains, green manufacturing, as well as Chinese developers, Chinese universities, and Chinese consumers.

He’s not a diplomat, yet he has long played the role of a bridge in a commercial sense. He’s not a historian, yet he has personally witnessed China’s long journey of transformation—from “world’s factory” to “innovation hub.” He’s not a poet, yet with the patience of an American supply-chain expert, he has forged his own path between the two sides of the Pacific. To understand Cook’s relationship with China, we can’t just look at how many iPhones Apple has sold in China, nor can we simply tally up the revenue that the Greater China region has contributed to Apple. Those are merely the surface-level aspects of the story. Beneath the surface lies three decades of industrial transformation, technological flows, trust-building, risk balancing—and the most complex and authentic business narrative of our globalized era.

Chapter One: The Era Before the iPhone

In 1996, Tim Cook, the procurement manager at Compaq Computers, came to China.

That year, it was still eleven years before the debut of the original iPhone, more than twenty years before Apple became a trillion-dollar company, and fifteen years before Tim Cook succeeded Steve Jobs as Apple’s CEO. Back then, Apple was far from being the tech giant with an enormous ecosystem that it is today. It was still struggling hard in the personal-computer market, with a complex product lineup, chaotic inventory, and low operational efficiency.

Cook joined Apple two years later.

He previously held a brief position at Compaq Computers, and even earlier, he spent twelve years at IBM. Unlike Jobs, whose leadership was renowned for his product intuition and ability to bend reality to his will, Cook’s temperament is more akin to that of a factory floor: He cares about inventory turnover, component costs, whether suppliers can deliver on time, and whether a production line can accomplish more tasks in less time.

Jobs had once publicly mocked IBM’s bureaucratic culture, but Cook learned from that large-company operating system how to turn chaos into order. Later, one of Apple’s most important strengths became precisely its ability to make complex manufacturing systems predictable, controllable, and scalable. And China is precisely the biggest stage for him to showcase this ability. In the late 1990s, China was at a critical juncture as its manufacturing sector accelerated its integration into the global division of labor. In factories across the Pearl River Delta and the Yangtze River Delta, machines roared, assembly lines operated around the clock, and countless workers migrated from rural areas to cities, embarking on an industrialization process of unprecedented scale. For many multinational technology companies, China represented cost advantages, production capacity flexibility, and—above all—a manufacturing potential that had yet to be fully tapped.

What Cook saw might have been more than just these.

What he saw was a possibility: What would happen if Apple’s stringent requirements for quality, efficiency, and precision manufacturing were combined with China’s enormous manufacturing capacity?

This is not a romantic fantasy—it’s the judgment of supply-chain experts. In Cook’s logic, manufacturing is never simply “contract manufacturing.” Whether a product can be mass-produced, whether its process can be upgraded in an extremely short timeframe, and whether it can strike the right balance among quality, cost, and delivery—all depend on the maturity of the supplier ecosystem, the sufficiency of engineers, the completeness of the industrial chain, the efficiency of logistics, and even the entire society’s ability to respond rapidly to changes.

China happens to possess these elements.

Thus, even in the days before the iPhone was born, Cook had already begun laying the groundwork for a deeper connection between Apple and Chinese manufacturing. At that time, Apple’s primary products were still computers, and China had not yet become the global hub for consumer electronics manufacturing. Yet the seeds had already been sown. Years later, recalling this experience at the China Development Forum, Cook said that in the late 1990s, when he first represented Apple in China, the company didn’t yet have the iPhone. He added that Apple had grown alongside China.

This sentence is worth pondering again and again. “Growing together” is not merely a polite cliché. It means that Apple doesn’t just come to China to procure, manufacture, and sell; nor does China simply provide Apple with labor and factory space. Rather, there exists between the two a long-term, two-way, mutually shaping relationship. Apple has brought the world’s most stringent consumer electronics manufacturing standards into China’s supply chain; in turn, China—with its astonishing learning speed, organizational prowess, and industrial scale—has driven Apple to make its products lighter, thinner, more precise, and more sophisticated.

This is a symbiotic relationship. However, in 1996, almost no one could have foreseen just how large this symbiotic relationship would grow over the next three decades.

Chapter 2: The Distance of the Jobs Era

If Cook was Apple’s most patient connector with China, then Apple during the Jobs era had a far more complex attitude toward the Chinese market.

Of course, Jobs was well aware of China’s importance. China boasts a massive consumer base, a fully integrated manufacturing system, and a rapidly emerging internet ecosystem. Yet for a long time, Apple in China was more like a company that existed largely in a “passive” state.

In 2007, the first-generation iPhone was released.

That was a moment that transformed the mobile phone industry. The touchscreen replaced the keyboard, and mobile internet began to shift from computer screens to the palm of your hand. The iPhone wasn’t just a product—it marked the dawn of a new computing platform.

However, while American consumers lined up to buy the iPhone, Chinese consumers mostly had no choice but to obtain it through gray-market products, jailbreaking, or unofficial channels. At the time, Apple had not yet established formal partnerships with China’s mainstream telecom operators. Although the iPhone held immense appeal in China, there was still no stable, legal, and large-scale pathway for its entry into the Chinese market. Apple under Jobs was accustomed to letting its products do the talking. In Silicon Valley’s logic, as long as the product is good enough, the market will naturally open up. Consumers will wait, channels will adapt, and operators will eventually give in. This confidence is not unfounded. The original iPhone, iPod, and Mac all won loyal users precisely because of their exceptional product quality.

However, the Chinese market does not entirely follow this logic.

China’s telecommunications market has its own regulatory framework, operator structure, and consumer habits. Selling mobile phones is not merely a product issue—it also involves network standards, tariff plans, channel management, local services, and after-sales support systems. If a foreign brand wants to truly enter the mainstream market, it must learn to coexist harmoniously with the local ecosystem. What Cook changed later was precisely this point.

He doesn’t deny the importance of products. On the contrary, he understands better than anyone that all of Apple’s commercial expansion must be built on product strength. Without the iPhone, none of what came after would have been possible. Yet Cook believes that the product is merely the starting point. Supply chains, distribution channels, partners, government relations, and localized operations—these seemingly less “sexy” aspects are precisely what determine whether a revolutionary product can truly turn into a global commercial triumph.

Jobs was like a conductor standing center stage, igniting the world with his products; Cook, by contrast, was more like the chief engineer behind the scenes, ensuring that the stage wouldn’t collapse, the lights wouldn’t go out, and the musical score could be executed precisely through the coordinated efforts of tens of thousands of people.

These two temperaments are not contradictory.

Apple needs Jobs’ vision, but it also needs Cook’s execution. In the Chinese market, Cook’s kind of patience is particularly important.

Negotiations with China Mobile began around 2007 and continued until 2013. Six years is more than enough time for an entire industry to undergo several major shifts. Although Jobs didn’t live to see this collaboration completed, Cook took over and kept pushing forward. Finally, at the end of 2013, Apple signed an iPhone sales agreement with China Mobile. In January 2014, Cook himself traveled to Beijing and stood alongside Xie Guohua, Chairman of China Mobile, officially launching the 4G versions of the iPhone 5s and 5c into China’s mainstream operator ecosystem. This is not a simple channel partnership.

It means Apple has finally admitted: In China, having a great product alone isn’t enough. You must enter the local ecosystem, understand local channels, and build long-term relationships with local partners. Over six years, Cook transformed the iPhone from a product that was “coveted but difficult to obtain officially” into a commodity that Chinese consumers can easily purchase through legitimate channels. Since then, Apple’s growth in China has rapidly accelerated. From 2010 to 2012, Apple’s revenue in China surged by approximately 600%, and its share of global total revenue jumped from about 2% to 12%. By the first quarter of fiscal year 2026—the quarter ending December 2025—Apple’s Greater China region revenue reached $25.53 billion, representing a 38% increase year-on-year and accounting for roughly 17.7% of its global total revenue.

In other words, for every $100 that Apple earns, nearly $18 comes from China. Behind this number lie countless Chinese consumers stepping into Apple retail stores, countless developers creating apps around the iOS ecosystem, and countless factories, engineers, and logistics personnel involved in manufacturing and delivery.

Cook didn’t view China merely as a sales market.

He regards it as the most important link in Apple’s global system.

Chapter 3: Symbiosis in the Supply Chain

In November 2024, Cook appeared at the 2nd China International Supply Chain Promotion Expo.

The Chain Expo is no ordinary consumer electronics show. Its focus isn’t on any single new product, but rather on the industrial chain itself: raw materials, components, manufacturing equipment, logistics, energy, R&D, and green transformation. For supply-chain managers like Cook, an event like this feels almost tailor-made for them.

He said there, “Without these Chinese partners, we wouldn’t have been able to achieve today’s accomplishments.” This statement is not an exaggeration.

Although Apple’s products seem to be born in design studios in California, they actually require extremely complex collaboration across a global supply chain. Every component—be it a piece of glass, a screw, a camera module, a battery, or a motherboard—depends on the coordinated efforts of dozens, even hundreds, of suppliers. And China occupies a special position within this network. As of 2026, more than 80% of Apple’s top 200 global suppliers have production facilities in China. Approximately 80% of iPhones are still finalized and assembled in China. Even though Apple has expanded its production capacity in recent years to places like India and Vietnam, China remains the most critical, most mature, and most irreplaceable component of its manufacturing ecosystem. This dependence is no accident.

China’s supply-chain advantages go beyond just labor costs. More crucially, it boasts a fully integrated industrial ecosystem, a massive pool of engineers, rapid iterative capabilities, and the ability to swiftly transform design blueprints into large-scale manufacturing. The precision, yield rates, environmental standards, and delivery schedules that Apple demands from its suppliers are exceptionally stringent on a global scale. Suppliers capable of consistently meeting these requirements often undergo self-improvemen in the process as well.

Lens Technology (蓝思科技) is a typical example.

In its early days, this company supplied glass cover panels to Apple, participating in the supply chain since the time of the original iPhone. As Apple’s products continued to evolve and upgrade, Lens Technology gradually expanded its capabilities—from simple glass processing to more complex protective panels, structural components, and precision manufacturing. By the era of the Apple Vision Pro, it had become an integral part of Apple’s cutting-edge product supply chain. Luxshare Precision(立讯精密)’s growth trajectory is equally intriguing.

Starting from relatively basic components such as connectors and cables, the company gradually moved into Apple’s more core manufacturing processes and eventually became one of the key manufacturers behind nearly all of Apple’s most advanced products. Its growth has not been merely driven by an increase in orders; rather, it reflects a comprehensive enhancement of its manufacturing, management, and engineering capabilities.

BYD Electronics (比亚迪电子) is the same way.

The outside world is more familiar with BYD because of its involvement in the new-energy vehicle sector. However, in the consumer electronics manufacturing field, BYD Electronics is also a key player in Apple’s supply chain. Starting from metal casing processing, it has gradually expanded into an even wider range of product lines and manufacturing stages, becoming an indispensable part of Apple’s manufacturing ecosystem.

These companies are not contract manufacturers passively fulfilling Apple’s orders.

In the course of collaborating with Apple, they have learned about the global top brands’ requirements for quality, efficiency, environmental sustainability, and innovation. At the same time, they have also leveraged China’s manufacturing strengths—such as scale, speed, and organizational capabilities—to benefit Apple in return. Cook is well aware of this.

In March 2025, during a media interview in Shanghai, he said that for Apple’s supply chain, there is no place more important than China. He also noted that over the past three decades, Apple has been steadily expanding its supply chain in China and continuously increasing its investment there. Even more noteworthy is the sentence he said afterward.

He said that if we look back at today in ten years, we might not think our current level of modernization is all that high, because everyone will keep making progress. This sentence feels like a prophecy.

It means that Cook doesn’t view China’s supply chain as a static “low-cost manufacturing base,” but rather as a continuously evolving ecosystem. Apple doesn’t simply draw production capacity from China, and Chinese suppliers don’t merely serve as contract manufacturers for Apple. Instead, the two sides are engaged in a long-term, mutually shaping relationship. Apple sets higher standards, forcing suppliers to upgrade; as suppliers enhance their capabilities, Apple can then design products that are more complex, more precise, and more advanced. This cycle makes Apple’s products increasingly difficult to simply replicate, and it also deepens China’s position in the global consumer electronics manufacturing supply chain. ———

Chapter 4: From Selling Products to Taking Root in China

The story of Cook and China would still be incomplete if it were limited solely to “selling products” and “making products.”

What truly makes this relationship long-term and stable is Apple’s gradual integration of R&D, the developer ecosystem, educational partnerships, green manufacturing, and local innovation into China’s narrative. In October 2025, Apple’s Chief Operating Officer Sabih Khan revealed that Apple has invested over one billion RMB cumulatively in its application research labs located in Shenzhen and Shanghai. These labs are not merely symbolic display spaces; rather, they are genuine technological facilities dedicated to product improvement and process development. They are equipped with cutting-edge instruments such as industrial-grade CT scanning devices, which enable three-dimensional imaging of the internal structure of iPhones without disassembling the phones. Using this technology, engineers can examine component assembly, thermal management structures, material stresses, and potential defects, thereby refining both design and manufacturing processes.

The enhanced durability of the Ultra Ceramic Glass panel, the structural design of the iPhone Air, and the VC vapor chamber cooling solution used in the iPhone 17 Pro series—all are related to the work conducted in these laboratories.

This means that in China, Apple is no longer just “producing products”; it’s also beginning to “improve products.”

The establishment of the R&D center has shifted the direction of knowledge flow between Apple and China. In the past, it was customary to think that Apple’s innovation took place in California, while China was responsible for implementation. Today, this division of labor is becoming increasingly complex. Chinese engineers, materials experts, and manufacturing specialists are increasingly involved in Apple’s early-stage product development and process optimization.

This is a deeper level of trust.

Because R&D collaboration involves sharing technology roadmaps, process secrets, and future product directions, it is more sensitive than a simple purchase order and also longer-term than ordinary manufacturing cooperation. Cook clearly recognizes that if Apple wants to achieve long-term success in China, it can’t just remain an outsider brand. It must demonstrate that it can bring value to China’s innovation ecosystem. In March 2025, Cook announced in Hangzhou that Apple would donate 30 million RMB to Zhejiang University to establish the “Apple Mobile App Incubation Fund.” Combined with its previous decade-long investment in co-hosting the Mobile App Innovation Competition with Zhejiang University, Apple’s cumulative donations to Zhejiang University have now reached 80 million RMB, benefiting over 30,000 students from nearly a thousand universities.

This is not an isolated charitable donation.

It points to the developer ecosystem.

Apple’s business model has long gone beyond just hardware sales. The App Store, the iOS ecosystem, developer tools, in-app purchases, and service revenues have become key components of Apple’s profit structure. And developers are the capillaries of this ecosystem. Without developers, iOS would merely be an operating system; with developers, it becomes the apps, games, tools, and services that countless people use every day.

By investing in developers in China, Cook is essentially investing in the future ecosystem foundation of Apple.

Meanwhile, green manufacturing has also become another important thread in Apple’s relationship with China.

In October 2025, Apple announced that more than 90% of its manufacturing in China has already been powered by renewable energy. Apple has also joined forces with Chinese suppliers to launch the “China Renewable Energy Infrastructure Fund,” with a total scale of 1 billion RMB. The fund aims to add 1 million megawatt-hours of clean electricity to China’s power grid by 2030. The significance of this set of numbers goes far beyond the polished language found in corporate social responsibility reports. The energy transition in manufacturing is no easy feat. Factories need to upgrade their equipment, supply chains must restructure their energy mix, and coordination is required among local governments, power grids, enterprises, and environmental protection goals. The fact that Apple has been able to drive a green transformation of this scale in China demonstrates that it has become deeply embedded in the operational logic of China’s industrial system.

Cook doesn’t treat environmental protection merely as a part of the brand image.

He turned it into part of a supply chain upgrade.

This point is crucial. In the future global trade and technological competition, carbon emissions, energy mix, and supply-chain transparency are emerging as new competitive thresholds. By proactively establishing green manufacturing capabilities in China, Apple is not only fulfilling its corporate commitments but also preparing itself for the industrial rules that will shape the next decade.

Chapter 5: Walking Between Two Worlds

The most delicate—and also the most wisdom-testing—aspect of Cook’s relationship with China lies beyond the realm of commerce.

Over the past decade, technological competition between China and the U.S. has steadily intensified. Issues such as chips, artificial intelligence, data security, supply-chain resilience, export controls, and market access have increasingly become part of the daily decision-making processes of global tech companies. Apple finds itself squarely in the middle of this dynamic and cannot afford to sidestep it. As a U.S. company, it must comply with U.S. laws and policy environments; yet it is also a company that heavily relies on Chinese manufacturing, places great importance on the Chinese market, and serves consumers worldwide. It cannot simply take sides, nor can it pretend that geopolitics doesn’t exist.

Cook’s approach carries a distinct personal style. He rarely issues fiery political statements and seldom uses emotional language to respond to complex issues. Instead, he prefers to explain Apple’s position using business-oriented language, collaborative narratives, and a focus on long-term interests.

In Washington, he emphasized Apple’s contributions to the U.S. economy: investment, job creation, R&D, the return of supply chains, and collaboration with U.S. suppliers. In August 2025, Apple announced that it would invest $600 billion in the U.S. over the next four years. This figure has been widely interpreted as part of Apple’s effort to demonstrate its value to the U.S. policy environment.

In Beijing, he emphasized the importance of China’s supply chain, the irreplaceable nature of the Chinese market, and Apple’s experience of growing together with its Chinese partners. He said that Apple and China share a relationship of “growing together,” remarked that China’s supply chain is “the most critical place there could be,” and added, “Many hands make light work.”

Both sides heard what they wanted to hear. This balance is not easy to maintain.

It demands that Cook neither over-compromise to any single side nor allow Apple to lose trust in key markets. His language is always cautious, restrained, and even somewhat vague. Yet precisely this vagueness sometimes becomes the very space in which multinational corporations can survive in complex political environments. Cook understands that Apple is not a political entity, but it also cannot remain detached from political realities. What he can do is to identify the point of least friction among business logic, technological realities, policy environments, and public sentiment.

He ensures that Apple remains as “indispensable” as possible: for the United States, it’s a symbol of innovation, job creation, and global competitiveness; for China, it’s a key player in high-end manufacturing, green transformation, developer ecosystems, and consumption upgrading; and for consumers worldwide, it’s a provider of products and services.

This “indispensability” is Cook’s most important moat.

But moats also have boundaries.

The technological competition between China and the U.S. is still ongoing, and the pressure to restructure supply chains has yet to dissipate. Geopolitical risks are unlikely to disappear simply because of a single visit or a mildly worded statement. What Cook leaves behind for his successor is not a smooth, straight path—but rather a complex web of relationships: one that is both profoundly deep and strikingly fragile; built on long-term trust yet requiring constant upkeep.

Chapter Six: The Glass Box and the Urban Living Room

In July 2008, one month before the opening of the Beijing Olympics, Apple opened its first Apple Store in mainland China, located in Sanlitun, Beijing.

What’s most striking about that building is its massive, curved glass façade. Sunlight streams through the glass, illuminating the white long tables inside the store, reflecting off the iPhones and MacBooks held by customers, and even lighting up the faces of the crowds waiting in line outside.

To Chinese consumers at the time, this store didn’t feel like a typical electronics retailer—it was more like a small art gallery or an entrance to the future. At the time, Cook wasn’t yet Apple’s CEO, but he already had a deep understanding of the significance of retail.

In Jobs’ vision, the Apple Store was never just a place to sell products. It was the physical embodiment of the brand experience—a direct point of contact between consumers and the Apple ecosystem. Cook brought this philosophy to China and infused it with a more pragmatic business logic: retail stores are not merely sales channels; they are the foundational infrastructure for building brand trust.

In the nearly two decades since then, Apple has opened dozens of retail stores across mainland China—from Beijing and Shanghai to Shenzhen, Chengdu, Hangzhou, Nanjing, Wuhan, Xi’an, and Qingdao. Each store’s location has been carefully selected, often in the most central and bustling commercial districts of the city. These stores are not merely counters tucked away in shopping mall corners; rather, they are standalone buildings, urban landmarks, and even popular tourist destinations.

In 2025, Apple opened a new retail store in Shenzhen. The store’s design incorporates Lingnan (岭南) architectural elements, engaging in a dialogue with the texture of the surrounding neighborhood. On opening day, long lines formed outside the store—some people were there to buy the latest products, others to take photos and check in, and still others simply wanted to step inside for a cup of coffee and a moment of relaxation. This kind of scenario would have been unimaginable ten years ago. A retail store of an American tech company has become an integral part of public life in Chinese cities. Here, people take “Today at Apple” classes to learn photography, programming, and music production; they book appointments at the Genius Bar to get their devices repaired; and they meet up with friends to spend weekend afternoons together. It’s no longer just a space for consumption—it’s evolved into something akin to a urban living room.

Cook’s emphasis on retail reflects his consistent management philosophy: user experience is not merely a marketing slogan—it’s a systematic endeavor that can be quantified, managed, and replicated. At each Apple Store, metrics such as sales per square foot, foot traffic, conversion rates, and customer satisfaction are all integrated into a meticulously designed operational framework. This management approach is entirely consistent with his style in the supply-chain domain. But retail stores have another layer of meaning as well.

In an era marked by heightened tensions in U.S.-China relations and frequent trust crises facing multinational corporations, a physical retail store is the most tangible “proof of presence.” It signifies that Apple isn’t just about remote sales, cross-border e-commerce, or a light-asset model that can be easily abandoned at any time. Here, Apple has invested in rent, manpower, interior design, and a local team—it has forged a long-term relationship with this city.

Chapter 7: Developers—The Invisible Capillaries

If retail stores are the bridge between Apple and Chinese consumers, then developers are the capillaries connecting Apple to China’s innovation ecosystem.

Cook recognized early on that Apple’s business model was undergoing a fundamental shift. While hardware remains central, software and services are becoming the primary drivers of profit growth. The App Store, iCloud, Apple Music, Apple Pay, and Apple TV+—these services together form a vast ecosystem, with developers being its most active participants.

In China, the developer community is particularly unique.

China boasts the world’s largest mobile internet user base and an exceptionally vibrant entrepreneurial culture. From WeChat (微信) and Douyin (抖音) to Xiaohongshu (小红书) and Pinduoduo (拼多多), Chinese developers have demonstrated astonishing innovation in the mobile app space. Rather than simply imitating Silicon Valley, they have cultivated unique product forms and business models that are tailored to the local market.

Cook is fully aware of this.

In 2016, Apple established its first developer center in China, located in Shanghai. This center is not an office space but rather a service platform geared toward developers, offering technical consulting, testing equipment, training resources, and community events. Its establishment marks Apple’s recognition of developers as a core strategic asset in China. Since then, Apple has continued to ramp up its efforts to build a robust developer ecosystem in China.

In March 2025, Cook announced in Hangzhou a donation of 30 million yuan to Zhejiang University to establish the “Apple Mobile App Incubation Fund.” This donation is not an isolated event. Over the past decade, Apple has collaborated with Zhejiang University to host the Mobile App Innovation Competition, reaching more than 30,000 students from nearly a thousand universities. Including previous contributions, Apple’s cumulative donations to Zhejiang University have now totaled 80 million yuan.

Behind these figures lies a clear strategic intent: to nurture the next generation of Chinese developers. In his speech in Hangzhou, Cook mentioned a specific detail: He said he had met many Chinese developers, many of whom began writing apps for the iOS platform during their university years. Some later went on to found their own companies, others joined large tech corporations, and still others continued as independent developers, contributing their creative ideas to the App Store.

These people are Apple’s deepest roots in China. They may not necessarily work directly for Apple, but their innovations have enriched the iOS ecosystem. Their apps have attracted more Chinese users to adopt iPhones and iPads, and their success stories, in turn, inspire even more young people to join the ranks of developers. It’s a virtuous cycle—and Cook is the one who keeps injecting energy into this cycle.

In March 2026, during his final visit to China, Cook once again highlighted Chinese developers. He said that Chinese developers are an indispensable part of Apple’s ecosystem, and their creativity is truly remarkable. This statement is not just a polite formality.

On the App Store, Chinese developers have consistently ranked among the world’s top contributors in terms of app downloads and revenue. From games to productivity tools, from education to health apps, from e-commerce to social platforms, Chinese developers are using code to redefine the possibilities of the mobile internet. And Apple, by providing a platform, tools, and distribution channels, has become the infrastructure provider fueling this wave of innovation. Cook’s relationship with Chinese developers is a delicate symbiosis. He needs them just as they need him.

Chapter 8: When AI Meets the East

In 2025, artificial intelligence has become the hottest topic in the global tech industry. From ChatGPT to Sora, from large language models to multimodal generation, AI is reshaping every industry—including Apple. In June 2024, at its Worldwide Developers Conference, Apple unveiled Apple Intelligence—a personal intelligence system deeply integrated into iOS, macOS, and iPadOS.

However, Apple Intelligence’s path to global adoption is not smooth.

In China, AI services face a unique regulatory environment. Requirements such as data localization, content censorship, and algorithm registration mean that Apple cannot simply transplant the U.S. version of Apple Intelligence into China. Instead, it must find a way to comply with Chinese regulations while maintaining the integrity of the product experience. Cook’s solution is collaboration.

In 2025, Apple confirmed that it is collaborating with Chinese AI companies to bring Apple Intelligence to China. The partners include Baidu and Alibaba. Baidu will provide AI technology capabilities and integrate them deeply with Siri, while Alibaba will be responsible for the content review engine, ensuring that AI-generated content complies with China’s regulatory requirements.

This collaborative model is yet another manifestation of Cook’s art of balance.

He didn’t choose to build a completely independent AI system of his own to evade regulation, nor did he opt to abandon the Chinese market. Instead, he took a more complex yet pragmatic approach: partnering with local collaborators, leveraging their technological expertise and compliance experience, while still maintaining Apple’s control over the user experience.

This strategy is not without risk.

Partnering with local AI companies means Apple must, to some extent, open up its system interfaces, share user interaction data, and accept the technological limitations imposed by its partners. For a company renowned for its closed ecosystem, this represents a significant concession. But Cook clearly believes that this concession is worth it.

China is one of Apple’s most important markets and also one of the most dynamic regions for AI innovation worldwide. If Apple Intelligence cannot be launched in China, Apple will lose a critical battleground in this AI race. Losing the Chinese market would not only mean a loss of revenue but also imply being marginalized in the AI era. At the China Development High-Level Forum in March 2026, Cook spoke about AI. He said that AI is a transformative technology, and Apple is committed to developing and deploying AI in a responsible manner. He also specifically highlighted the progress being made in China’s AI sector, praising the achievements of Chinese AI companies such as DeepSeek.

These remarks have attracted widespread attention. The CEO of a U.S. tech company publicly praised the technological capabilities of Chinese AI firms—a move that takes on particularly significant meaning against the backdrop of the current U.S.-China tech rivalry. Cook isn’t expressing a political stance; he’s sending a business signal: Apple respects China’s technological capabilities, is willing to collaborate with Chinese companies, and believes in the potential of China’s AI ecosystem.

This posture is the culmination of Cook’s thirty years of experience in China.

He knows that when doing business in China, you can’t just bring products—you must also bring respect. Respect for the local market, respect for local partners, and respect for local innovation. This kind of respect isn’t weakness; rather, it’s the wisdom of a long-term perspective.

Chapter 9: Walking on a Tightrope

The most challenging aspect of Cook’s relationship with China lies beyond the commercial sphere.

Over the past decade, technological competition between China and the U.S. has shifted from simmering undercurrents to open confrontation. Chip export controls, the Entity List, data security reviews, and pressure to decouple supply chains—these issues have continuously disrupted the operating environment for multinational tech companies. Apple finds itself at the heart of this storm and cannot afford to turn a blind eye. It is a U.S. company headquartered in California and governed by U.S. law.

It is also one of the world’s leading consumer electronics manufacturers, with its supply chain heavily reliant on China. It remains a key player in the Chinese market, with the Greater China region accounting for nearly one-fifth of its revenue. These three identities mean that Apple faces difficult choices at every decision-making juncture.

Cook’s approach is marked by a distinct personal style: cautious, restrained, and pragmatic. He rarely makes public comments on specific policies related to the U.S.-China tech rivalry, and even less often responds to sensitive issues with emotional language. Instead, he prefers to explain Apple’s position in business terms: Apple contributes jobs and investment to the U.S. economy; Apple’s relationship with China’s supply chain is symbiotic; and Apple is committed to serving consumers worldwide.

This linguistic strategy is sometimes criticized as “vague” or “evasive.” But in Cook’s view, ambiguity itself is a strategy.

In a highly polarized political environment, any clear statement could provoke one side or the other. By remaining vague, however, Apple can keep open room for maneuver. This allows Apple to maintain communication with both Washington and Beijing without having to make an either-or choice between the two. This balancing act requires extremely high political sensitivity and business acumen.

In Washington, Cook told the White House that Apple has pledged to invest $600 billion in the U.S. over the next four years, supporting the return of U.S. manufacturing and driving R&D innovation. This figure has been widely interpreted as part of Apple’s effort to demonstrate its value to the U.S. policy environment. In Beijing, he emphasized the irreplaceable nature of China’s supply chain, highlighted Apple’s journey of growing together with its Chinese partners, and underscored the saying, “Many hands make light work.”

Both sides heard what they wanted to hear. Meanwhile, Apple continues to operate globally. But this balance does not come without a price.

Apple faces criticism in the U.S. for being “too soft on China,” and in China, it occasionally runs into questions about its “lack of localization.” The pressure to shift supply chains persists, and geopolitical risks cannot be entirely eliminated. What Cook leaves behind for his successor is not a smooth, straight path but rather a complex web of relationships—deep enough yet fragile enough.

During his last visit to China in the spring equinox of 2026, Cook did not shy away from these challenges. During his meeting with Ren, Chairman of the China Council for the Promotion of International Trade, he said that the stability of global supply chains is crucial to Apple, and Apple is committed to working together with its Chinese partners to address the challenges ahead.

This sentence sounds plain, yet it carries profound meaning. It means that Cook acknowledged the existence of challenges, yet also expressed his confidence in the future. He believes that as long as the symbiotic relationship between Apple and China’s supply chain remains deep enough and both sides continue to create value, this relationship will be able to withstand external shocks.

This belief stems from thirty years of experience.

Chapter 10: One Person’s Style, a Company’s Destiny

Another hidden thread in the story of Cook and China is how his personal management style has shaped Apple’s narrative in China. Unlike Jobs, Cook is not a leader known for his intuitive flair for products. He rarely takes center stage at product launch events and seldom announces new products in a dramatic fashion.

His language is more like that of financial reports: precise, calm, and quantifiable. But this very style happens to be his most powerful weapon in China. Chinese business culture places great importance on relationships, trust, and long-term commitments. Cook’s patience, pragmatism, and sustained dedication are highly aligned with this cultural ethos. He doesn’t pursue one-off transactions; instead, he focuses on long-term partnerships. He doesn’t make promises lightly, but once he does commit, he always delivers on his word.

This style is particularly evident in relationships with Chinese suppliers. Apple has extremely stringent requirements for its suppliers. Every aspect—quality, yield rate, delivery times, and environmental standards—is subject to rigorous benchmarks. Yet Cook is also willing to provide suppliers with technical support, management training, and long-term orders. He’s not squeezing suppliers; rather, he’s helping them grow. The growth trajectories of companies such as Lens Technology, Luxshare Precision, and BYD Electronics all bear this out.

Cook’s management style has also influenced Apple’s organizational culture in China. Apple’s China team is not merely a sales branch; rather, it’s a locally rooted organization with a high degree of autonomy. They deeply understand Chinese consumers, are intimately familiar with local distribution channels, and can swiftly adjust their strategies in response to market changes.

This localized capability is the key to Apple’s continued success in China. In an interview in March 2025, Cook discussed Apple’s organizational culture. He said that Apple’s success stems from its relentless attention to detail, its cross-departmental collaboration, and its relentless pursuit of the ultimate user experience.

These words may sound like corporate propaganda, but in Cook’s mouth, they carry a more specific meaning. Every detail means that every link in the supply chain must be precise and flawless. Every collaboration means that design, engineering, manufacturing, marketing, and retail must seamlessly integrate with one another. Every user experience means optimizing the entire value chain—from product design all the way to after-sales service.

This dedication to detail is one of the most important legacies Cook left to Apple. In China, this heritage manifests itself as a unique business relationship: it’s not simply a matter of buying and selling, nor is it one-way control; rather, it’s a deep collaboration rooted in shared interests and long-term trust. Over thirty years, Cook nurtured this collaboration from a single seed into a towering tree.

Chapter Eleven: The Language of the Silent

Tim Cook is not a person who’s easy to read.

He rarely talks about his personal life, seldom displays emotions in public, and rarely uses exaggerated rhetoric to define what he’s doing. Unlike Jobs’ “reality distortion field”-style charisma, Cook is more like a mirror: What you see reflected in him is often not himself, but rather your own expectations and anxieties. Yet it is precisely this silence that constitutes the most distinctive aspect of his relationship with China.

In Chinese business culture, there’s a subtle consensus: the fewer words you say, the more room there is for interpretation. Overcommitting often implies excessive risk, whereas restraint is seen as a sign of maturity. Cook naturally aligns with this cultural ethos. He doesn’t sketch out grand visions at the first meeting, nor does he clothe his business intentions in flowery rhetoric. Instead, he prefers to let the numbers do the talking: investment amounts, employment figures, the number of suppliers, and the percentage of renewable energy used.

This linguistic style might be seen as “lacking vision” in Washington, but in Beijing it’s interpreted as “trustworthy.” Cook’s image in China differs subtly from his image in the United States. In the eyes of U.S. media, he is Jobs’ successor—a seasoned operations expert, a “guardian” who shifted Apple from being product-driven to supply-chain-driven. But in China, he’s more like an old friend, a firsthand witness to the rise of Chinese manufacturing, and a calm helmsman who remains composed even amid storms.

This difference in image wasn’t deliberately created—it’s the natural outcome of thirty years of interaction.

When Cook first came to China, he was a middle-aged supply-chain manager with thick hair and sharp eyes. Thirty years later, he’s an elderly man with silver hair, speaking more slowly and smiling more gently. His body may be aging, but his relationship with China has only grown deeper with the passage of time.

This deep-rooted bond isn’t built through just one visit or a single statement—it’s the result of countless handshakes, countless negotiations, and mutual support accumulated over countless crises.

Chapter 12: Disputes and Costs

Cook’s story with China has not been entirely smooth sailing.

Over the past three decades, Apple has faced numerous controversies in China. From labor rights to data security, from after-sales service to marketing strategies, each controversy has dealt a blow to Apple’s brand image and local trust. Around 2010, labor issues in Apple’s supply chain drew global attention.

A series of incidents at Foxconn’s factories prompted the outside world to start scrutinizing the human costs behind the consumer electronics manufacturing industry. At the time, Cook’s response was cautious. He didn’t shy away from the issue, but neither did he make any dramatic statements. Instead, he chose a more pragmatic approach: urging suppliers to improve working conditions, introducing third-party audits, and gradually enhancing transparency and standards.

This approach was criticized by some at the time as “not forceful enough.” However, in the long run, it helped Apple gradually improve its manufacturing environment without compromising supply-chain stability.

In 2013, Apple faced a crisis in its after-sales service in China. The CCTV “3·15” Gala exposed issues with Apple’s warranty policies in China, sparking consumer dissatisfaction. Cook’s response was equally intriguing. Rather than issuing an immediate public apology, he first sent a team to investigate the problem and then adjusted China’s warranty policies to bring them more in line with global standards.

This “investigate first, then act” approach reflects Cook’s consistent management style: staying unswayed by emotions and making decisions based on facts and data. But not all disputes can be resolved so smoothly.

Data security is one of the most complex challenges Apple faces in China. With the implementation of China’s Cybersecurity Law and Data Security Law, multinational corporations are required to store user data within the country and submit to regulatory scrutiny.

This poses a direct challenge to Apple’s commitment to privacy protection. Apple ultimately chose to establish a local data center in China, to be operated by Cloud on Guizhou. This decision has sparked controversy. Some privacy advocates argue that Apple has yielded to regulatory pressure, thereby undermining its brand image as a defender of privacy. However, Apple counters that this is a necessary condition for operating legally in China, and that Apple still retains cryptographic control over user data.

Cook’s stance on this issue reflects his consistent approach of striking a balance.

He didn’t openly defy regulators, nor did he completely abandon the principle of privacy. Instead, he chose a middle path: while complying with the law, he sought to protect users’ rights as much as possible. This approach couldn’t satisfy everyone, but it allowed Apple to continue operating in China. Controversy is an unavoidable reality for multinational corporations. Cook’s approach may not be dramatic, but it is certainly pragmatic. Over the course of thirty years, he has demonstrated that in a complex environment, what’s needed isn’t heroism—but rather patience and sound judgment.

Chapter 13: Apple in China in the Post-Cook Era

In June 2026, Tim Cook will officially step down as Apple’s CEO and assume the role of Executive Chairman of the Board. John Ternus will succeed him as Chief Executive Officer.

Tennus, a manager with a background as a hardware engineer, has long been responsible for Apple’s product design and engineering teams. His appointment marks Apple’s shift from an operations-driven approach—characteristic of the Cook era—to one driven by technological integration. But the Chinese legacy Cook leaves behind will not disappear simply because he’s gone.

Apple’s supply chain network in China is deeply integrated into China’s manufacturing ecosystem. Over 80% of iPhones are still assembled in China, and hundreds of Chinese suppliers are involved in the production of Apple products. This dependency cannot be easily replaced in the short term. Apple’s retail network in China has become an integral part of urban consumer culture. Dozens of Apple Stores are scattered throughout China’s major cities; they’re not just sales channels—they’re also brand experience hubs and community spaces. Apple’s developer ecosystem in China has nurtured tens of thousands of Chinese developers. These developers create apps, games, and services around the iOS platform, generating revenue for Apple and injecting vitality into China’s mobile internet innovation. Apple’s R&D investment in China is shifting from the manufacturing end to the innovation end. The application research labs in Shenzhen and Shanghai are now involved in the early-stage R&D and process optimization of Apple products.

These assets are the result of Cook’s three decades of dedicated effort. They won’t fall apart simply because of a change in CEO—rather, they’ll continue to flourish under the new leader’s guidance. But challenges also exist.

The technological competition between China and the U.S. continues to intensify, and the pressure to restructure supply chains has yet to dissipate. In recent years, Apple has been expanding its production capacity in places such as India and Vietnam, aiming to reduce its reliance on Chinese manufacturing. This “de-risking” strategy is a necessary move for Apple to address geopolitical uncertainties. But completely decoupling is neither realistic nor in Apple’s best interest.

China’s supply-chain advantages lie not only in cost competitiveness but also in its comprehensive industrial ecosystem, its large pool of engineers, and its rapid iteration capabilities. These advantages are the result of decades of industrialization and cannot be replicated in the short term. Cook knows this very well.

During his last visit to China, he said that Apple and China have a relationship of “growing together.” This statement serves both as a summary of the past and an aspiration for the future. Whether Ternus can sustain this relationship remains to be seen.

But what Cook left behind is not just a set of strategies—it’s a way of thinking: finding balance in a complex environment, maintaining patience amid uncertainty, and seeking win-win outcomes in the face of conflicting interests. This way of thinking is the most valuable legacy that Cook left to Apple.

Chapter 14: The Ferryman of the Pacific

Looking back over the past thirty years, the relationship between Cook and China has long since transcended mere commercial cooperation. It’s a story of trust, symbiosis, and long-term commitment. In 1996, when Cook first came to China, Apple was still a struggling computer company, and China was far from being the global manufacturing hub it is today. The connection between the two was weak and uncertain.

Thirty years later, Apple has become deeply intertwined with China. Apple’s products are manufactured, sold, and developed in China; Chinese developers contribute their creativity to the Apple ecosystem; China’s supply chain supports Apple’s products; and Chinese consumers account for nearly one-fifth of Apple’s revenue.

This kind of bond wasn’t established through a single transaction or a single policy; rather, it has been built up over three decades through countless interactions, countless compromises, and countless instances of mutual growth.

Cook is not the sole driving force behind this relationship.

The rise of China’s manufacturing sector, the growth of Chinese consumers, the innovation of Chinese developers, and the evolution of China’s supply chain—these forces have collectively shaped the relationship between Apple and China. Yet it was Cook who made the right choices at the critical moment.

He chose to invest in Chinese manufacturing even before the iPhone was born. He chose to promote Apple’s collaboration with Chinese carriers during the Jobs era.

He chose to remain patient amid the supply chain dispute rather than simply cutting ties. He chose to collaborate with Chinese companies in the AI era rather than building his own closed system. He chooses to strike a balance in the U.S.-China tech rivalry, rather than opting for an either-or approach. At the time, these choices may have seemed less dramatic—or even somewhat conservative. But in the long run, they enabled Apple to establish a deep-rooted presence in China and secured for China an irreplaceable position on the global technology map.

Cook is not a hero.

He didn’t change the world, nor did he create any revolutionary products. He was simply an exceptionally rational manager who spent thirty years building a bridge between the two sides of the Pacific Ocean.

Share

Leave a comment

Subscribe now

Financial infrastructure is a critical component of international financial governance, and cross-border payment lies at the very heart of financial infrastructure. Yet, looking at the current state of development, cross-border payment is precisely one of the weakest links in the existing international financial governance system.

Today, Agent AI is rapidly integrating into the global financial system, and payment scenarios are the key entry point for its real-world application. Over the past eight months, competition in the global intelligent payment space has suddenly intensified, with a flurry of innovative products and new business models being rolled out in rapid succession, delivering a fresh shock and posing serious challenges to the governance system of international financial infrastructure.

The Rise of Intelligent Payment and the Intensifying Competition

To understand intelligent payment, we need to grasp three layers: the structural characteristics of traditional cross-border payment, the white-hot market competition that has emerged with the rise of intelligent payment, and the fundamental shift in the value proposition (价值定位) of payment itself. The competition is so fierce precisely because payment is no longer just a simple transaction tool; it is becoming the underlying foundation that disrupts existing economic and business models, touching nearly every facet of the commercial world. This is the fundamental reason for the major transformation now underway.

(I) The Three-Tier Split Architecture and Inherent Risks of the Traditional Cross-Border Payment System

Today’s global cross-border payment system can be broadly divided into three tiers: messaging relies on SWIFT; fund transfers depend on commercial banks and their correspondent banking networks; and final settlement of currencies falls under the purview of individual sovereign central banks.

SWIFT handles messaging, covering customer access points across the United States, Europe, China, and elsewhere, but it is a private institution registered in Belgium and is subject only to the operational guidance of the National Bank of Belgium. Under this system, messaging, fund transfers, and currency settlement are separated from one another. Cross-border transactions are overwhelmingly denominated in U.S. dollars, and the final settlement chain is tied to the Federal Reserve and the U.S. Treasury, forming a key pillar of dollar hegemony and enabling payment infrastructure to be weaponized as a tool for financial sanctions. After a cross-border transaction sends a message via SWIFT, banks drive the movement of funds, and various currencies ultimately return to their respective central banks for final settlement. The flow of funds must pass through a lengthy chain of correspondent banks, leading directly to a host of pain points such as high costs and long processing times.

This system lacks unified global governance. SWIFT itself, as a private institution, is a globally critical piece of financial infrastructure, yet its governance foundation is extremely weak. The primary constraints on SWIFT come from national laws and regulations, mainly in the areas of anti-money laundering and counter-terrorism financing.

Reports from the Financial Stability Board (FSB) have long pointed out the chronic problems inherent in traditional cross-border payments: low processing efficiency, high transaction costs, insufficient financial inclusion, data fragmentation, high compliance costs, and inadequate market competition.

In a relatively benign geopolitical environment, these systemic flaws could be tolerated by the market. But against the current backdrop of heightened geopolitical risks, the fragility of cross-border payment infrastructure has been fully exposed, with direct implications for global financial stability and the financial security of individual nations.

(II) Agent AI Brings a Paradigm Shift in Payments

The most important technological change over the past eight months has been the emergence of Agent AI—intelligent agents capable of autonomous decision-making, negotiation, and optimization. Both the financial and AI communities have widely recognized that payment is the most promising use case for Agent AI, and payment is now evolving from automated payment to AI-agent-driven intelligent payment.

Traditional automated payment follows preset rules and requires no autonomous machine judgment; it executes one-way instructions with fixed amounts, fixed counterparties, and fixed times.

Intelligent payment, by contrast, relies on data to make autonomous judgments: payment amounts, routing paths, counterparties, and execution times can all be adjusted dynamically. Intelligent agents are capable of reasoning, negotiation, and optimization. Everything from small subscription bills to cross-border investments can be delegated to intelligent agents, enabling machine decision-making, machine optimization, machine-to-machine negotiation and routing, and machine-to-machine final settlement.

The traditional cross-border payment system, which has been in place for decades, has suddenly encountered the shock of an entirely new technological paradigm. The monopoly barriers of the SWIFT messaging network and sovereign currencies are now at risk of being broken.

Blockchain technology can reconstruct messaging and ledgers, while digital currencies can change the form of money itself. The market has come to realize that payment is no longer just a back-end transaction process; it can reach upward to engage customer needs, assist in iterative customer decision-making, and even steer consumer behavior.

At the same time, platform companies have also recognized that they must combine their AI capabilities with payment capabilities to build a complete industrial chain and form a closed commercial loop. In the past, payment was merely a passive third-party service.

Now, with its active decision-making attributes, payment can influence business models and commercial behavior, evolving into the foundation of a complete business system and ecosystem. Payment has officially shifted from a “passive execution tool” to an “active decision-making vehicle”—an unprecedented change that has given rise to intense global market competition.

(III) The Reconfiguration of Intelligent Payment’s Value Proposition

The current global competition in intelligent payment is unfolding across multiple fronts, focusing on core layers such as protocol standards, platform gateways, and machine-to-machine settlement, with the competitive landscape growing increasingly fierce.

First, competition over protocol standards, vying for the “infrastructure gateway” and the “trust layer.” To achieve interoperability between different systems, a foundational framework protocol that can be executed must be established. A variety of technical protocols have already emerged in the market, including ACP, UCP, and NPP.

One particularly noteworthy development is the practice of Google and the FIDO Alliance: Google donated its proprietary protocol to the FIDO Alliance and, leveraging its industrial influence, pushed its corporate private protocol to become nearly a global commercial standard on GitHub.

A fundamental shift is taking place here: in the past, governance rules were textual terms mutually agreed upon by multiple parties; now, rules can be executed directly through program code. Different platforms and different business scenarios can generate differentiated code-based rules, which not only pushes the system toward decentralization but also intensifies competition over underlying technical standards.

Second, competition at the platform and gateway layer, capturing “user intent.” The biggest change with intelligent payment is that it extends upward to directly interface with and understand user intent. What the user outputs is merely a natural-language expression of intent with certain constraints, while the payment process must output clear transaction objects, goods, amounts, and settlement instructions. This process relies on large language models to interpret vague subjective needs and then translate them into definitive payment execution actions.

Third, competition in machine-to-machine (M2M) settlement, building a “moat of trust.” The core transformation in M2M settlement is the shift in risk control from “identifying people” to “identifying machines,” and from customer credit governance to intelligent-agent behavior governance.

This is the entirely new underlying logic of compliance and security in the era of intelligent payment. The traditional payment compliance framework is relatively mature, centered around anti-money laundering, counter-terrorism financing, transaction transparency, and the like, with KYC (Know Your Customer) as the core of trust. But when intelligent agents replace humans in transaction decision-making, the traditional trust mechanism is fundamentally overhauled, and KYA (Know Your Agent) becomes the new core principle.

The most profound and essential change brought by intelligent payment does not lie in the payment action itself, but in the fundamental repositioning of payment’s value. Intelligent payment can be deeply integrated with enterprise resource planning (ERP) and customer relationship management (CRM) systems, embedding itself across the entire business processes of procurement, production, and cross-border supply chains.

Traditional back-end financial functions such as clearing, settlement, foreign exchange trading, and risk management are gradually being handled by decentralized networks. Intelligent agents can autonomously select optimal cross-border payment routes, completely breaking free from the traditional model of cascading correspondent-bank layers.

Over the past eight months, the race for product innovation in global intelligent payment has not been about a single payment tool, but about the underlying gateway to future digital business models and business ecosystems—a move of profound strategic significance.

Governance Challenges for International Financial Infrastructure

In the face of the transformation of the payment system driven by artificial intelligence, the governance system of global financial infrastructure has shown significant lag, with existing rules and frameworks clearly unable to keep pace with real-world developments. At present, major international institutions have begun to establish preliminary research frameworks on intelligent payment, but they have yet to develop mature governance or implementation plans.

The G20 launched a project to improve the cross-border payment system in 2020, but looking at the current reality, the original goals of the project are already difficult to achieve. The Financial Stability Board (FSB) systematically outlined the pain points of cross-border payments in 2022 and proposed 15 priority actions to enhance interoperability, legal and regulatory frameworks, supervisory frameworks, data exchange, and messaging standards, setting quantitative targets for the end of 2027.

Based on the assessment results from two months ago, these targets are largely unlikely to be met on schedule. More critically, this framework was built on the traditional payment system; even if fully realized, it would already be outpaced by reality and unable to address the new issues raised by intelligent payment.

In March 2026, the International Monetary Fund (IMF) proposed a three-tier theoretical framework for intelligent payment: the first tier relies on AI and large language models to parse and understand user intent; the second tier is a human-machine collaborative control and authorization layer, responsible for risk management; and the third tier entrusts deterministic intelligent agents with final transaction execution.

This framework is logically clear and theoretically compelling. It aims to fully leverage AI’s reasoning and judgment capabilities while incorporating human oversight to balance transaction speed and compliance risks. However, the core difficulties in practice remain unresolved. The IMF has offered more of a conceptual framework and guiding opinions, but has not provided mature answers on how exactly to implement it.

The Bank for International Settlements (BIS) Agorá project has explored an entirely new clearing model, achieving near-instant atomic settlement through the tokenization of central bank money and commercial bank money on a unified platform.

The core advantage of this model is that it does not alter the boundaries of central bank sovereignty or change the rights and responsibilities of commercial banks; it only requires the tokenization of transacted currencies, enabling centralized information processing and settlement on the platform while eliminating the complex correspondent-bank chain.

On July 30, the BIS released the latest results from its real-value testing: in collaboration with five central banks and 22 financial institutions, it completed 30 real transactions totaling approximately US$1 million, covering six currencies. The average time from transaction initiation to final settlement was 80 seconds, with no need for cross-system reconciliation, no correspondent-bank chains, no waiting for overnight clearing windows, and multi-currency settlement completed simultaneously.

On the surface, these results are nothing short of remarkable. However, there are critical constraints: 30 small-scale transactions executed on a dedicated network can be highly efficient, but if scaled up to billions of real-world transactions, relying on the BIS-built system while at the same time ensuring that central bank sovereignty is not undermined and that commercial banks’ existing business boundaries are not upended—this goal is extremely difficult to achieve in practice.

Currently, a set of noteworthy phenomena is emerging in global digital currency regulation: the United States, the European Union, and Hong Kong (China) are accelerating the improvement of stablecoin (稳定币) regulatory frameworks, while some central banks (such as those in Canada, Malaysia, and Thailand) are slowing down or pausing the development and rollout of central bank digital currencies (CBDCs).

Looking at these two trends together, it appears that a preliminary consensus has formed in global digital currency theory: stablecoins and digital currencies are market-driven commercial behaviors, not sovereign currencies, and central banks do not intervene—this is a particularly important distinction from sovereign currency.

The United States is comprehensively promoting the development of market-driven dollar stablecoins, with a very clear policy direction. At the regulatory level, constraints have been moderately relaxed; in certain cases, U.S. stablecoin issuers have even been granted special access to the Federal Reserve’s discount window, so that in crisis situations they can receive Fed liquidity support, even though stablecoins themselves are not sovereign currency. At the same time, the eSLR (enhanced Supplementary Leverage Ratio) capital regulatory rules have been adjusted, imposing lower capital requirements on relevant institutions compared to ordinary commercial entities.

Behind the U.S. support for stablecoins lies a clear strategic calculation: stablecoins can serve as a new vehicle for perpetuating dollar hegemony. Stablecoin issuers place a large portion of their reserves in U.S. Treasury bonds; today, the stablecoin sector as a whole has become the eleventh largest holder of U.S. Treasuries. The global stablecoin market is dominated by dollar stablecoins, which are being exported outward from the United States, serving as a new vehicle for the expansion of the dollar system. Yet dollar stablecoins themselves carry significant risks. From the publicly disclosed reserve structures of USDT and USDC, the actual cash component of their reserves is only about 6%, with the bulk of the remainder consisting of overnight repurchase agreements, primarily U.S. Treasuries of various maturities. The overall reserve portfolio is essentially a large money market fund. Historically, in every financial crisis, money market funds have been highly vulnerable to market liquidity and credit risk shocks.

China has adopted a dual-track approach, with CBDC and the digital renminbi (e-CNY) developing in parallel. The e-CNY serves as both a payment instrument and a reserve instrument, as reflected on commercial banks’ balance sheets. A critical point is that CBDCs and the e-CNY are sovereign currencies, fundamentally distinct from market-based stablecoins.

In the future, the world will see a long-term tripartite competition among three major digital currency tracks: the U.S.-led market-based dollar stablecoin system, Europe’s plan for a central-bank-backed euro stablecoin (expected to be launched by 2029), and China’s sovereign digital renminbi and CBDC framework.

The three paths are logically independent and mutually incompatible in terms of rules, and the high-level global contest over digital currencies and intelligent payments has fully begun. Yet, currently, the financial regulatory frameworks and international governance mechanisms for digital currencies remain absent—a major challenge we now face. Stablecoins are not sovereign currency; they are neither central bank money nor commercial bank money, and they carry inherent additional risks.

Moreover, their safety boundaries, governance rules, and management systems remain unclear. If stablecoins are to be used as a primary vehicle for cross-border payments, the entire cross-border payment system would need to be restructured to address the exogenous risks they introduce.

The technical architectures for cross-border CBDC interaction (peer-to-peer direct connections, hub-and-spoke models, and third-party intermediaries) have yet to reach global consensus, severely constraining the standardized development of global digital financial infrastructure.

China’s Current Cross-Border Payment Landscape and Challenges

China has already established a three-tier, parallel cross-border payment infrastructure system: the traditional RMB cross-border settlement system (CIPS), the “Shuzhida” digital RMB cross-border platform, and the multi-CBDC bridge project mBridge, forming a differentiated and complementary strategic layout.

The CIPS system continues to expand its business volume, with transaction volume last year reaching approximately RMB 190 trillion (USD 28.22 trillion), yet its scale remains limited within the broader global cross-border payment system.

During the Lujiazui Forum on June 18 this year, “Shuzhida” was officially launched—a comprehensive platform for digital asset and payment services that leverages blockchain capabilities for cross-border use of the digital RMB. This is a strategically significant move. We anticipate a highly uncertain external environment ahead, with the world moving toward a pattern of multiple coexisting currencies, where market-issued currencies and central bank sovereign currencies will develop in parallel.

Therefore, from the very outset, the platform was designed with inclusiveness in mind, capable of accommodating multiple models. Some functions of the platform have already been deployed, while a number of blockchain-related modules are still under development; traditional non-blockchain supporting services are already in place, and the overall platform remains in a state of continuous iterative construction.

The mBridge project, co-initiated by the People’s Bank of China, the Hong Kong Monetary Authority, the Bank of Thailand, and other institutions, is building a cross-border payment interoperability network and transaction hub to help address the fragmentation of global cross-border payments.

It is worth noting that China’s domestic payment system is primarily built on traditional technological architecture and is not fully based on blockchain, while the future trend in cross-border payments is increasingly moving toward blockchain-based paradigms.

Therefore, how to effectively bridge China’s mature digital payment capabilities with cross-border blockchain financial networks is currently a core challenge. Some leading domestic payment institutions have already taken the lead in piloting tokenized cross-border payment projects overseas, joining the global competition in intelligent payment, but overall, China is still in the early catch-up phase.

Overall, the traditional cross-border payment system is clearly lagging behind; intelligent payment has only just emerged but competition is already fierce; a financial infrastructure framework based on digital currencies and blockchain is gradually taking shape; yet the field of global financial infrastructure governance remains weak.

How China can improve its own positioning and play a role in the governance of global financial infrastructure is both a major challenge and a vast area for research.

Share

Leave a comment

For the past several years, one question has persistently shadowed discussions about China’s role in the global economy: Will the world’s supply chains finally break free from the Middle Kingdom?

The term “decoupling” first gained traction during Donald Trump’s first presidency, when Washington launched a full-blown trade war against Beijing. Then came the pandemic, which exposed the fragility of global supply chains and prompted many governments to rethink their over-reliance on a single manufacturing hub. The consensus in many Western capitals seemed clear: it was time to diversify, and that meant moving production out of China.

Yet reality has stubbornly refused to follow the script. When COVID-19 upended logistics networks worldwide, countries that had talked tough about reducing dependence on China found themselves scrambling for goods that only Chinese factories could supply in sufficient quantity and quality. The pandemic proved an uncomfortable lesson: China’s place in the global industrial ecosystem is not easily erased.

Decades of building an integrated manufacturing system—from raw materials and components to final assembly and after-sales support—have created a web of interdependencies that cannot be replicated overnight, nor by relocating a few assembly lines.

Subscribe now

The “Big” Advantage That Is Often Overlooked

Let’s start with the sheer scale of Chinese manufacturing. By some estimates, China now accounts for roughly 45 percent of global manufacturing output in real terms—soon to surpass the combined share of all G7 nations.

For years, Chinese observers have lamented that their country is “big but not strong,” as if size itself were a liability. But in manufacturing, scale is a capability in its own right. This is not a contest between individual firms or products; it is a contest between entire industrial ecosystems. To produce a finished good competitively, a nation needs not just final assemblers but also a vast army of suppliers, subcontractors, skilled workers, and accumulated production know-how. That is precisely what makes China so difficult to replace.

Yet size alone does not capture the most telling shift in recent years. In the early post-WTO era, Chinese exports were synonymous with labor-intensive goods—textiles, toys, and footwear—earning the label “China Shock 1.0” in Western economies.

Today, we are witnessing “China Shock 2.0”: an ever-growing stream of high-quality, technologically sophisticated products, from electric vehicles to advanced machinery, flooding developed markets. Chinese companies no longer compete on low wages alone; they compete on innovation, speed, and quality.

Crucially, China has upgraded its industrial structure without abandoning its traditional strengths. While the share of garments and footwear in total exports has declined, the country remains a formidable producer of these everyday goods. That matters because a truly resilient manufacturing power cannot rely solely on a few high-tech sectors; it needs depth across the board. The ability to churn out both cutting-edge EVs and low-cost consumer goods with equal efficiency is what gives China its unique industrial resilience.

Consider the disposable lighter. Once synonymous with Wenzhou’s bustling workshops, lighters are now produced in places like Shaodong, Hunan, where an entire ecosystem of mold makers, injection-molding specialists, and logistics providers ensures that millions of units roll off production lines at pennies apiece, year after year. The product itself is simple; the system behind it is anything but. That system, not any single factory, is China’s true competitive edge.

To understand why this industrial prowess persists, however, we must look beyond scale and look at what drives it: sustained technological progress.

The Engine Behind Upgrading: Long-Term Investment

In economics, total factor productivity (TFP) is a key measure of how efficiently an economy uses its inputs. Since 2009, China’s TFP has grown at an average annual rate of roughly 2.2 percent—a sign that growth is not merely about pouring in more capital but about genuine efficiency gains and technological improvements.

Such gains do not materialize from thin air. China’s R&D spending now accounts for about 2.8 percent of GDP, placing it among the world’s top spenders in absolute terms. Given the size of China’s economy, even a moderate R&D-to-GDP ratio translates into a massive pool of research funding.

Over decades, this has created a virtuous cycle: more investment begets more patents, more scientific publications, and more top-tier talent. Several Chinese universities now rank among the world’s top 50, and their graduates are feeding directly into industrial R&D labs.

This is why many multinational companies—especially from Europe—have not abandoned China despite political pressures. They see not just a vast consumer market but also a rich ecosystem of suppliers, engineers, and research partners.

For any rational business, supply chain decisions are ultimately driven by cost, efficiency, technology, and market access—not by campaign rhetoric. The real-world logic of commerce has so far proven more resilient than the political logic of decoupling.

It is also this industrial and technological foundation that has given China far more staying power than many observers anticipated. U.S. sanctions, from trade tariffs to semiconductor export controls, have not pushed China out of global value chains.

Instead, they have accelerated indigenous innovation in precisely those areas where foreign technology was once dominant. The Sino-American rivalry has thus entered a new phase—one in which Chinese firms are no longer passive rule-takers but active competitors rethinking their global strategies.

”Re-Creating China Overseas”: The New Wave of Outbound Investment

What we are witnessing today is not the world abandoning China; it is Chinese companies going global on their own terms. The narrative of “decoupling” often misses this point entirely. While trade frictions have altered the geography of production, they have not diminished China’s connection to global markets—they have merely rerouted it through new channels.

The shift is most visible in the wave of outbound direct investment. For decades, Chinese firms built their success on domestic manufacturing and exports. Now, facing both intense competition at home and rising tariff barriers abroad, they are setting up factories, supply chains, and R&D centers in Southeast Asia, Latin America, Africa, and even Europe. This is not a flight of capital but a strategic redeployment—what some call “re-creating China overseas.”

In the past, the mantra was “China + 1”: keep production at home while adding a backup facility elsewhere. Today, it is evolving into “China + N”: building multiple nodes in different countries to create a flexible, risk-diversified network. This is precisely how Britain, the United States, and Japan expanded their economic influence during their own heydays. By comparison, China’s overseas assets remain relatively small as a share of GDP, suggesting that this globalization wave has only just begun.

But going abroad is not simply a matter of transplanting the Chinese factory model to foreign soil. It requires learning new skills—how to navigate local regulations, how to build trust with foreign communities, how to balance short-term losses against long-term gains. These are not problems that China’s manufacturing prowess can automatically solve.

Learning to See the World Eye-to-Eye

The first mental adjustment is what might be called “learning to see the world horizontally.” For generations, China looked outward with a mixture of awe and anxiety, absorbing foreign technologies and management practices while often deferring to Western models.

That era is ending. As Chinese firms venture into every corner of the globe, they are discovering that every country has its own strengths and weaknesses. Neither adulation nor condescension serves the purpose; what works is a pragmatic, respectful engagement with local realities.

Second, Chinese companies must master the art of using rules—not just following them. For too long, the conversation about international rules has focused on how others use them to constrain China.

Now that Chinese businesses are global players, they need to understand how to advocate for their interests within legal and regulatory frameworks, how to shape standards, and how to defend their rights in foreign jurisdictions. This is not about confrontation; it is about becoming sophisticated stakeholders in a multipolar world.

Third, there is the question of costs. Entering a new market inevitably involves upfront investments, cultural adaptation, and sometimes uncomfortable compromises.

Chinese firms will need to learn the difference between visible losses and hidden costs—the latter often being far more damaging in the long run. A willingness to bear short-term “visible” expenses in order to build sustainable local relationships will be the hallmark of truly globalized Chinese enterprises.

A Long Road Ahead

The phrase “re-creating China overseas” is not about exporting a development model wholesale, nor about hollowing out the domestic economy. It describes a natural evolution: as Chinese firms mature, they are stretching their production, supply, and asset networks across the world, just as earlier industrial powers did before them. Manufacturing remains the bedrock, technology remains the driver, and globalization opens the new frontier.

But this journey is still in its early stages. The capabilities that made Chinese firms unbeatable at home—efficiency, cost control, speed—will not automatically guarantee success abroad.

The next decade will test not just their industrial might but their cultural intelligence, legal savvy, and strategic patience. If they can pass that test, then “China overseas” will be not a slogan but a new reality—one that reshapes the global economy as profoundly as China’s own industrialization did over the past half-century.

The road ahead is long, but it is Chinese companies themselves that are paving it.

Share

Leave a comment

Last year, China exported 249.8 gigawatts of solar modules—an almost unimaginable amount, enough to carpet half the planet. It’s a story we’ve all heard in passing.

But for the average Chinese person, “solar power” doesn’t carry much weight in daily life. At best, it means a few panels on the rooftop, feeding a little daytime juice back to the grid, skimming a few hundred yuan off the annual electricity bill—a nice supplement, but nothing essential. The grid is still there, the AC still hums, takeout orders still fly in; the panels might keep the lights on during a rare outage, but that’s about it.

Ship that same Chinese-made solar panel to Cuba, to a farm in rural Pakistan, or to a telecom tower in the African bush, and the picture changes completely. In these places, solar isn’t about saving money—it’s about having electricity at all.

With it, you can cook, keep your business running, charge the fan in your child’s room; without it, you’re plunged back into a pre-industrial darkness.

The 2025 data tells the story. China’s solar exports used to go mainly to developed countries in Europe and America. But for the first time this year, developing countries bought more than their wealthier counterparts. In just twelve months, the number of countries importing over 1 gigawatt annually jumped from 39 to 47. Each new addition represents another power system being rebuilt from the ground up.

Subscribe now

Havana’s Electric Tuk-Tuks

In 2026, Havana’s streets are swarming with Chinese-made electric tricycles—tuk-tuks. They haul freight, carry passengers, run fixed routes as minibuses; in some neighborhoods, they even collect garbage.

Prices range from 10,000 yuan (1,484 U.S. dollars) to nearly 30,000 yuan. That’s exorbitant, because many are shipped back via Panama by relatives or importers; some buyers sell off their gasoline cars to afford one. Many owners shell out another 3,000 or 4,000 yuan to mount a solar panel and rack on the roof, so they can charge on the go.

Just a few days ago, Johnny (小钟), a Bilibili vlogger from China who roams the globe, posted a travel video from Cuba. After dinner, he hopped onto a tuk-tuk—instantly recognizable to any Chinese viewer—and remarked that these vehicles have become an everyday sight in Cuba.

Why have tricycles and solar panels become such necessities in Cuba? Because the country’s energy system has deteriorated to a state outsiders can barely fathom.

First, oil. Cuba has long depended on Venezuela and Mexico for petroleum, but U.S. sanctions have choked off supply. In the first ten months of 2025, total fuel imports from all sources averaged just 45,400 barrels per day—a 35% drop year-on-year.

By 2026, official gasoline prices hover around $2 a liter; on the black market, severe shortages have driven prices as high as $8. Meanwhile, the average monthly salary in Cuba’s state sector is around $10; in the private sector, it’s better, but still under $45.

Do the math: one liter of black-market gas can eat up nearly a month’s wages for a state worker. Gasoline, in Cuba, has simply become unaffordable for ordinary people.

Then there’s electricity. In May this year, some neighborhoods in Havana experienced outages of 20 to 22 hours a day—leaving just two or three hours of power, barely enough to top up a phone charger.

As of March, 9 of the 16 generator units across the country’s seven thermal plants were shut down. They burn Cuba’s high-sulfur domestic crude, which wears out equipment at an accelerated rate; spare parts are nearly impossible to import under sanctions, so broken units stay broken. In May, the Minister of Energy and Mines announced that the country’s diesel and fuel oil reserves were completely depleted.

Solar capacity, meanwhile, stands at 1,300 megawatts—but the grid is so antiquated that dispatch and transmission can’t keep up, and much of that solar output is simply wasted. As of July 2026, nationwide blackouts continue, and there’s no quick fix.

With oil unaffordable and power unreliable, a single solar panel takes on a meaning utterly different from what it has in China.

In the central city of Santa Clara, a free solar charging station has appeared: 20 device outlets, 16 vehicle charging spots, and 12 cooking stations, all running simultaneously. A modest 30-kilowatt solar system powers an entire block’s worth of tricycle charging, rice-cooker meals, phone recharging, and micro-businesses. One mother who sells sandwiches used to have to get up at 2 a.m. to cook when power was available; now she brings her pots to the station and fries fish alongside her neighbors.

In a country where oil is out of reach and the grid teeters on collapse, a solar panel plus an electric tricycle amounts to a self-sufficient mini-survival system.

Cuba is the most extreme example. Elsewhere, Chinese solar hardware does different jobs, but they all share a common thread: replacing diesel.

Africa has about 500,000 telecom towers, many of them previously powered by diesel generators, with fuel costs alone eating up 60% of operating expenses. When Nigeria removed its fuel subsidies, diesel prices doubled within a year. Switching to solar slashes fuel bills by 30–50% at a stroke.

And these towers aren’t just for phone calls—they support mobile payments, medical emergencies, and the community’s link to the outside world. If a tower goes dark, the whole neighborhood goes dark. Some sites are now being developed as small solar microgrids, with surplus power shared with nearby villages.

Panels at 7 Cents a Watt—Better Than Oil

In 2025, Chinese solar module prices bottomed out at around 7 U.S. cents per watt. A standard 550-watt panel comes out to under $40.

For that same $40, you can buy a few days’ worth of diesel—and then it’s gone. A solar panel lasts 20 years or more. Turning a monthly fuel expense into a one-time equipment investment: install once, stop buying oil every month. The arithmetic is undeniable.

In many developing countries, the grid is either intermittently functional or nonexistent; daily electricity comes from diesel generators, with fuel trucked in from far away. The farther the site, the higher the transport cost—and the fuel can be stolen, adulterated, or simply fail to arrive during the rainy season.

Over just a few years, Chinese module factory prices dropped 70% from $0.25 per watt. The brutal price competition among Chinese manufacturers has effectively opened a procurement window for these countries. Here, solar isn’t competing with grid power—it’s competing directly with diesel.

So solar spreads through these regions in a way completely different from China: not from national power plans, but from the bottom up—from individual users who do the math and realize solar is cheaper than fuel.

Pakistan has taken this logic the furthest. Electricity prices tripled in three years, while Chinese module prices halved in the same period. With tax-free import channels open, households, businesses, and farmers began installing rooftop solar on their own.

According to the International Energy Agency, Pakistan added around 10 gigawatts of solar in 2025—almost all of it distributed, meaning it wasn’t a government-planned rollout, but the collective decision of millions of households after running their own numbers.

Pakistan also has 1.5 to 2 million agricultural tube wells, about 80% of which used to run on imported diesel. Community electricians and small traders learned to install solar pumps through online videos, delivering them farm by farm. Solar here spreads like mobile phones—from the bottom up, without waiting for a master plan.

And its impact runs deeper than phones. If a phone breaks, you replace it; but once energy infrastructure is plugged in, the entire system transforms. In Pakistan, as rooftop panels multiplied, daytime grid power became surplus. The government realized it didn’t need to buy as much natural gas, renegotiated LNG contracts with Qatar, and cancelled some cargo orders from Eni. A tiny rooftop panel, installed by enough people, can shift a nation’s LNG procurement.

From Module Exporter to Energy-System Integrator

China now has 47 major buyers of solar panels. But selling panels is just the first layer—and only the beginning.

In 2025, exports of solar cells nearly doubled; inverter exports surpassed $9 billion and are growing. Panels remain the mainstay, but the supporting hardware is catching up fast.

The reason is simple: a solar panel only works during the day. When the sun sets, photovoltaic output drops to zero. To keep the lights on at night, you need batteries to store daytime surplus and power-control equipment to manage the flow. Chinese companies are now packaging these together for export.

The UAE offers a telling example. They’re planning a massive solar complex with 5.2 GW of PV capacity paired with 19 GWh of energy storage—the goal being 24/7 stable power even after sundown.

The supplier list reads like a who’s-who of Chinese industry: panels from JinkoSolar and JA Solar, storage batteries from CATL, construction by PowerChina and India’s L&T. For the two core components—generation and storage—Chinese firms dominate.

This trajectory may feel familiar. Japanese automakers followed a similar playbook: first they shipped cars, then they handled repairs, then they supplied parts, and finally they built factories locally. Chinese solar is treading the same path.

Electric mining trucks are following the same route. In a case we covered earlier, hundreds of mining haulers were shipped overseas, but months before they arrived, the manufacturer’s team was already on-site building a solar-plus-storage microgrid and laying power lines.

From an empty pit, they built out generation, storage, charging and dispatch—a whole system stacked with Chinese vendors. The trucks couldn’t charge without the grid; the panels would be meaningless without the trucks. The whole package moves together.

The panel itself is already the cheapest and most commoditized link. The real competition now lies beyond the panel: after the modules arrive, how do you configure storage, connect to the grid, handle local installation and maintenance, and manage daytime surplus for nighttime use? That work is just beginning.

We don’t yet know when the UAE’s 5.2-GW mega-plant will come online. But in Cuba, the sandwich-selling mother no longer has to wake up at 2 a.m. to cook.

It’s just an unassuming sheet of glass and silicon. But its significance transcends any grand narrative.

Adapted from a blog by Coolplay Lab (酷玩实验室).

Share

Leave a comment

Subscribe now

If you have been following the AI hardware story, you already know about the GPU shortage, the HBM memory crunch, and the power infrastructure bottlenecks. But there is another, less visible supply chain crisis quietly building—one that involves a chemical compound most people have never heard of.

That compound is “indium phosphide (InP) .”

In February 2025, the Ministry of Commerce and the General Administration of Customs of China issued Announcement No. 10, placing InP, organometallic precursors such as trimethylindium and triethylindium and related production technologies under export control.

In fact, InP sits at the very top of the optical module supply chain. High-speed optical transceivers and co-packaged optics (CPO) require optical chips, and more than half of those optical chips are manufactured using indium phosphide substrates. As AI data centers scale from 800G to 1.6T and eventually 3.2T, the demand for InP substrates is exploding—and global supply is nowhere close to keeping up.

The Numbers Are Staggering

According to multiple industry estimates, the global supply-demand gap for InP substrates will exceed 70% in 2026. Let us break down the numbers.

2025 global demand for InP devices was approximately 2 million wafers per year (measured in 2-inch equivalent). Actual production capacity was only about 600,000 wafers per year .

\For 2026\, third-party consulting firm Yole projects that global demand will climb to 2.6 million to 3 million wafers per year, while effective capacity will only increase to about 750,000 wafers per year . That translates to a supply-demand gap of more than 70% .

In July 2026, Lumentum CEO Michael Hurlston made a stunning public statement at the Paris European AI Summit. He said that the current supply-demand gap for InP substrates has already surpassed that of mainstream memory chips like DRAM and NAND flash .

“I don’t believe our capacity—together with Coherent’s—can meet NVIDIA’s and other customers’ demands,” he said .

This is a remarkable admission from the CEO of one of the two companies that together control more than 80% of the high-end InP optical chip market . The other is Coherent.

In fact, the situation has become so dire that Nokia—one of Lumentum’s largest customers—has been forced to get into the manufacturing business itself. Last year, Nokia acquired Infinera and with it, Infinera’s InP wafer fab in California. More recently, it acquired NXP’s Arizona plant with the explicit purpose of converting it into an InP optical chip production line .

Downstream complaints are mounting, and pressure is being passed up the chain to the InP wafer manufacturers. The dominant suppliers—Sumitomo Electric of Japan and AXT of the United States—have become the targets of industry frustration. But AXT has pushed back, repeatedly emphasizing at recent earnings calls that the root cause of the shortage is actually something else entirely .

A Three-Link Chain with Three Different Realities

To understand the InP shortage, you need to break down the InP wafer production process into three distinct steps:

1. **Primary indium smelting** —extracting raw indium from ores

2. **Indium purification** —upgrading primary indium to high-purity semiconductor-grade indium

3. **InP wafer manufacturing** —growing single crystals and fabricating substrates

Each of these three links has a completely different competitive landscape, and each is affected differently by the current crisis.

Link One: Primary Indium — China’s Stronghold

China’s biggest advantage in this chain is pure geography and scale.

Global indium reserves are estimated at just 50,000 tons, of which only about half is economically recoverable. Of the proven reserves, China accounts for more than 70% .

But reserves alone do not tell the full story. Indium is rarely found in concentrated deposits—no rich indium ores have ever been discovered. Instead, it is almost always produced as a byproduct of zinc and tin smelting, like the dregs left over from making “tofu”(豆腐). No one builds a tofu factory just to get the dregs; similarly, indium production is tied to zinc and tin supply, not to indium prices.

China produces the most “tofu” in the world, and therefore also produces the most “dregs.” China’s primary indium production accounts for more than 50% of global output. At its peak, 80% of the 400 tonnes of indium produced globally each year were exported from southwest China’s Yunnan Province alone .

Link Two: Indium Purification — A Tale of Two Markets

Primary indium typically has a purity of 95% to 99%. It is first refined to 4N–5N grade indium (purity of 99.99% to 99.999%), and then further purified to 6N grade (99.9999%) before it can be used to manufacture InP optical chips.

China accounts for 70% to 80% of global refined indium production, making it the absolute dominant player in this segment.

However, the 6N and above high-purity indium market—the semiconductor-grade material essential for InP single-crystal growth—is still dominated by Japanese manufacturers such as Dowa and JX Metals . These Japanese players have long controlled the high-end purification market. But here is the irony: almost all of the primary indium they use for purification comes from China.

Link Three: InP Wafer Manufacturing — The Weakest Link with a Twist

InP wafer manufacturing is China’s weakest link in the chain—but there is one major exception.

Beijing Tongmei (北京通美) is the Chinese operating entity of U.S.-based AXT, which is currently the world’s second-largest InP wafer manufacturer (after Japan’s Sumitomo Electric) . Years ago, AXT relocated all of its InP wafer production lines to China specifically to be closer to the primary indium supply and save on transportation costs. Today, Beijing Tongmei handles all of AXT’s physical business operations, including manufacturing. The North American parent company is primarily responsible for capital operations.

This creates a fascinating situation: every InP wafer AXT sells—even those shipped back to the United States—is theoretically manufactured in China and must go through Chinese export procedures.

Together, Sumitomo Electric, AXT, and JX Metals control more than 90% of the global InP substrate market . All Chinese domestic manufacturers combined (excluding Beijing Tongmei) account for less than 10% of global supply.

One of China’s largest domestic producers is Xinyao Semiconductor (鑫耀半导体), a subsidiary of Yunnan Lincang Xinyuan Germanium (Yunnan Chihong Zinc & Germanium) . Its current annual production capacity is approximately 150,000 wafers per year.

The Export Control Wild Card

This is where the story takes a geopolitical turn.

In February 2025, China placed indium phosphide and related products on its export control list. The first major casualty was Beijing Tongmei (AXT). Although the company reportedly obtained an export license by June 2025, the allowed quantity was “less than expected,” which caused its fourth-quarter revenue to miss targets .

The long-term uncertainty created by these controls has placed a heavy burden on an already fragile supply chain. At its second-quarter earnings call, AXT said that even with strong demand and sufficient capacity, its ability to ship to customers outside China depends entirely on the Chinese government’s licensing decisions.

In early 2026, Beijing extended export controls further upstream to indium itself. The government imposed a cap limiting indium exports to 30% of total domestic output. Moreover, export approvals for 6N-grade and above high-purity indium were further tightened, effectively putting a stranglehold on the arteries of the InP supply chain.

This was a direct hit on Japanese suppliers like Sumitomo Electric and JX Metals, which source more than 90% of their high-purity indium raw materials from China. According to industry reports, Sumitomo’s annual production capacity has already dropped from 180,000 wafers to fewer than 100,000 wafers. Delivery lead times have stretched from under six months to 18 months.

Given this context, it is easy to understand why Coherent’s CEO was one of the first to secure a seat on the Air China flight accompanying President Trump’s delegation to Beijing in May 2026 . Officially, the trip was about “promoting U.S.-China optical chip cooperation.” But industry insiders knew the real purpose: to secure supply.

Coherent, along with Lumentum, is one of the world’s largest suppliers of InP optical chips. NVIDIA has already made a $2 billion strategic investment in Coherent to secure its CPO supply chain. In June 2026, Jensen Huang personally attended the topping-out ceremony of Coherent’s new factory, encouraging the partner to accelerate capacity expansion .

The “Resources Strong, Manufacturing Weak” Syndrome

On the ground in China, industry insiders describe the current situation as strong in resources, weak in manufacturing.

This is how the three links shape up in reality:

- **Primary indium: Abundant and well-supplied. Nearly 60% of China’s refined indium production is sold to the ITO target market. Total reserves are more than adequate.

- **High-purity indium (6N/7N): Supply is tightening rapidly. With the rise in AI computing demand, the high-purity indium market is under severe strain. Domestic effective mass production capacity is still limited in the short term. However, the actual consumption is tiny—a 2-inch InP wafer requires only 1 to 2 grams of high-purity indium. Even at Yunnan Germanium’s current capacity of 150,000 wafers per year, the total high-purity indium consumption would only be 300 to 400 kilograms . This small volume has, until recently, given little incentive for upstream refiners to invest in high-purity purification capacity.

- **InP wafers (3–4 inches): Domestic players are now competitive and have captured most of the Chinese market.

- **InP wafers (6 inches): The domestic localization rate is less than 5%. Yunnan Germanium is currently the only Chinese company capable of mass-producing 6-inch InP wafers, and even that is still in early stages .

According to one InP substrate company executive, the Chinese InP industry is just getting started. “In the past, the overall market demand for InP was not this large. Now, AI infrastructure construction is driving a rapid increase in InP material demand, but capacity expansion takes time.”

Regarding the widely cited “70% supply-demand gap,” industry players are taking a cautious approach. The securities department of Yunnan Germanium said that the company does not make major investment decisions based solely on market hype or third-party research papers.

“We make expansion decisions based on two things,” the company said. “First, we reference the likely order volume our downstream customers can give us. Second, we assess our own production and operational conditions to determine whether we have the capability to complete the project.”

The company’s current expansion plan aims to add 300,000 wafers per year in equivalent capacity, with a construction period of 18 months.

Real Demand vs. Speculative Hype

Behind the hot market expectations, industry observers note the need to distinguish between real demand and speculative hype.

The “fundamental demand” for InP substrates is driven by AI data center construction, which has a degree of sustainability. According to Ministry of Industry and Information Technology data for January–February 2026, national computing infrastructure investment grew 42% year-over-year. The “Eastern Data, Western Computing” project, the eight national computing hub nodes, provincial-level AI computing centers, and major internet companies’ AI cluster construction are all accelerating, creating visible and tangible demand growth.

By contrast, areas like 6G and automotive LiDAR are more about strengthening long-term expectations. These fields are still in early industrialization stages, far from large-scale commercial deployment. The market may be overinterpreting the impact of these long-term applications on the current supply-demand balance .

The High-Purity Indium Puzzle

From a supply chain perspective, the InP wafer manufacturing process begins with high-purity indium and phosphorus as raw materials, which are synthesized into InP polycrystalline material, then grown into single crystals, and finally processed into wafers through cutting, grinding, and polishing.

China’s primary indium is mainly derived from lead-zinc-tin smelting byproducts, and China accounts for more than 70% of global primary indium production.

However, the critical raw materials for InP single-crystal manufacturing—high-purity indium and electronic-grade red phosphorus—remain heavily dependent on imports.

Companies like Tin Industry Co., Hua Xi Nonferrous Metals, and Zinc Industry Co. can stably supply large quantities of 4N–5N industrial-grade refined indium, with nearly 60% of sales going to the ITO target market. Total reserves are plentiful.

But with the rise in AI computing demand, supply of 7N and above semiconductor-grade ultra-high-purity indium is tightening. Domestic effective mass production capacity is limited in the short term.

According to Tin Industry Co., InP wafers for optical communications require indium purity at the 6N (99.9999%) to 7N (99.99999%) level. The company’s associated enterprise, Yunnan Tin New Materials, is advancing its high-purity indium project and already has some 7N-grade high-purity indium capacity, with expansion underway.

Yuguang Gold & Lead (豫光金铅), another high-purity indium refiner, has built a 7N-grade high-purity indium pilot production line with an annual capacity of 300 kilograms. The output has been supplied to downstream customers for trial verification, with feedback expected by the end of 2026 or early 2027. The company will evaluate expansion based on market conditions after receiving customer feedback.

Hua Xi Nonferrous Metals has also conducted R&D work on 7N and above high-purity indium for semiconductor applications, but the products remain in the research and validation phase.

A Shanghai-based small metals industry analyst told reporters that while China has abundant refined indium resources, there are not many domestic companies that can stably produce 6N or 7N semiconductor-grade high-purity indium.

“Ordinary refined indium and semiconductor-grade high-purity indium may seem like just a few ‘9s’ of purity difference, but behind them are completely different process systems.”

The analyst further noted that there is another reason for the lag in China’s high-purity indium purification development: the demand for indium from InP wafers is not large. The consumption of high-purity indium in InP wafer manufacturing is only 1 to 2 grams per 2-inch wafer. Using Yunnan Germanium’s 150,000-wafer annual capacity as an example, the total consumption would only be 300 to 400 kilograms per year. This small volume has historically given little incentive for upstream refiners to invest in high-purity purification capacity.

“From a pricing perspective, ordinary refined indium currently sells for just over 5,000 yuan per kilogram. High-purity indium above 6N grade sells for perhaps just over 6,000 yuan. The price difference is only about 1,000 yuan—the premium is not as attractive as one might imagine,” the analyst said .

The Geopolitical Dimension: Trump, Trade, and Technology

This brings us to the broader geopolitical context.

In May 2026, as part of the Trump administration’s visit to China, Coherent’s CEO Jim Anderson was among the first executives to secure a seat on the delegation’s Air China flight . The visit was framed as a diplomatic mission, but the semiconductor industry understood it for what it was: a desperate scramble to secure InP supply.

Coherent is one of NVIDIA’s most critical CPO partners. Earlier, NVIDIA had made a $2 billion strategic investment in Coherent to ensure its supply chain resilience. In June, Jensen Huang personally attended the topping-out ceremony of Coherent’s new factory, a clear signal of the strategic importance of InP optical chips to NVIDIA’s AI roadmap.

The export controls imposed by China have effectively weaponized the country’s upstream dominance. With indium and InP both under export restrictions, overseas manufacturers face critical supply uncertainty—while Chinese companies gain a structural advantage.

But the irony is not lost on industry observers. AXT’s InP wafers are manufactured in China. Sumitomo’s high-purity indium comes from China. The entire global InP supply chain is built on Chinese raw materials and Chinese production capacity. The export controls have simply made that dependency explicit and precarious.

The Historical Parallel

This is not the first time China has faced this kind of situation.

In 2011, a news story in China’s financial media reported that the trading price of indium had “surged” from 4.6 yuan to 6.2 yuan per gram within a month. This was unprecedented in China’s domestic indium market. For the previous decade, even though Yunnan had exported 80% of the world’s indium, domestic indium prices had remained as steady as an electrocardiogram. The reason? Japanese companies.

From the 1980s, Japan had been deeply involved in the indium industry but lacked domestic indium resources. China, despite holding most of the world’s indium reserves, had fragmented production capacity and small-scale manufacturers with poor bargaining power. Japanese companies exploited this by buying indium cheaply in China, processing it into components in Japan, and selling it back to Chinese companies at prices dozens of times higher.

That was an early scar on China’s indium industry—and a warning about the value chain. Downstream always has more pricing power and higher profit margins. Controlling upstream resources does not mean you can rest on your laurels; it should be used as a foundation to develop processing and manufacturing capabilities.

In 2011, the Yunnan Pan-Asia Nonferrous Metals Exchange was established, consolidating supply, centralizing pricing, and allowing institutional and individual investors to participate in trading. This brought pricing power back to Chinese producers. That was the beginning of China’s effort to regain its voice in the indium industry.

Today, the optical module boom represents another beginning—a chance for China to transform its upstream resource dominance into downstream manufacturing leadership. Whether it succeeds will determine not just the future of the InP supply chain, but the balance of power in the AI hardware industry for years to come.

Conclusion: A Crisis, an Opportunity, a Test

The InP shortage is a supply chain crisis, a geopolitical flashpoint and an industrial test all rolled into one.

The numbers are clear: demand is surging, supply is constrained, and the gap will not close overnight. The Chinese government has used export controls to leverage its upstream dominance, creating opportunities for domestic manufacturers while imposing severe constraints on overseas competitors.

But the gap between Chinese production capacity and global demand remains enormous. The validation cycles are long. The technical barriers are high. And the geopolitical landscape is unpredictable.

What remains to be seen is whether Chinese manufacturers can translate resource dominance into manufacturing leadership—or whether the bottlenecks will simply move further up the chain. The players that can validate fast, scale fast, and navigate the export control maze will be the ones that win.

For now, the shortage is real, the prices are rising, and the industry is watching—and waiting.

Share

Leave a comment

He does not chase the spotlight. He does not court headlines. And yet Liang Wenfeng—a 42-year-old former math prodigy from a small coastal county in southern China—has become one of the most talked-about figures in the global technology industry.

If you have heard of DeepSeek, you have heard of Liang. If you have not, you will soon enough. His company’s artificial intelligence models have sent shockwaves through Silicon Valley, erased nearly $600 billion from Nvidia’s market capitalization in a single trading day, and forced the world’s largest technology firms to reconsider virtually everything they thought they knew about AI.

But who is the man behind the machine?

Chapter One: A Teacher’s Son

Liang’s father is named Liang Dong. He was a teacher at Meilu Elementary School in the county seat of Wuchuan, Guangdong Province. Over the years, he also held leadership positions in the school’s academic affairs office and general affairs office, and ultimately retired as the school’s director of general affairs. Liang’s mother was a teacher at Bianling Elementary School in Dashanjiang Subdistrict, Wuchuan County.

Although Liang is originally from Mililing Village, Tanba Township, Wuchuan County, he did not actually spend very much time living in the village as a child. He moved to the county seat with his parents at a very young age and grew up in the staff dormitory of Meilu Elementary School. The village was about six miles away, and with poor transportation at the time, he could only return during school holidays to visit his grandparents.

Years later, after DeepSeek exploded into global prominence, Liang himself made every effort to stay out of the public eye—but his hometown took immense pride in him. From the entrance of Mililing Village all the way through, banners and signboards related to him appeared everywhere, including a blue roadside sign that read, “I Miss You at the Home of the DeepSeek Founder.” The road in front of his family home has now been renamed “Number One Scholar Road.” On his grandfather’s four-story house, the windows and walls are plastered with highly conspicuous signs that say “Home of the DeepSeek Founder.”

In reality, however, Liang and his parents spent very little time in that house. As a child, Liang lived in Meilu; after he left for college, he essentially left Wuchuan altogether. His parents also spent most of their time in the county seat. The house was occupied by his grandfather, uncle, and aunt. Liang would still return during the Lunar New Year to visit his grandfather.

About twenty kilometers from Mililing Village, and about ten kilometers from Meilu Elementary School, there is a place called Xiajie in Wuyang Township, Wuchuan County. This was the hometown of Lin Zhaotang, the only imperial examination zhuangyuan (状元) —the highest possible scholarly honor in imperial China—produced by western Guangdong during the Daoguang era of the Qing Dynasty. Because western Guangdong produced only this one *zhuangyuan* in all of Chinese history, his status in the region is extraordinarily revered, and Wuchuan’s Xiajie became known as the “Hometown of the Number One Scholar.”

But now, with Liang’s rise, the title of “Hometown of the Number One Scholar” has gained a formidable challenger. To outsiders, when they hear that Wuchuan is the “Hometown of the Number One Scholar,” most would likely assume it refers to Mililing Village—not Xiajie.

Chapter Two: The Prodigy

Liang started school remarkably early. In 1990, at just five years old—not yet having reached the legal school age—he was admitted early to first grade at Meilu Elementary School. This was possible only because his father was a teacher there and could arrange for his early enrollment. As a result, whether in elementary school, middle school, or university, Liang was always one to two years younger than most of his classmates.

Despite his youth, his academic performance never lagged behind—he was always first in his class. According to a summary compiled by his elementary school teachers (a summary that is now prominently displayed on the wall of Meilu Elementary School), Liang exhibited four defining traits in his studies:

First, he was intensely focused in class. He would listen with rapt attention to every word the teacher said, afraid of missing even a single piece of knowledge. Second, he actively answered questions. Whenever the teacher posed a question, he was always the first to raise his hand. Third, he loved tackling difficult problems and possessed a tenacious spirit of inquiry. Fourth, he excelled at synthesizing learning methods. He could quickly distill the teacher’s lessons and form his own knowledge framework.

Liang, however, had his own summary of what made him successful. He believed his approach to learning rested on two pillars:

First, broad interests. He did not confine himself to textbook knowledge but read extensively outside class, devouring knowledge about science, technology, and general subjects. Second, unique methods. He had his own approach to self-study and the ability to learn independently, rather than simply following the teacher’s lead.

Liang lived by these principles. As a child, he loved participating in various science and technology activities, enjoyed building model ships and airplanes, and especially liked tinkering with electronic devices. As he grew older, whether in academic study, research, investment, AI development or even company management, he always had his own distinctive approach rather than following others.

In the early 1990s, when Liang was in elementary school, Guangdong’s economy was booming at breakneck speed, and money-making opportunities were everywhere. At the time, many parents would come to Liang’s home and tell his parents that “studying is useless”—the local conventional wisdom was that doing business was the only real path to success.

But as teachers, Liang’s parents did not see it that way. They believed in the power of knowledge and in education as a means of changing one’s destiny. They quietly and diligently did their jobs, spending their entire careers as teachers, and always encouraged Liang to study hard.

Compared to the business-minded villagers, Liang’s family circumstances could fairly be described as modest. His father once offered this assessment of his son: “He has been sensible since childhood, very frugal, and a very obedient, good child.” While this evaluation is filled with pride, anyone familiar with Chinese society could also read between the lines—Liang’s childhood was marked by financial constraint.

Frugality and modesty became lasting traits. Even after he had made a great deal of money, he continued to maintain simple living habits.

His father’s other observation: “He has always been confident in his studies, very ambitious, with his own goals and ideals. When he studied, he always aimed to be first.” This confidence likely came from consistently ranking first throughout elementary school.

Chapter Three: The Early Academic Journey

Liang participated in the Mathematical Olympiad and won a national second prize. For an elementary school in a small western Guangdong county, this was almost certainly a school record—at least, from publicly available records, we have not found any better result.

For academically gifted children in Wuchuan during that era, there was another crucial examination: the “June 1st” subject competition, which determined who would be directly admitted to Wuchuan No. 1 Middle School—the county’s best secondary school. Liang won first place in the Meilu Township “June 1st” competition, and subsequently placed within the top forty in the county-wide competition, earning him direct admission.

In middle school, however, Liang’s overall performance was not particularly outstanding. For most of the three years, he hovered around the middle-to-upper tier, ranking roughly around 100th in the entire school. His high school entrance exam score of 819 was not exceptional either—though it cleared Wuchuan No. 1 Middle School’s admission line, it did so by only about twenty points.

Liang’s strength lay in the sciences, especially mathematics. In both seventh and eighth grade, he won first place in the school’s math competitions. This indicated that his intelligence and learning ability were exceptionally strong—it was just that, for various reasons, he had not yet translated that potential into other subjects.

Fortunately, his parents did not put much pressure on him. On the whole, they respected Liang’s own choices and gave him plenty of freedom. In his spare time, Liang enjoyed playing Ping-Pong, soccer and chess, and loved reading extracurricular books. His parents did not intervene even during the period when his grades were less impressive.

In terms of lifelong impact, the most significant development of Liang’s middle school years was likely his first exposure to computers. His interest in computers was not the kind that involved going to internet cafes to play games—it was a genuine thirst for knowledge. He specifically found a teacher to teach him about computers. In a remote county in the late 1990s, this was exceptionally rare.

Liang quickly learned to use computers to look up information, find materials, study English, and even repair computers. According to the later principal of Meilu Elementary School, whenever the school’s computers had problems, Liang would volunteer to fix them.

Chapter Four: The High School Ascent

After entering high school, Liang’s grades began to climb. By his sophomore year, he had risen from around 100th in the grade to the top ten. By his junior year, he was consistently in the top three.

In 2002, Liang took the national college entrance examination and scored 806. Guangdong Province at the time used a standardized scoring system—raw scores were converted based on provincial rankings using a normal distribution table. The maximum possible score was 900. Based on the conversion table of the time, Liang’s 806 placed him roughly between 280th and 320th in the province.

That year’s Guangdong exam format was: Chinese, Mathematics, English, Comprehensive Subjects, and an elective subject (X). To be fair, Liang’s individual subject scores were not exceptionally outstanding across the entire school—except for English, where he scored 141 raw, which converted to 842 standardized, ranking him 48th in the province (tied), 5th in Zhanjiang, and 1st in Wuchuan. His other subject scores were less remarkable:

At Wuchuan No. 1 Middle School alone, many classmates outperformed him in individual subjects. Long Renhai tied his English score (842). Cai Yun and Guo Liangming scored higher in Chinese (833). Li Kun and Deng Wendong scored higher in Comprehensive (Li Kun 842, Deng Wendong 804). Wu Kangzuo scored higher in Physics (811). Ling Da scored higher in Mathematics (834).

But what set Liang apart was that he did not have any weak subjects—he was relatively balanced across the board. So when totals were calculated, he ranked first in both Wuchuan No. 1 Middle School and Wuchuan City.

It is worth noting that many media outlets and self-media later reported that Liang was Zhanjiang City’s top scorer in the college entrance exams. This is incorrect.

Zhanjiang’s top scorers that year were two: Li Wei from Zhanjiang No. 1 Middle School and Zhou Tian from Zhanjiang No. 2 Middle School. Both scored 900, making them two of Guangdong’s seven perfect scores that year. Li Wei chose Physics as his elective subject. After completing his undergraduate and master’s degrees at Tsinghua University’s Computer Science department, he went to the United States for further study and remains there—he is now a major figure in computational biology. Zhou Tian chose English, went to Fudan University’s Law School, and her subsequent whereabouts are unknown.

Liang was actually the top scorer of Wuchuan City—which had been upgraded from a county to a city in 1994. So when local media in 2002 reported that “Liang is the city’s top scorer in the college entrance exams,” they were referring to Wuchuan City, not Zhanjiang.

That year, Guangdong’s admission policy required students to submit their university preferences before taking the exam. Based on his usual grades, Liang had a slim chance of getting into Peking University or Tsinghua University, but would have struggled to secure a desirable major. If he chose Zhejiang University, he could study Electronic Information Engineering—his long-time dream program. So he gave up on Tsinghua and Peking and chose Zhejiang University.

When the final admission cutoffs were announced, Peking University was 807 and Tsinghua was 805. Liang’s 806 fell just short of Peking University. Tsinghua was a bare possibility—but only for its least desirable majors. He chose Zhejiang University. This was a wise decision, and it demonstrated another of Liang’s defining traits: a clear and accurate understanding of his own abilities.

Chapter Five: The Zhejiang University Years

In October 2002, at seventeen, Liang packed his bags, left Wuchuan, and arrived in Hangzhou, beginning eight years of study at Zhejiang University.

The reason he arrived in October was that ZJU’s Zijingang Campus (紫金港校区) was being used for the first time that year. Due to certain delays in construction, the campus was not ready on time, and that year’s freshmen could not check in until October 14.

Before arrival, students admitted to ZJU through early admission or those with outstanding exam scores were required to take the selection exam for the prestigious Chu Kochen Honors College (竺可桢学院). For example, Xu Jin—Liang’s future business partner—was admitted to Chu Kochen’s “Mixed Class” that year. Chu Kochen College is ZJU’s elite program for top freshmen; Pinduoduo (拼多多) founder Huang Zheng is an alumnus. Liang was likely not admitted (the top 15 Guangdong students admitted to ZJU were eligible for the selection; Liang should have qualified), or he may have chosen not to participate.

Liang entered the Department of Information and Electronic Engineering (now the College of Information Science and Electronic Engineering) and was housed in Ziyun Dormitory Building 3. They were the first batch of freshmen to set foot on Zijingang Campus and the first students to live in Ziyun Dormitory.

However, this cohort only stayed at Zijingang for two years and nine months, moving to Yuquan Campus (玉泉校区) in the summer of 2005. Female students were placed in the famous “Princess Building” (Building 8), while male students moved into Building 5, next to the basketball court below Building 8.

By then they were seniors, and the most important task was the graduation thesis. The department’s experimental building—the Information and Electronics Building—sat at the foot of Laohe Mountain beside Lover’s Hill, nicknamed “ZJU’s Potala Palace.” Outside the windows, couples strolled on the grass; squirrels leaped among the ancient trees. Sometimes, if they forgot to close the windows at night, the little creatures would jump inside looking for food.

Zhejiang University is a very special school in China. Former president Zhu Kezhen (Chu Kochen) distilled ZJU’s spirit into two characters—*cheng* (integrity,诚) and *qin* (diligence,勤)—and observed that “outsiders who meet ZJU people always praise the school’s down-to-earth character.” This quality is very apparent in ZJU alumni.

In his 1936 freshman welcome address, President Zhu posed two questions that every ZJU student would ask themselves countless times: “Why have you come to Zhejiang University? And what kind of person do you wish to become after graduation?”

Since that speech, these two questions have become the classic guiding questions for every ZJU student. While they might joke about the “standard answer”—“To goof off, to drift”—in reality, most follow the answers Zhu himself provided at the end of his address: “First, your studies should not be confined to the subjects themselves, but should train you in how to properly train your own thinking. Second, the purpose of our lives is to serve, not to enjoy.”

As a result, ZJU people tend to set high goals while keeping their feet planted firmly on the ground. In academics, this translates to a strong sense of purpose and self-motivated learning. In society, it produces people who get things done and become influential entrepreneurs.

At the same time, ZJU’s atmosphere is remarkably free. Its distance from China’s political and commercial centers allows students to follow their own hearts with less outside influence. The highly flexible “complete credit system” and the four short semesters per year give students ample independent time to pursue what they love. Professors are also tolerant and supportive of students’ various explorations.

All these factors have fostered a rare entrepreneurial culture at ZJU. The number of billionaires, listed company founders, and unicorn founders produced by ZJU ranks among the top in the nation.

Looking only at Liang’s own college, notable alumni include Duan Yongping (段永平)of BBK (步步高), Chen Mingyong (陈明勇) of OPPO, and Zhu Jiangming (朱江明) of Leapmotor (零跑) (originally from the Radio Electronics program at the old Hangzhou University, later merged into ZJU’s Information and Electronics Department). Looking across the entire university, among entrepreneurs born in the 1980s around Liang’s cohort, there is Huang Zheng of Pinduoduo, Chen Qi (陈琪) of Mogujie (蘑菇街), Chen Weixing (陈伟星) of Kuaidi (快的), Huang Xiaohuang (黄晓煌) and Chen Hang (陈航) of Manycore Tech (群核科技), Xu Chi (徐驰) of Xreal, Zhu Qiuguo (朱秋国) of DEEPRobotics.

There’s one small anecdote from Liang’s story to highlight ZJU’s special character: When Liang enrolled, he was in Class 0203 of the Electronic Information Engineering program. His class advisor was the highly respected retired professor Chen Kangsheng (陈抗生).

Professor Chen was already 63 at the time. He had previously served as head of the Information and Electronics Department and director of the university’s Personnel Office. Earlier students like Duan Yongping and Chen Mingyong were all taught by him.

Professor Chen, at over sixty, with decades of teaching experience, in retirement, took on the role of freshman class advisor—not just lecturing on academic subjects, but guiding students on how to adapt to university life, how to balance life and studies, and how to live a meaningful life. This is something rarely seen at any other university in China.

After enrollment, Liang remained characteristically low-key. Thanks to ZJU’s liberal and tolerant atmosphere, he could skip classes he found uninteresting and devote his time to what he truly loved. As a result, his GPA was not particularly outstanding. According to classmates in neighboring classes, he ranked somewhere around the middle-to-upper tier in the department. By the time graduate school admissions came around, his grades were not good enough for a direct recommendation.

So what did Liang do with all that free time? He taught himself digital circuits and analog circuits, and engaged in hands-on engineering projects. At one point, he designed a circuit and programmed a microcontroller to convert an acoustic guitar into an electric guitar that could be controlled by a computer to produce special effects.

This reveals not only that Liang has great ideas and strong practical skills, but also that he plays guitar—anyone who knows nothing about guitars would not know how to tune one.

Converting an acoustic guitar to an electric guitar had nothing to do with exam scores and no commercial purpose whatsoever. It was purely driven by interest. The fact that Liang was willing to spend his time on this shows that he is fundamentally motivated by passion, not utility—a trait that would later guide his entrepreneurial direction and style.

Another very “Liang Wenfeng-style” endeavor was his solo bicycle trip across several eastern provinces. To save money, he often skipped hotels, pitching a tent and sleeping outdoors.

Years later, in 2011, after graduating from graduate school, he did the same thing again—alone, driving a small Fiat, he ventured into the desolate Ali region of Tibet (西藏), exploring deep into the wilderness, and nearly lost his life there.

Though he was eventually rescued by local Tibetans and survived, his beloved Fiat remained forever in that vast, endless plateau.

At the time, he could not have known—or perhaps already knew—that “deep exploration in uninhabited territory” would one day become a metaphor for his life’s work.

It is worth noting that the photographs Liang took during his Tibet trip demonstrate that he is not merely an engineer who spends all his time coding—he possesses a very strong aesthetic sensibility. His photos are remarkable in their framing, composition, and lighting—almost every one could be a wallpaper.

This is the side of Liang that the outside world never knew.

In fact, he has many more hidden dimensions. The “geek introvert” is merely Liang’s surface; a more complete, richer, more accurate portrait reveals a young man who loves solitary adventure, plays soccer, plays guitar, excels at tennis, has high-level photography skills, is full of compassion, and loves life.

Return to his university days.

In his junior year, Liang, along with classmates Qi Diefeng and Liu Wei, participated in the 7th “Sony Cup” National Undergraduate Electronic Design Competition. They chose the topic “Simple Spectrum Analyzer.” Though they did not ultimately win the Sony Cup, they were among the 173 first-prize winners nationwide (out of 4,662 teams from 525 universities).

This first prize carried significant weight for both the students and the school. Liang and his teammates received eligibility for graduate school admission without examination.

However, the awards were announced in mid-October 2005, while ZJU’s graduate recommendation list had already been finalized and submitted to the Ministry of Education in late September. Apart from Liu Wei, who had previously qualified for 2006 graduate admission based on his GPA, both Liang and Qi missed that year’s window and could only be considered for the 2007 admission cycle.

Consequently, Liang and Qi started graduate school a year later than Liu. As it happened, ZJU was reforming its graduate program structure at the time: those entering in 2006 only needed two years to graduate (in 2008), while those entering in 2007 faced a three-year program (for most engineering disciplines). Both Liang and Qi graduated in 2010.

From his graduation in July 2006 to his graduate enrollment in September 2007, Liang had more than a year of free time—likely the most carefree period of a person’s life. All academic work was complete, and no work had yet begun, meaning zero academic or professional pressure. For a 21-year-old, this was pure, unadulterated freedom.

We have little knowledge of how Liang spent those fourteen months—whether he traveled, whether he experienced romance. From fragmentary accounts from classmates, it seems he worked on an electronic sensing system related to marine navigation.

Chapter Six: The Graduate Years and the Birth of a Vision

In 2007, Liang began graduate school under advisor Xiang Zhiyu, focusing his research on machine vision—specifically, target tracking algorithms for PTZ cameras (pan-tilt-zoom cameras, the spherical surveillance cameras commonly seen in security systems).

His thesis tackled a core problem: how to use computers to efficiently process the massive stream of image data generated by thousands of surveillance cameras, detect anomalies in real time, and enable rapid response.

Liang proposed two solutions: first, an automated switching matrix that would route footage from cameras detecting suspicious activity to a monitoring wall, ensuring nothing was missed. Second, each camera would be equipped with intelligent tracking capabilities, eliminating the bottleneck of one operator controlling only one camera, allowing a small team to operate an entire surveillance center.

We will not delve into the technical details here. But for those familiar with DeepSeek’s later innovations—particularly its Multi-head Latent Attention mechanism—a pattern emerges. Though the technical domains are completely different, the underlying philosophy and logic are remarkably consistent: how to allocate limited resources intelligently, sense the full picture while focusing selectively, and synthesize optimal results.

Of course, Liang in 2007-2010 could not have foreseen developments more than a decade away. But in his second year of graduate school, he applied these concepts—machine recognition, intelligent tracking, data analysis, global awareness, selective focus, and optimal computation—to an entirely unexpected domain, with unexpected results. That decision would shape the rest of his life.

Chapter Seven: The First Fortune

In 2008, during his second year of graduate school, Liang developed a keen interest in stock and futures trading. He invested 80,000 yuan in principal and began his journey as an investor.

Liang’s approach was different from most. He did not analyze market sentiment, company fundamentals or rumors. Instead, he used mathematics, probability theory, and computer code to analyze trading data and identify arbitrage opportunities.

Combining the machine vision techniques he had learned during graduate school, he discovered that by using digital image processing to extract data from trading software, he could write his own programs to automate trading—profiting in both bull and bear markets.

Using data, statistics and mathematical models to drive investment decisions, replacing subjective experience, intuition and emotion, and executing trades automatically via computer programs—this approach was already popular in the United States and had become mainstream among hedge funds.

But in China at the time, it was still relatively rare. Even the terminology was not unified: the market called it “quantitative trading,” while regulatory documents referred to it as “programmatic trading.” Eventually, it was unified under the simpler name: quantitative trading.

Liang and his partners—referred to in sources as Mr. S and Mr. J—were among the first in China to pursue quantitative trading. Through this method, Liang quickly earned his first fortune.

Initially, however, Liang did not see investing as his career—it was merely a hobby and a way to earn extra money during his student years. In 2009, he even interned at a senior’s company to get a taste of corporate life.

The company was Aiqi Information (艾麒信息), based in Shanghai, founded by Zhou Chaoen (周朝恩), a ZJU Computer Science alumnus. Though Liang was just an intern, Zhou quickly recognized his abilities and appointed him manager of the New Technology Department, offering a monthly salary of 16,000 yuan—quite generous for an intern in 2009.

At Aiqi, Liang was responsible for R&D in AI video and image technologies. He worked with extreme dedication, losing track of time, spending almost the entire day in front of his computer and rarely stepping out for breaks. His passion for and commitment to technology left a deep impression on Zhou.

For his colleagues, perhaps an even more memorable moment was Liang’s impromptu performance at the company’s annual party. Liang had not prepared anything, but after the host’s repeated invitations, he reluctantly took the stage.

To everyone’s surprise, his performance was not only creative but full of energy—it wowed the audience and instantly electrified the room. When everyone later voted on the best act of the evening, Liang won first place without any competition.

As his graduate graduation approached, Liang did not stay at Aiqi for long and returned to school to write his thesis. Before leaving, he discussed his future plans with his senior. Liang said he wanted to start a business.

Zhou advised him to look for businesses with high gross margins. Liang took this advice to heart; more than a decade later, when he met Zhou again, he specifically thanked him for that advice, saying it was “very valuable and meaningful” and that he had always remembered it.

In June 2010, Liang successfully defended his thesis, graduated, and began his entrepreneurial journey.

He did not jump straight into full-time quantitative trading. Instead, he first went to Chengdu and spent some time trying to find real-world applications for AI technology. He experimented with many scenarios but could not achieve the progress he had hoped for. Eventually, he returned to the financial sector and quantitative trading.

On one hand, his various entrepreneurial ventures faced setbacks; on the other hand, his quantitative trading was becoming increasingly profitable. Ultimately, Liang decided to commit fully to quantitative trading, left Chengdu, and returned to Hangzhou.

There is a telling detail: in 2011, Liang narrowly escaped the Ali Plateau, but his car could not be recovered and remained there forever.

He posted on Weibo at the time:

“Stranded in the uninhabited area for a week. Thankfully I made it out. But my Fiat didn’t. It will stay forever on that endless grassland. That’s the END.”

Looking back years later, that Fiat with the Sichuan license plate served as a metaphor—a symbol of the dreams Liang had once explored but ultimately could not carry all the way, left behind in the wilderness.

Chapter Eight: The Birth of High-Flyer

In 2013, Liang, together with his graduate school colleague Xu Jin (who entered ZJU the same year as Liang in 2002, graduated from Chu Kochen College’s Mixed Class, and started his Ph.D. in 2006, graduating a year ahead of Liang—so from an undergraduate perspective they were classmates, but from a graduate perspective Xu was the senior; he was also one of four classmates Liang thanked in his master’s thesis acknowledgments) and Computer Science alumnus Zheng Dazuo, founded Hangzhou Jacobi Investment Management Co., Ltd. This marked Liang’s transition from a solo quantitative trader to an institutional player.

In 2015, they established Hangzhou High-Flyer Technology Co., Ltd. (later renamed Zhejiang Jiuzhang Asset Management Co., Ltd.), obtained a private equity license, and launched the “High-Flyer Quant” (幻方量化) brand.

According to High-Flyer’s official public account, the name “High-Flyer” is a homophonic translation of “Huan Fang” (幻方), which comes from the ancient Chinese Luoshu (洛书) (Nine-Palace Diagram)—a special matrix that represents the crystallization of science and an auspicious symbol.

Liang hoped that High-Flyer would one day rival the world’s quantitative trading titan—Simons’s Renaissance Technologies.

By this point, Liang had been in the game for seven years. His initial 80,000 yuan had grown through multiple market cycles to over 100 million yuan.

Seven years from 80,000 to over 100 million—even taking the lowest estimate of just over 100 million—represents an astonishing compound annual return of 177%.

Consider this: he had only been out of graduate school for five years. Within five years of leaving academia, he had accumulated over 100 million yuan in cash assets—not company stock, not options, but cash! This wealth creation was nothing short of legendary, demonstrating the power of Liang’s quantitative trading abilities.

But for Liang, this was just the beginning. In the next five years of High-Flyer’s existence, he would become the legendary “head of a 100-billion-yuan quantitative private equity fund.”

The reason for institutionalizing operations was Liang’s belief that China’s quantitative trading had moved beyond the era of “lone warriors” into the age of elite-driven private funds. Therefore, finding the best talent and recruiting them was the most important task.

In a 2015 recruitment post that stirred excitement, High-Flyer summarized the characteristics of the talent they sought:

First, they possess deep mathematical knowledge and programming skills.

Second, they share similar personalities—low-key, steady, disciplined, with regular lifestyles and strict self-discipline.

Third, they develop their trading strategies entirely on their own, from writing to maintenance—each is a lone ranger in the market.

Fourth, their strategies generate excess returns in both bull and bear markets.

On why these people can make money, they argued:

Quantitative trading relies on mathematics, probability theory, and computer code—it is a rigorous science and systematic engineering.

Quantitative methods and strategies remain sensitive topics. Simply copying from books or the internet will not work; successful traders will not give valuable answers because strategy convergence leads to failure.

Given trading costs like commissions and stamp taxes, trading is a negative-sum game. Only a minority can profit—it is a competition among high-IQ individuals, but the winners reap extraordinary returns.

To attract talent, they launched the High-Flyer Quant public account, telling the story of “From 80,000 to 100 Million—The Mysterious and Low-Key World of Quantitative Trading,” and spread it across the bulletin boards of Tsinghua, Zhejiang University, Shanghai Jiao Tong, and other schools.

Beyond academic credentials, candidates had to meet one of three criteria: “Outstanding grades in computer-related courses,” “International Collegiate Programming Contest regional silver medal or higher,” or “Extraordinary talent in a particular area.” The hard requirement: “Must be a high-level programmer.”

During this period of rapid growth, the priority was scale and profit. If Liang (and High-Flyer) had a dream at this stage, it was simply: “To become a world-class quantitative investment firm.”

Consequently, High-Flyer’s recruitment copy was straightforward and frankly all about money:

“The myth of wealth creation in the internet industry continues, and the mysterious, low-key quantitative trading industry will be a surprising dark horse, breeding more and more super-rich.”

This marked a stark contrast with DeepSeek, which would come later. We will return to this when we reach DeepSeek.

Of course, beyond financial incentives, they offered the usual perks: excellent office environment, advanced equipment, comprehensive support—and, playfully, “high-definition photos of goddess-level female colleagues.”

The “goddess” was clearly a lighthearted gimmick to signal the company’s open, inclusive culture and relaxed atmosphere. The combination of financial freedom, good treatment, and appealing culture proved highly successful. High-Flyer quickly assembled a team of top talent from various fields, including some of China’s earliest quantitative traders, Golden Bull Award winners, experts in intelligent robotics, and specialists in internet big data and pattern recognition.

One amusing detail: Liang owned a black coat emblazoned with the words “RESPECT THE OTHERS” in large letters, which he wore around the office—whether to remind himself or his colleagues, no one knew.

With institutional operations and the addition of super-talented members, High-Flyer raced toward its dream of becoming a “world-class quantitative investment firm.”

By the end of 2015, High-Flyer managed nearly 500 million yuan of proprietary capital. In 2016, it began external fundraising, managing 1 billion yuan. In 2017: 3 billion. In 2018: 6 billion.

In August 2019, High-Flyer officially became a “10-billion-yuan private fund,” entering the top 30 private funds nationwide. Since only 4 of those 30 were quantitative funds, High-Flyer also became one of China’s “Big Four Quant Giants.”

Ten billion yuan was already enough to be called a “private equity giant” or “financial whale.” But to Liang, it was merely a waypoint.

In July 2020, High-Flyer’s assets under management exceeded 20 billion yuan. In July 2021, it became one of only five “100-billion-yuan private funds” nationwide—and the first 100-billion-yuan quantitative private fund.

After that, High-Flyer did not continue pursuing growth but voluntarily reduced its AUM (Assets Under Management), at one point dropping to roughly 45 billion yuan, before gradually rising back to 80-90 billion by 2025.

Despite ups and downs, High-Flyer had become China’s most representative and best-known quantitative investment firm. Two indicators demonstrated its and Liang’s industry standing.

First, in 2019, High-Flyer won two Golden Bull Awards—”Three-Year Golden Bull Private Fund Management Company (Relative Value Strategy)” and “Most Trusted Golden Bull Private Fund Management Company.” Liang was invited to the Golden Bull Awards ceremony and Private Fund Summit Forum.

After the morning’s ceremonial proceedings, the afternoon featured three “Golden Bull Sharing Sessions” of the highest value. The forum on quantitative investing—”Quantitative Investing and Financial Technology”—featured Liang as the most prominent speaker, delivering the first address. His talk, “New Trends in A-Share Quantitative Investing,” became the most widely circulated speech from that year’s Golden Bull Awards (retitled in circulation as “The Future of Quantitative Investing in China, Through a Programmer’s Eyes”).

Second, in 2021, when the biography of “quant trading godfather” James Simons—The Man Who Solved the Market—was introduced to China, the publisher invited three people to write forewords: first, economist Professor Ba Shusong (巴曙松); second, Liang; third, Qiu Huiming (裘慧明) of Minghong Investment (明汯投资).

The most widely quoted passage came from Liang’s foreword, “Creating a Model for Understanding the Market”—especially the closing line: “Whenever I encounter difficulties in my work, I think of Simons’s words: ‘There must be a way to model prices.’” This became Liang’s most famous quote and the best recommendation for Simons’s biography.

But despite High-Flyer becoming China’s number one, despite Liang amassing immense wealth through it, despite his becoming the undisputed face of quantitative investing in China—for Liang personally, this was merely the overture to his life’s work.

His true mission would begin at High-Flyer’s peak. But it would not be quantitative investing. It would be AI.

Chapter Nine: The Crossroads

While High-Flyer was charging ahead at breakneck speed, Liang had already seen through the fog and glimpsed the future of quantitative investing.

That future was not optimistic.

Liang’s 2019 Golden Bull Award speech had already laid out his views with considerable clarity. Broadly speaking, his predictions encompassed both short-term and long-term horizons.

In the short term, quantitative private funds could grow enormously—there was still room for dozens of times expansion (relative to the 10-20 billion yuan base of China’s top quantitative firms at the time), and they could eventually reach a scale of over 1 trillion yuan.

But in the long term, the ultimate trajectory of quantitative investing was this: the average returns of quantitative strategies would converge toward the market average return, eventually declining to the point where excess returns—what the industry calls “alpha”—would fall to zero.

Why would such a trend emerge?

Liang believed that quantitative investing broadly follows Moore’s Law—roughly every 18 months, investment capability doubles. But on the other hand, market efficiency is constantly improving (to a significant degree driven by quantitative trading itself), which causes the marginal returns on increased capability to diminish. In other words, even if investment capability doubles, the profits generated do not double.

This means that as market efficiency increases, the gap between what quantitative strategies earn and what the average market participant earns narrows, and eventually that gap converges to zero.

When the market becomes fully efficient, all quantitative strategies fail. Index investing becomes the only truly valuable approach—everyone simply buys the index. Wealth ends up in the hands of ordinary people.

Liang did not say this explicitly, but we can reasonably infer: by that point, the quantitative investing industry would gradually fade away, much like the technical analysis school of trading has today.

The last and most critical question remaining was this: from market inefficiency to full efficiency—how long would that take?

In his speech, Liang made three judgments:

1. “In the coming years, the efficiency of China’s stock market will further increase. This is an irreversible historical trend.”

2. “We are still far from a fully efficient market—at least for the next few years, we don’t need to worry about this.”

3. “Fundamental quantitative investing… may take several more Moore’s Law cycles to mature. But the direction is certain.”

There are three time-related phrases here: “coming years,” “still far,” and “several more Moore’s Law cycles.” One Moore cycle is 18 months, or 1.5 years. “Several Moore cycles”—if we take three, that’s 4.5 years; if nine, that’s 13.5 years.

Regardless of the exact number, what is certain is that Liang and his team had seen this inevitable trend. And they already knew, with clarity, that quantitative investing’s golden days were numbered.

In fact, High-Flyer was already feeling the chill of the market.

In 2019, High-Flyer experienced its largest drawdown. At the start of 2020, in a letter to investors issued by High-Flyer Quant CEO Lu Zhengzhe (陆政哲), there was an explicit mention of the need to “strive to explore the frontiers of technology and mine the deep alpha in the financial markets”—this statement alone fully demonstrates that the team had already realized there was no longer enough alpha to be found in the shallow layers of the financial markets; they had to go deeper.

Since Liang and his team had all seen this trend, the next logical step was that they had to confront two soul-searching questions:

1. Is the quantitative investing industry still worth dedicating my entire life to?

2. If quantitative investing is not the future, then in which direction should my future go?

I believe these were the questions Liang thought about most, and most deeply, around 2018.

Fortunately, the direction of the future had already revealed itself naturally through High-Flyer’s own development.

Chapter Ten: AI First

High-Flyer’s initial quantitative trading models used linear models and traditional machine learning algorithms. During the era of rapid market growth, these models had supported the compound annual returns of well over 100% that Liang achieved in his early personal investing and that High-Flyer achieved in its early institutional investing.

But as industry competition intensified and market efficiency improved, the limitations of these models became increasingly apparent. If the models weren’t fast enough or accurate enough, High-Flyer couldn’t compete.

So in October 2016, High-Flyer underwent a critical transformation: it upgraded its quantitative trading models from linear models and traditional machine learning algorithms to deep learning algorithm models (DL models), and upgraded the processors used for model computation from CPUs to GPUs.

Since all of High-Flyer’s models were developed in-house, this transformation meant more than just giving High-Flyer’s quantitative trading more powerful computational tools. The more important and far-reaching implication was this: the adoption of deep learning models inherently contained a path of self-development—as long as you gave a deep learning model sufficiently good algorithms, sufficiently powerful computing power, and sufficiently abundant effective learning resources, it was highly likely to evolve into a powerful AI.

In 2016, Liang probably hadn’t thought that far ahead when he adopted deep learning models—he was probably just thinking about having a more effective tool for quantitative trading. But this seed, planted almost inadvertently, would in the not-too-distant future grow into a towering tree that would capture the world’s attention.

In 2017, High-Flyer continued to expand its AI algorithm research team and its AI hardware and software R&D team. By the end of the year, nearly all of High-Flyer’s quantitative strategies were being computed using AI models. In external communications, High-Flyer explicitly stated: “Investing has entered the AI era.”

In 2018, High-Flyer’s AI development encountered some difficulties: the complex computational demands of the models meant that single-machine training hit a computing power bottleneck; at the same time, the growing training demands were coming into conflict with limited computing resources. Both problems required large-scale computing power solutions.

This brought the company to a critical crossroads: how should the relationship between AI and quantitative trading be structured?

If quantitative trading was primary and AI was secondary—AI serving quantitative trading—then as the cost of AI continued to rise, the company would have to weigh trade-offs and decide whether it was necessary to invest more in AI. The key consideration would be whether the incremental investment returns generated by improved AI capabilities were sufficient to cover the costs of that investment.

In other words, the envisioned “large-scale computing power solution” would be constrained by the “return-cost” framework of quantitative investing, with “maximizing quantitative investment returns” as the key decision metric.

But if AI were freed from the constraint of “serving quantitative trading,” and instead became the development goal itself—or even the reverse, with AI as primary and quantitative trading as secondary, with quantitative trading serving AI—then the decision-making logic would change entirely.

Not only would it break free from the quantitative investing “return-cost” framework, but all of the profits from quantitative trading could be burned for the sake of AI development itself, and if that wasn’t enough, external funding could be sought.

Earlier, this question hadn’t needed to be considered, because the returns from model upgrades far exceeded the costs. But by 2018, the tension between “AI development” and “quantitative returns” had become extremely acute. Liang and his partners had to make a decision on the most important question:

Does AI serve quantitative trading, or does quantitative trading serve AI?

At its deepest level, this question was: Are we continuing to be a quantitative firm, or are we transforming into an AI company?

Or: Are we a quantitative firm, or are we an AI company?

The question circled back to President Zhu Kezhen’s classic two questions: “What are we going to do? What kind of people are we going to become?”

From the official statements on High-Flyer’s website, we can see that Liang and his team answered these two questions with action that year.

The “High-Flyer Milestones” section’s description of 2018 has two sentences: the first is “Won the Golden Bull Award”; the second is “Established AI as the company’s primary development direction.”

The first sentence is a record of past honors; the second is a description of the company’s future.

If future historians were to describe the history of DeepSeek’s development, this would be the moment of “the Dao (道) generating the One”—the moment of “the Wuji (无极) generating the Taiji (太极)”—the moment of “Let there be light.”

Once this most consequential decision was made, everything became different.

Not only did this company from then on become fundamentally different from every other quantitative firm in the world—more importantly, what this company was about to do would make Liang and his team into completely different people from who they had been before, and would make the world into a completely different world.

Before, it was just about making money, plus enjoying the intellectual superiority. Afterward, it was about using their own creations to change the entire world and create a new future for humanity.

Chapter Eleven: The Firefly Project

If 2018 was the year the decision was made, then 2019 was the critical year when High-Flyer’s AI ambitions officially took shape.

Three months after Liang’s Golden Bull Award speech, in December 2019, High-Flyer formally incorporated High-Flyer AI (High-Flyer Artificial Intelligence Basic Research Co., Ltd.), with the explicit goal of “committing to research in AI algorithms and foundational applications,” and launched the “Firefly I” AI supercomputing cluster project.

“Firefly I” was initially designed to be equipped with 500 graphics cards, but when it was actually completed in early 2020, it was expanded to 1,100. It occupied an area roughly the size of a basketball court and could perform 184 quadrillion floating-point operations per second. It supported ultra-large-scale model training, dramatically boosting High-Flyer’s AI capabilities—a large model that previously would have taken two months to train could now be completed in just three and a half days on “Firefly I.”

The total investment in “Firefly I” was close to 200 million yuan, all from High-Flyer’s own funds. This did not even include the ongoing maintenance and operational costs. At the time, there was a saying circulating online: “High-Flyer’s annual electricity bill alone is more than the total annual operating expenses of other private funds.”

High-Flyer was the first quantitative firm in China to invest this much money in AI.

In fact, the number of companies in the entire Chinese market at the time that could do large-model training at the thousand-card scale could be counted on one hand—and they were mostly a handful of internet giants or native AI startups like SenseTime. High-Flyer was the only fintech company among them.

Compared to the internet giants and native AI startups, High-Flyer AI’s path was far more difficult.

At the time, the entire team were complete novices when it came to supercomputing—they didn’t even know how to cool a supercomputer, and had to figure everything out on their own. This work was at the very frontier of global exploration, with almost no reference materials available. Even if there was some industry “know-how” out there, why would anyone be kind enough to share it? So almost every problem had to be solved by designing their own solutions and putting them into practice.

Making matters worse, the end of 2019 and the beginning of 2020 coincided with the outbreak of the COVID-19 pandemic, which caused significant disruptions to personnel movement and material transport.

But High-Flyer’s team managed, against all odds, to complete and launch “Firefly I” by March 2020.

What was especially remarkable was that the High-Flyer team did not lower their standards just because the work was difficult. On the contrary, they aimed from the very beginning for the industry’s limit—not the limit of the quantitative investing industry, but the limit of the AI industry. They used the most powerful hardware and the most aggressive solutions on the market. Though they encountered many problems along the way, they ultimately found solutions to all of them.

Through this willingness to break through constraints and challenge limits, High-Flyer forged a team that was technically formidable and fearless. The team that would later build DeepSeek grew out of this foundation.

The 200 million yuan investment and 1,100 graphics cards in 2020 were already enough to astonish the entire industry at the time. But just a year and a half later, the world would truly see the extent of Liang’s ambition.

In 2021, High-Flyer launched “Firefly II” The investment in “Firefly II” reached 1 billion yuan and occupied an area roughly the size of ten basketball courts.

According to a paper later uploaded to the arXiv platform by the DeepSeek team under the username Wenfeng Liang:

“For deep learning training, we deployed the Firefly II system equipped with 10,000 PCIe A100 GPUs. Its performance approaches that of NVIDIA’s DGX-A100 AI supercomputer, while achieving a 50% cost reduction and 40% lower energy consumption.”

In 2021, deploying 10,000 A100 graphics cards for a self-built AI training cluster—not to mention that no other quantitative firm in the world had done anything close to this—even among global technology companies and AI firms, such a feat was rare.

Later, Yu Lili, a reporter from 36Kr’s “Dark Waves” column, asked Liang why he spent so much money on this. His answer was telling:

“An exciting endeavor perhaps shouldn’t be measured purely in monetary terms. It’s like buying a piano for your home—first, because you can afford it, and second, because there are people eager to play music on it.”

When a piano is bought because people are eager to play music on it—not for taking exams, not as a decoration—that is when it can truly fulfill its greatest purpose.

You might even have bought the piano originally just to play Für Elise. But once you have it, with your passion and your daily, joyful practice, you end up playing those profound, powerful, soul-stirring masterpieces by Bach and Beethoven.

Liang and his partners had originally built their deep learning AI training model to do quantitative investing better. But once that AI model was built, what it could do extended far beyond quantitative investing.

Before long, the world would be amazed by the beautiful music they played on that piano.

Chapter Twelve: The ChatGPT Moment

In 2022, human history experienced a watershed moment.

At 3:38 AM Beijing time on December 1—11:38 AM PST on November 30—a Twitter user named Sam Altman posted a brief message: “today we launched ChatGPT. try talking with it here: chat.openai.com.”

When people clicked through the link, they saw a welcome message that read “Welcome to ChatGPT.” After registering and logging in as prompted, they entered a chat window. They typed some text, and the page returned some content in response.

Since the internet existed, humans had done this billions of times. But this time, everyone felt something different, because what this webpage returned was unlike anything before.

It could answer your questions. It could help you find all kinds of information. It could write poems and essays for you. It could write code for you. It could even—with something approaching emotion—chat with you.

It seemed to have… more intelligence?!

This product, called ChatGPT-3.5, attracted one million users in five days, and reached 100 million monthly active users within two months—becoming the fastest-growing consumer application in human history.

Later, people would call this moment the “ChatGPT moment.”

It may have been one of the most important moments in human history, because it pulled back the curtain on humanity’s formal entry into the AI era.

And High-Flyer in 2022 remained in silence.

The biggest news from High-Flyer that year was its charitable donations.

The “High-Flyer Milestones” section of the company’s website recorded only two major events for 2022: “守望相助” (Helping Each Other) and “突破极限” (Breaking Limits).

“Helping Each Other” referred to the charity work High-Flyer did that year. According to the website: in 2022, High-Flyer donated a total of 221.38 million yuan to charitable organizations, and an employee identified as “一只平凡的小猪” (An Ordinary Little Pig) personally donated 138 million yuan, supporting 23 public welfare projects across 15 charitable organizations.

It is particularly worth noting that the 221.38 million and the 138 million are counted separately and are parallel items—the latter is not part of the former. That is to say, High-Flyer plus its employees donated at least 359.38 million yuan that year.

This staggering figure sparked widespread discussion online.

At the time, High-Flyer was still seen by the public as a “quantitative private fund,” and the dominant tone of media and self-media coverage was: “Mysterious private equity giant donates billions.”

The individual donation by “An Ordinary Little Pig” became a trending search topic under the headline “Private equity employee donates 138 million.” This was likely the first time High-Flyer Quant had become so publicly known (previously, news about High-Flyer had mainly circulated within investment circles).

Because “An Ordinary Little Pig” donated anonymously, the donor’s true identity sparked much speculation. High-Flyer itself never revealed it—the official line was, “All donations are anonymous, so we don’t know who it was either.”

In fact, High-Flyer’s enthusiasm for charity had been baked into its DNA from the very beginning. For example, the recruitment copy published when the company was first founded mentioned:

“Mr. S, jokingly called a ‘money-printing machine’ by his friends, is also a leading figure in the quantitative trading world… Like Buffett, he devotes a portion of his returns to charitable causes.”

Later, High-Flyer established a dedicated “Public Welfare Working Group,” with Yin Mingze, who had previously been deeply involved in charitable work, serving as project coordinator.

High-Flyer’s philanthropic efforts had no CSR targets and sought no publicity—it was simply a group of down-to-earth people doing down-to-earth things. The projects they prioritized most were those related to education, development, and critical illness support for children and women.

The other major event of 2022—”Breaking Limits”—was described as follows:

“Firefly II achieved a breakthrough in hardware and software architecture through the interconnection of over 800 switches plus core expansion sub-trees, breaking through the physical limitations of Phase One and doubling its computing capacity. The new hfai framework accelerated models by 50-100%. The cluster ran continuously at full capacity, with average utilization above 96%. Over the year, it ran 1.35 million tasks, totaling 56.74 million GPU-hours. Idle computing power used for research support reached 15.33 million GPU-hours, accounting for 27%.”

A series of technical terms and metrics, presented with objective description and no grand proclamations. There was no hint that a major breakthrough—one that would upend the industry and shake the world—was about to happen.

Meanwhile, the world was in tumult.

Shortly after ChatGPT-3.5 was released, Anthropic released Claude, Google released Gemini, Elon Musk founded xAI and launched Grok… these large models began influencing people’s work and lives at an astonishing pace.

Compared to all this, the two major events that High-Flyer had solemnly listed—corporate charity and the Firefly II upgrade—seemed almost trivial.

But I believe it was precisely because High-Flyer treated these two things as the company’s most important events that DeepSeek’s meteoric rise was ultimately made possible—and that, in the future AGI landscape, Liang and his team are destined to be among its most important architects.

These two major events:

Charity represents love, emotion, compassion—the human soul.

The Firefly upgrade represents curiosity, technology, the desire to explore—the source of human evolution.

They involve not just the 359.38 million yuan in donations, 23 beneficiary projects, and an upgrade to a technological tool—they clearly symbolize something deeper: *What are we doing AI for? How should we do it?*

Many people think about this question too simplistically. Even many companies that currently hold advantageous positions in the industry’s competitive landscape have not thought it through clearly—or have never thought about it at all.

AI is not simply technological progress or productivity enhancement. AI is the first time in human history that we are creating new intelligence—and intelligence that might surpass our own. AI is the entity to which we will soon collectively entrust the entirety of humanity’s future.

Shortly after ChatGPT’s release, Musk commented on Twitter: “We are not far from dangerously strong AI.”

He believed that powerful AI has the potential to destroy civilization. Within ten years, humans may no longer be able to control the world. Because “if the intelligence gap between AI and humans is far greater than the gap between humans and chimpanzees, it’s hard to imagine chimpanzees still being in control.”

“AI godfather” Geoffrey Hinton continued to warn people about AI’s enormous risks, saying: “Humans have never had to deal with anything smarter than themselves. There are almost no real-world examples of a less intelligent species controlling a more intelligent one.”

Hinton compared developing AI to raising a tiger—cute when young, but it might eat you when it grows up. And unlike a tiger, AI is smarter than humans. The idea that we can simply turn off super-intelligent AI at any time is naive.

Yoshua Bengio, another of the “deep learning triumvirate,” OpenAI’s former chief scientist Ilya Sutskever, DeepMind founder Demis Hassabis share largely the same view.

This raises a question of immense significance for humanity: If AI is so dangerous, should we still develop it? And how should we develop it?

The question of “whether to develop AI” no longer needs an answer—it is already an unstoppable historical tide that cannot be reversed by anyone’s capability, will, or action.

The question of “how to develop AI” is the real, meaningful, and civilization-defining issue.

This is not a philosophical question, but a real-world question about how those who have the power to determine humanity’s fate will actually act.

The most important determining factor is this: to what extent will those who are creating and developing the AI that will one day take over humanity’s future consider “AI for good”? How much weight does “AI for good” carry when they are determining technical roadmaps, business models, and even in the corporate processes of recruiting, daily management, product positioning, and marketing?

Chapter Thirteen: The Turning Tide

In 2023, the United States fully led the AI wave. ChatGPT-4.0, Claude, Gemini, and other state-of-the-art AI large models were released one after another, and the pace of AI evolution was breathtaking.

In China, the venture capital and startup circles—which had not seen any “super sexy” projects for a long time—were jolted by ChatGPT’s sudden emergence. Everyone knew that a new era had arrived, and new opportunities were here.

Wang Huiwen, Wang Xiaochuan, Kai-Fu Lee all announced their entry into AI. Baidu launched ERNIE Bot, Alibaba launched Tongyi Qianwen, Tencent launched Hunyuan, iFlytek launched Spark, and Moonshot AI launched Kimi…

But behind all the apparent hustle and bustle, there was still, for many Chinese, a lingering anxiety or worry: *America is still ahead of us!* If AI is the future, then in the China-U.S. competition, what are our chances? Can we even compete?

Although domestic large models were developing rapidly, compared to ChatGPT-4.0, Claude, and Gemini, they were still far behind—and the gap was even widening.

The so-called “hundred-model war” contenders could only show off in the Chinese market. On the international stage, aside from occasionally topping a single metric in open-source model evaluations, they were comprehensively behind in general benchmarks.

Even the most advanced Chinese large models could only barely approach the level of ChatGPT-3.5, still far from 4.0. The “generational gap” between Chinese and American AI was real and palpable.

This was also reflected in the consumer market.

In 2023, any user who could access ChatGPT would not choose to use domestic large models, because the quality gap was simply too vast.

By 2024, with Kimi’s upgrades, and ByteDance and Tencent launching Doubao and Yuanbao, China’s consumer-facing AI made significant strides. But OpenAI’s combination of multimodal capabilities (GPT-4o), deep reasoning (o1/o3), and world simulation (Sora), along with Google, Anthropic, and Meta’s close pursuit, plunged the domestic industry back into a dual state of “excitement + despair”—excitement that AI could do all this, and despair that we were falling even further behind.

What was Liang and High-Flyer doing during this time?

In April 2023, High-Flyer Quant’s official public account published an article titled “A New Journey for High-Flyer,” announcing:

High-Flyer will concentrate its resources and efforts to fully commit to artificial intelligence technology that serves the common interests of all humanity, establish a new independent research organization, and explore the essence of AGI.

We will invest fully and continuously, avoid doing mediocre things, and use the longest-term vision to answer the biggest questions.

At the beginning of this manifesto, a quote from the famous French filmmaker François Truffaut was used as an epigraph:

“You must be wildly ambitious, and at the same time, wildly sincere.”

Not a single sentence in the entire article mentioned business. Not a single sentence mentioned profit. Instead, it emphasized “serving the common interests of all humanity” and “using the longest-term vision to answer the biggest questions.”

On July 17, 2023, Liang officially registered “Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.” and used the English translation of “DeepSeek” (深度求索—deep exploration) as the name for the AI large model he was developing.

On November 2, 2023, DeepSeek released its first open-source code large model, DeepSeek Coder—free for commercial use and completely open-source. Its official public account published its first post, claiming the model “might be the strongest open-source code large model!” High-Flyer Quant’s public account reposted it simultaneously.

On the internationally authoritative HumanEval programming multi-language test, DeepSeek Coder outperformed existing open-source models in every language—including Meta’s CodeLlama, which had previously led the pack. This attracted significant attention in major international developer communities like HuggingFace and GitHub. Overseas tech media and AI bloggers began reviewing it, and domestic tech communities followed up with coverage.

However, because this model was primarily for coding, it only gained traction within technical circles and was not known to the general public. In China, only a few financial media outlets reported on it from a curiosity angle, with headlines like “Quantitative Giant Releases Large Model.”

But it had already left a deep impression on developers both at home and abroad: this large model from China was completely open-source, free for commercial use, and highly capable! Many developers abandoned their previous paid commercial code large models and switched to DeepSeek Coder, exclaiming, “This is amazing!”

On November 29, 2023, DeepSeek launched DeepSeek LLM 67B, its first publicly released general-purpose large language model. The performance of LLM 67B was already approaching that of GPT-4.0, and in nearly 20 Chinese and English public benchmark rankings, it surpassed Meta’s LLaMA2 70B, which had been the benchmark for open-source models.

But the most striking thing was this: a product with performance approaching that of top-tier closed-source models was actually completely open-source, required no application for access, and was free for commercial use!

DeepSeek was the first in the world to push both “high performance” and “open-source friendliness” to such extremes.

But at this point, DeepSeek had only made a deep impression on technical circles and a handful of industry observers. The general public was still largely unaware that China had such a powerful large model.

Moreover, these earlier products had merely announced its existence and attracted a few glances—they had not yet had a substantive impact on the industry.

At this point, DeepSeek was still a minor player—promising, but still dispensable.

Chapter Fourteen: The Tipping Point

On May 6, 2024, DeepSeek finally began to become a major force in the industry—that day, it released DeepSeek V2. This was DeepSeek’s first product that truly shook the industry and deeply impacted the competitive landscape.

V2, with its Multi-head Latent Attention (MLA) and Mixture of Experts (MoE) architecture, achieved performance on par with GPT-4, but its API pricing was only 1% of GPT-4’s—widely described as “ridiculously cheap.”

V2’s high performance and low price triggered an earthquake in the industry. That month, domestic large models were forced to slash prices across the board. Alibaba cut prices by 97%, Zhipu AI by 80%, ByteDance’s newly launched Doubao general model was priced 99.3% cheaper than the industry average, and Tencent, Baidu, and iFlytek’s flagship models were directly announced as free.

DeepSeek was thus dubbed the “price butcher” and the “Pinduoduo of the AI world.”

Internationally, the technical community showered V2 with praise—at levels unprecedented for a Chinese large model. “Possibly the best paper of the year.” “Full of astonishing wisdom.” “DeepSeek has hired a bunch of unfathomable geniuses.” “Chinese-made large models will be as impossible to ignore as drones and electric vehicles.” These were not casual comments from forum users—they came from recognized authorities and top technical experts in the field.

Unlike DeepSeek’s previous models, which had only attracted attention in developer communities, V2 directly became the hot topic of the week across many forums. Twitter saw a flood of reviews and discussions. Many domestic tech media outlets also began reporting on it, and it even made it onto Zhihu’s trending list.

Overall, V2 became a sensation in technical circles, gained a large number of users and a strong reputation in the B2B space, and through V2, DeepSeek truly established itself in the industry.

If previously many people had, to some extent, regarded DeepSeek as just a quantitative investment firm’s side project, then after V2, everyone knew that DeepSeek had become a global AI player capable of shaking the entire market, deeply influencing the industry landscape, impossible to ignore, and able to represent Chinese large-model technology and commercial potential on the international stage.

But for DeepSeek, this was still just a small step.

It would take another seven and a half months for the general public to truly know its name.

By then, the entire world would be shaken.

Chapter Fifteen: The DeepSeek Moment

On January 27, 2025, the U.S. stock market opened after the weekend. Nvidia, the market’s benchmark, saw its stock price plunge from the opening bell. By the closing bell, the company’s market capitalization had fallen 16.97%—Nvidia’s largest single-day drop since 2019. The $589 billion wiped off its market value was also the largest single-day market cap loss for any single stock in U.S. history, a record that still stands today.

Nvidia’s stock crash had nothing to do with its own operations. It was caused entirely by a development from across the Pacific: DeepSeek had released its R1 large model a week earlier.

Within just a few days, R1 had ignited the entire U.S. tech scene. By the weekend of January 25-26, even Twitter (X) was overrun—one user described it as “nine out of ten posts are discussing DeepSeek,” a level of buzz that even ChatGPT-3.5’s launch hadn’t achieved.

Not only that, but before the U.S. markets opened on January 27, R1 had already reached number one on both the U.S. and China App Stores. Shortly after, it became the number one downloaded app in 140 countries and regions.

The most important financial media in the U.S. and Europe all followed up with coverage.

CNBC was first: A little-known Chinese AI lab has released an AI that costs less but outperforms America’s best models, sending panic through Silicon Valley.

The Wall Street Journal reported: Chinese AI model shakes global tech markets, arguing that it forces investors to reassess the core assumptions underpinning trillions of dollars in AI capital spending.

Bloomberg argued that with R1’s emergence, “the entire AI investment thesis needs to be re-examined.”

The Financial Times of the U.K. stated: R1 represents a disruptive inflection point. For the first time, an open-weight model, with extremely low training costs, can match top-tier closed-source reasoning systems.

The Economist argued: R1’s real revolution is AI democratization… reshaping the global geography of AI innovation.

The New York Times thought even bigger, framing it from the perspective of geopolitics and AI sovereignty: This Chinese company open-sourcing a powerful reasoning model changes the balance of power: Western developers may increasingly build on model weights originating from China, reshaping the global landscape of AI progress. It also argued that this outcome exposes the limitations of U.S. semiconductor export controls, while raising troubling questions: who will set the rules for future global AI innovation?

While the media focused more on DeepSeek’s impact on markets, geopolitics, and national competition, the tech community paid more attention to the technology itself—its power and its disruptive nature:

Marc Andreessen, Silicon Valley’s most influential venture capitalist, said: DeepSeek-R1 is one of the most amazing and stunning breakthroughs I’ve ever seen. Their decision to open-source the weights is a profound gift to the world. This is AI’s Sputnik moment.

“Sputnik” was the name of the first artificial satellite launched into space—and because it was the Soviets who launched it first, it caused immense shock in the United States. Americans felt that America’s advantage in space had suddenly been seized by the Soviets. Afterward, the U.S. threw itself into a race with unprecedented urgency and massive investment to reclaim its space superiority.

Since then, “Sputnik moment” has been used to describe inflection points where a competitor suddenly produces a disruptive breakthrough, forcing a country to reassess its advantages and accelerate its efforts. After Andreessen used this metaphor for R1, “Sputnik moment” became one of the most frequently used phrases in American media whenever DeepSeek was mentioned.

Elon Musk commented: “For most of human history, China was the most powerful country on Earth. So you can expect them to do many great things—DeepSeek is one of them. This is the result of China having a large number of outstanding talents, and it’s very impressive.”

Microsoft CEO Satya Nadella said: DeepSeek’s new model is impressive, especially in terms of how efficiently they develop open-source models and perform inference. Their computational efficiency is outstanding. We should take these technological developments from China very, very seriously.

Meta’s Chief AI Scientist and Turing Award winner Yann LeCun argued that this is not China surpassing the U.S. in AI, but rather open-source models surpassing closed-source models.

We need not cite more—the above comments from media and tech leaders are more than enough to describe the shock that DeepSeek R1 caused in the minds of Americans. This was the first time a Chinese large model had caused such a psychological impact on American media, the tech community, and ordinary people.

In China, recognition of DeepSeek came a little earlier.

On January 20, 2025—the same day R1 was released—Liang was invited to attend the State Council Premier’s Symposium, and appeared on that evening’s *Xinwen Lianbo* (China’s national evening news broadcast).

We need not say much more about this. Anyone with even a basic understanding of China knows what this means, and what it says about how the nation views Liang and DeepSeek.

And this was on the very day R1 was released—before its full impact had been felt and before the media had extensively reported on it. It shows that the state recognized Liang and DeepSeek’s importance earlier than the public did.

For ordinary Chinese users, starting from January 20—first, due to the Premier’s Symposium and Xinwen Lianbo (新闻联播); second, because of the foreign tech and media reactions being relayed back to China; and third, because of the sheer surprise and delight DeepSeek’s product brought to users—Liang finally began to enter the public consciousness, sparking a phenomenal wave of online frenzy.

During the Lunar New Year holiday, DeepSeek flooded virtually every information platform and social media platform we can think of. DeepSeek’s website traffic and mobile app downloads surged by hundreds of millions, and its mobile app’s monthly active users quickly reached tens of millions. Even Mililing Village—Liang’s hometown—became a popular tourist destination.

Compared to other AI products on the market at the time, DeepSeek’s performance brought immense surprise. It could even be said that many Chinese users, until they used DeepSeek, had never truly experienced an AI product whose quality approached or matched international first-tier models like ChatGPT.

Previously, only a tiny minority of domestic users had access to foreign products like ChatGPT; the vast majority could use domestic AI products, but the quality of what those AI produced was far inferior to ChatGPT. DeepSeek was the first time many people felt: *China’s large models aren’t bad either—they can compete with the world’s best.*

Later, just as the release of ChatGPT-3.5 had been called the “ChatGPT moment,” people also called the arrival of DeepSeek-R1 the “DeepSeek moment.”

Chapter Fifteen: A Nation-Defining Product

The impact of the DeepSeek moment on China—and on the world—may be more profound than we imagine.

Before this, U.S. AI had raced far ahead globally, and China was scrambling to catch up. Though there had been sporadic moments of brilliance (such as certain Chinese models occasionally ranking high in particular benchmarks, or a particular company’s product receiving brief acclaim), as a whole, China still lagged far behind the U.S. Many even believed there was a clear generational gap between China and the U.S., and that the gap was widening.

At its core, a nation’s development is driven by technological revolutions. When a technological revolution occurs, the countries that catch the wave develop; those that miss it fall behind.

When the Industrial Revolution and the Electrical Revolution occurred, China missed out, fell behind, and endured a century of humiliation. It was only after the founding of New China and the reforms and opening-up that we caught up on these two waves.

By the Computer Revolution and the Internet Revolution, China was somewhat behind but gradually caught up. By the Mobile Internet Revolution and the New Energy Revolution, China was not far behind the U.S. in these two fields, and in some areas was neck and neck.

In contrast, Japan—which had industrialized by capitalizing on the Industrial and Electrical Revolutions—failed to seize opportunities in the Computer, Internet, Mobile Internet, and New Energy revolutions. Japan’s economy peaked in the mid-1990s and has since stagnated or declined.

This time, the newest and most significant technological revolution is AI. And the AI revolution is fundamentally different from previous technological revolutions—AI can learn by itself, develop by itself, improve by itself, and can deeply affect every industry related to “intelligence.”

The higher the technical level of one generation of AI, the more efficiently it can create the next generation of AI, and the more deeply it can transform other products and industries. After a few rounds, the Matthew Effect reaches its extreme—”winner takes all” dominates the industry and the global economic order.

Therefore, AI development is not just a matter of which company can capture how much market share, how much profit, or how much market capitalization from an enterprise or industry perspective. It concerns a nation’s future economic development, industrial layout, high-tech competition, and even military and political dynamics.

To put it bluntly, AI is a major factor in national destiny. If China loses in the AI competition, it would not just be a matter of Chinese tech companies’ financial reports lagging behind American ones, or Chinese AI practitioners earning less than their American counterparts. It wouldn’t even just be about whether China would fall into a “lost thirty years” like Japan. It would be about whether we would be permanently suppressed and exploited in global power competition.

From November 2022 to December 2024, the scales of national destiny—with AI as the core influencing factor—swung violently, gradually tilting toward the U.S., and China faced extreme peril. During this period, China desperately needed an AI that could truly compete, to shoulder the burden of reversing the nation’s fortunes.

Unfortunately, we waited two years, and that “competitive AI” did not appear.

It was not until January 2025, when DeepSeek delivered R1, that we could truly say we had gained a counterweight to balance the scales. While it wasn’t that China’s large models had immediately caught up with the U.S., the violent swing of the scales had stabilized. The trend of continuing to tilt toward the U.S. was halted, and China’s side began to slowly rise.

In this sense, DeepSeek fully deserves the title of “a nation-defining product,” and Liang fully deserves the honor of being one who “shifted the nation’s destiny.”

What exactly made DeepSeek-R1 so special that it shook America’s tech community and Wall Street so profoundly, and why is its impact on the future so far-reaching?

The core keywords are three: exceptional performance, extremely low cost, and extreme open-source.

In terms of performance, R1 surpassed all existing open-source models in tasks such as mathematics, coding, and natural language reasoning, and even approached—and in some metrics surpassed—OpenAI o1, the most advanced closed-source reasoning model at the time.

This was the first time a Chinese large model had reached this level of performance. From then on, the claim that “Chinese large models can’t match American ones” became a question that could be challenged and overturned, rather than an accepted fact.

In terms of cost, R1’s usage cost was only about 3% of o1’s. Its training cost was even more astonishing—from V3’s pre-training to R1’s reinforcement learning, it cost only $5.776 million, with the incremental training cost alone coming in at just $294,000.

The numbers $5.576 million and $294,000 became two figures that American media emphasized repeatedly. Because in American companies, achieving the same training results would typically cost tens of millions or even hundreds of millions of dollars.

This enormous disparity forced American companies to urgently study DeepSeek to figure out how it achieved such low costs. But these American companies would soon fall into despair—they simply could not achieve such low costs, because in these companies, “every generative AI organization leader’s salary is higher than the entire V3 model’s training cost, and we have dozens of such leaders.”

More critically, R1’s low cost overturned the long-held assumption in the American AI industry—that AI progress must depend on piling up computing power. It created a new narrative: through engineering and algorithmic innovation, even with less computing power, you can achieve top-tier results. This was the primary reason Nvidia’s stock lost nearly $600 billion in a single day.

From the perspective of national competition, the strategy the U.S. had been using—controlling computing power, banning high-end chips from being sold to China, to block China’s AI development—suddenly revealed a fatal flaw. The U.S. had thought the chokehold was airtight, but unexpectedly there was still so much room to breathe.

To this end, the U.S. Congress convened a special hearing titled “DeepSeek: A Deep Dive,” to assess DeepSeek’s impact on America’s AI industry and chip export controls. At hearings convened by the U.S. Homeland Security Committee and the U.S.-China Economic and Security Review Commission on China’s AI development, DeepSeek was also the top topic of discussion.

In terms of open-source, DeepSeek-R1’s degree of openness far exceeded that of all other major large models at the time.

R1 used the extremely permissive MIT License, allowing anyone to freely use the model (including for commercial purposes) without authorization, to further develop and fine-tune it according to their needs, and even to republish modified versions, including for commercial closed-source use.

More importantly, R1 not only published the model weights but also released a detailed technical paper and provided complete reasoning code and tools—essentially open-sourcing the “know-how” to the entire world.

This paper would later appear on the cover of Nature, the world’s most prestigious academic journal, in September 2025, making R1 the first mainstream large language model to pass Nature’s independent peer review. Liang himself was named one of *Nature*’s Top 10 Scientific People of 2025, described as a “Tech-disruptor.”

For the world, DeepSeek’s large model—with its performance rivaling the best closed-source models, its extremely low training and usage costs, and its extreme open-source nature—advanced humanity’s journey toward “AI for all” by a giant leap.

Previously, closed-source APIs were luxury goods available only to a very few companies—even for small and medium-sized enterprises, accessing an API was expensive, let alone training a large model from scratch. But starting with R1, cutting-edge AI technology truly became a “public good” available to everyone.

The same was true for C-end users. Previously, they had to spend tens or even hundreds of dollars per month to access top-tier AI; now, they could get products of comparable quality completely free of charge.

This was a major shock to the business model that American companies were accustomed to.

Previously, American companies were used to spending enormous sums to build a moat, establish a monopolistic market position, and then charge high prices to reap high profits.

But DeepSeek, with its extremely low training and usage costs, posed a serious threat to this model. It was as if, just as American AI companies had set the table and were about to enjoy a fine steak dinner, DeepSeek barged in and flipped the table.

The data speaks for itself.

After R1’s release in 2025, DeepSeek continued to roll out a series of new products and has consistently remained one of the world’s most popular large models.

From the mobile perspective, although DeepSeek never deliberately pursued user numbers, its app’s monthly active users have long remained in China’s top three. From the web perspective, DeepSeek leads by a landslide—ranking first in China, with traffic higher than the second, third, and fourth combined, and globally ranking in the top five.

From the perspective of the global AI ecosystem, according to data from OpenRouter, the world’s largest AI model aggregation platform, DeepSeek models’ API call volumes have consistently ranked first globally, and have driven the total call volume of Chinese models to more than three times that of the U.S., surpassing the U.S. for sixteen consecutive weeks.

Among these, a single DeepSeek model—V4-Flash—alone accounted for 70% of the total token volume of all models from all U.S. companies combined in a single week (U.S. total: 10.26 trillion; V4-Flash single model: 7.22 trillion).

This demonstrates that DeepSeek’s models have become the primary “default option” for global developers when building applications. It has become an indispensable core unit of the global AI infrastructure—a domain that had previously been the exclusive preserve of American AI companies, the altar where American firms sat on high and received offerings from the world.

Chapter Sixteen: Flipping the Table

“Flipping the table” was never Liang and DeepSeek’s original intention. They simply wanted ordinary people to be able to afford better AI at lower cost.

More than twenty years ago, in order to start a business, Liang had diligently studied various management and business knowledge. At the time, his greatest idol was Jack Welch, the CEO of General Electric, and he treated many of Welch’s ideas as gospel.

Years later, as Liang grew and as the world changed, much of what Welch had said no longer held the same weight for him. But one most important statement—Liang always believed it was true—was what Welch said: A company’s most important thing is its vision.

When he was building High-Flyer Quant, Liang’s vision was “to rival the world’s quantitative trading titan—Simons’s Renaissance Technologies.”

This vision had once inspired Liang to lead High-Flyer Quant on its meteoric rise, creating a legend in China’s quantitative private fund industry.

But for a more mature Liang, pure scale and money alone had long ceased to hold any appeal. This vision still existed for High-Flyer Quant, and High-Flyer Quant was still advancing toward it, but for Liang’s personal ambitions, it had become too light, too small, too uninteresting.

So, moving from quantitative trading to AI, from High-Flyer to DeepSeek—what was Liang’s new vision?

We need to trace back to when he first entered AI.

Liang later explicitly stated: he and dozens of his colleagues were not doing DeepSeek for money at all. When they founded the company, not one of them had ever thought about going to the capital markets or going public—if that was what they wanted, they wouldn’t have come.

If not for money, then for what?

DeepSeek has never published a standardized statement of its vision. In Liang’s words: “A vision is not a slogan hung on the wall.” “This vision isn’t even written down—it’s not something that was put in writing; nothing has ever been written out.”

But from his many comments about vision, and from what DeepSeek is doing and how it does it, Liang and DeepSeek’s vision can be summarized in one sentence:

Using originality and open-source as the path, explore the essence of intelligence, and create AGI that everyone can afford.

Liang believes that although everyone at DeepSeek may understand the vision differently, they are aligned on one big direction: “We are doing this with a great deal of kindness toward this world, and we believe this is useful to humanity.”

The entire DeepSeek company was created and operates around this principle. The company has no organizational hierarchy, no KPIs, no performance reviews. Everyone decides for themselves what to do each day, driven entirely by vision. Even more remarkably, as an AI company, DeepSeek barely requires overtime—a stark contrast to other internet and AI giants.

Liang’s “great kindness toward this world” is not just empty talk—it is practiced in every company decision and in everyone’s daily work.

One example proves the point: when DeepSeek launched the DDCP model, they initially priced it relatively high because they were worried that if demand was too high, the service wouldn’t be able to handle it. But the team members were all unhappy about it—because none of them wanted to set a high price. Later, when conditions matured and the concerns were resolved, they cut the price to a quarter of the original, and everyone suddenly became very happy.

Liang and the company, coming from a quantitative finance background, were entirely capable of precisely calculating the elasticity of user demand—how much sales would increase for every 1% price drop, or how much they would decrease for every 1% price rise—and then setting the price that would maximize company profit. But they never did that.

Even when Liang knew perfectly well that for some products, within a certain price range, demand elasticity was zero—meaning that doubling the price would not cause any users to leave, effectively allowing them to double profits with virtually no risk—he absolutely refused to do it.

Because his team’s real purpose was never about making more money. It was about how to make the models “very cheap, very high-quality, and accessible for everyone to fully use.” That is why when the company lowered prices—meaning reduced profits—everyone cheered in the company group chat and felt genuinely happy.

For the same reason, DeepSeek has shown extreme commercial restraint. There were many times when money was practically lying on the ground for them to pick up, but they chose not to take it.

For example, during the 2025 Lunar New Year, DeepSeek’s users suddenly increased by hundreds of millions overnight. For any other company, this level of traffic would have been a once-in-a-lifetime windfall—many companies would do anything to achieve such results. But Liang’s approach was exactly the opposite. He did not pursue how to retain these users, did not try to monetize them, and certainly did not use this momentum to build a super app like Douyin or WeChat…

DeepSeek was entirely capable of doing so, but it simply didn’t. They spent very little cost on maintaining consumers daily active users. Other companies fought tooth and nail for consumers, while DeepSeek didn’t compete at all. Instead, it opened up its DeepSeek models as foundational technology to all companies—including competitors—and let them take users away. For example, Tencent’s Yuanbao at the time made its integration of the full-power DeepSeek R1 its biggest selling point. Alibaba’s DingTalk, Baidu Search, and ByteDance’s Feishu all integrated DeepSeek models as well.

If you look at this purely from a commercial interest perspective, it’s baffling: isn’t this helping competitors? To put it more bluntly, isn’t this “aiding the enemy”? But this is Liang and DeepSeek’s style. They are willing to be extremely open-source and extremely inclusive, hoping to help everyone—even competitors—advance AI capabilities faster together, so that AI can better serve the development of all companies. In the long run, this is more conducive to everyone moving together toward AGI that is accessible to all.

What’s funny is that some media and self-media have used “DeepSeek’s monthly active users declined” or “DeepSeek’s user numbers only rank X” to try to prove that “DeepSeek is failing.” These people simply don’t understand: first, DeepSeek never pursued those numbers in the first place; second, many users, though using other apps, are fundamentally still using DeepSeek—it’s just not reflected in the statistics (for example, for a long time, most Yuanbao users were by default using its built-in DeepSeek, but those daily or monthly active numbers were counted as Yuanbao’s, not DeepSeek’s).

For Liang, competing for C-end users or building a super app could generate money—even a lot of money—but compared to what they ultimately aim to achieve, those are just sesame seeds. The real watermelons lie ahead. Why fixate on those tiny sesame seeds?

The “watermelons ahead” that Liang talks about are ultimately AGI.

Chapter Seventeen: What Is AGI?

AGI (Artificial General Intelligence), while a form of AI in the broad sense and the ultimate direction of current AI development, is—in any context where the term is used—a fundamentally different thing from narrow AI.

According to the paper “A Definition of AGI” by Yoshua Bengio, one of the “deep learning triumvirate” and a Turing Award laureate, and his co-authors:

AGI is an artificial intelligence that can match or even surpass a well-educated adult in both the breadth and depth of cognition.

Specifically, AGI should possess these characteristics:

- Cross-domain learning and reasoning abilities comparable to or exceeding human capabilities;

- The ability to learn from small samples, extrapolate, and solve never-before-seen problems;

- Complete and balanced comprehensive cognitive abilities, including common sense judgment, long-term autonomous planning, and the ability to reflect and correct errors;

- The ability for continuous learning and experience accumulation, integrating new information into long-term memory.

In other words, AGI must have intelligence and autonomous learning/problem-solving abilities equal to or greater than humans.

Creating such AGI has been humanity’s long-standing dream. But from the definition and characteristics above, one can also see that this is an extremely difficult path. Giving machines intelligence equal to or surpassing human intelligence is almost like the work of a “god.” Humanity has pursued all sorts of explorations, but progress had always been limited.

It was only after the “ChatGPT moment,” with the surge of large models emerging and competing with each other, and with AI evolving at an explosive, unprecedented pace, that humanity truly began to see the dawn of AGI.

Currently, the prevailing view is that AGI is within reach. Apart from a small minority of pessimists who believe that humanity is still far from AGI—needing more than ten years or perhaps impossible to ever achieve—most industry insiders and scientists estimate 5 to 15 years. The more optimistic even believe that AGI could appear within 1 to 5 years—Elon Musk, Dario Amodei, Sam Altman, Demis Hassabis, Jensen Huang, and others generally hold optimistic views.

Liang has not given a specific timeline, but he has explicitly stated: “From a technical roadmap perspective, the path to AGI is actually quite clear.” This indicates that Liang already has a confident plan for how to achieve AGI.

Liang believes that AGI can be reached step by step. Each step paves the way for the next. Humanity has already climbed many steps—after the ChatGPT moment, we reached language models; in 2025, we reached Chain of Thought (CoT); in 2026, we reached Agents. The next step after that is continuous learning. Once we step onto the continuous learning stair, humanity will arrive at the singularity.

In Liang’s view, the singularity is not a point, not a sudden mutation, but a continuous, gradual process. After overcoming the “continuous learning” bottleneck, AI enters an acceleration phase of self-iteration through continuous learning—it can develop its own new versions, research on its own, and then develop further AI models, ultimately achieving AGI. This process is the singularity.

Of course, different companies and different people have different views and approaches to achieving AGI. But in Liang’s view, the roadmap he has outlined—”building on the steps already taken, first focusing on solving continuous learning, passing through the singularity, and ultimately reaching AGI”—is the smartest and easiest route. Because every step taken paves the way and accelerates the next step. When the continuous learning problem is solved, AI can assist or even lead the creation of the next generation of AI—which will be orders of magnitude more efficient than humans creating the next generation of AI.

In this process, user acquisition and commercial monetization are natural byproducts—they are merely side effects on the road to AGI.

This is fundamentally different from the vast majority of other companies. Most companies do AI to attract B-end or C-end customers with better products, achieve commercial monetization, and make more money. Liang (and DeepSeek) does AI purely for the sake of doing AGI—it just happens that in the middle of the process, some products emerge, and they commercialize them along the way.

This vision, this scale of ambition, this restraint, this focus—essentially amounts to a form of “dimensionality reduction attack.” Companies driven by capital, with commercial monetization as their primary motive for doing AI, simply don’t compete in the same dimension as DeepSeek.

But the path Liang has chosen is not without obstacles. The biggest problem now—and in the long run—is the scarcity of talent.

For this endeavor, there is no shortage of money, no shortage of resources—only a shortage of truly exceptional people.

From the very first day High-Flyer Quant was founded, Liang had valued talent above all. By the DeepSeek era, that emphasis had multiplied many times over.

As someone who deeply dislikes public exposure, his only two media interviews were both for the purpose of recruiting. Even his decision to go from being unwilling to raise external capital to eventually accepting it—the core consideration was never money or resources, but people.

Although DeepSeek already has a remarkably low employee turnover rate by industry standards, and although those who join DeepSeek tend to be more idealistic, when every other company in the industry is competing for talent with massive salaries and tempting stock options, the top talent in the market won’t all flock to DeepSeek just because Liang’s vision is grand or because DeepSeek’s path to AGI is bright.

Every truly talented person in this field is being fiercely pursued by countless tech giants—they don’t all need to crowd into DeepSeek. Even those already working at DeepSeek could, for various reasons, be poached by other companies. This has happened multiple times at DeepSeek, and every time it happens, it is an immeasurable loss.

That is why Liang believes that the company’s most core interest is the stability of its employees, and its greatest risk is employee attrition.

In reality, DeepSeek already offers very high salaries—more than enough to secure a decent, comfortable life. But the market always offers higher pay to poach talent, and not everyone can resist such temptations.

I believe that under the current model of corporate and market operations—pioneered by the United States and driven by capitalist values—this is a nearly unsolvable problem. There will always be exceptionally talented people who choose other big companies over DeepSeek for higher pay, and there will always be people who have already taken on significant responsibilities at DeepSeek who leave because of external temptations. This is a natural law of the market.

But in the long run, this will not affect the final outcome.

Because there will always be many truly great people—especially the greatest among them—who value ideals more than material gain.

As long as Liang and his team continue to insist on “doing this with a great deal of kindness toward this world,” as long as their goal remains “creating AGI that everyone can afford,” as long as they still cheer in their group chats when they lower costs for users (even if they make less money themselves), and as long as they are genuinely on a credible path to AGI—I believe that those who truly have the ability and the will to change the world will be more than willing to walk alongside Liang.

After all, the immense sense of accomplishment that comes from personally participating in the creation of an AGI built on love and kindness, one that can deeply influence every aspect of the world and create a better future for humanity—that is something no amount of money can buy.

Chapter Eighteen: The Man Behind the Legend

Liang’s intelligence, his intense focus, his grand vision—these have all been discussed at length. But few people know about Liang’s deep sense of loyalty and devotion to relationships.

Just a few details, and you’ll get the picture.

In November 2011, Liang went to Tibet and lost his Fiat in the Ali wilderness. Though the car was only worth a few tens of thousands of yuan, he grieved over it for a long time and even posted a dedicated memorial on Weibo. At the time, Liang had been out of graduate school for just over a year, living in a rented room in Chengdu, exploring his dream of bringing AI to reality.

In October 2012, his old classmate and business partner Xu Jin got married. Liang served as the best man, standing behind the bride and groom with a silly, happy grin. At the time, Liang had returned to Hangzhou, fully focused on quantitative trading, and had already made a great deal of money.

In November 2016, the ZJU Information and Electronics Department Class 0203 organized its tenth reunion. Only 14 of the 39 classmates made it back from various places. Liang was among them. From Yuquan to Zijingang, from the soccer field to the dormitories, from visiting their old class advisor to visiting the dormitory matron—he was there from start to finish. At the time, Liang had already founded High-Flyer Quant, managing 1 billion yuan (with over 500 million yuan of his own capital), and was a rising star in the industry.

In January 2025, Liang returned to Wuchuan to spend the Lunar New Year with his parents, visited his grandfather and uncle and aunt in Mililing Village, and played soccer with his childhood friends.

By then, he was already the head of a “100-billion-yuan quantitative private fund,” had attended the Premier’s Symposium, appeared on the national evening news, and had created DeepSeek—which had already shocked the world.

— END —

Reference: 何加盐 | 梁文锋的来时路与明日歌

Share

Leave a comment

The 2026 World Robot Conference, which closed in Beijing this weekend, sent an unmistakable signal: the AI race is moving beyond a narrow contest of large-model parameters and benchmark scores, and into the domains of robotics, manufacturing, and the physical world. The question of whether AI “works and delivers value” is displacing pure technological showmanship as the industry’s central preoccupation.

Among the exhibitors, Unitree Robotics once again commanded the spotlight. In a keynote address, founder Wang Xingxing argued that the robotics industry is approaching its own “ChatGPT moment” — once critical breakthroughs in software and embodied intelligence are achieved, robots could experience the kind of explosive growth that large language models have seen over the past several years.

That prospect raises a pressing question in the context of U.S.-China AI competition: with American heavyweights like OpenAI, Google, Anthropic, and Meta dominating the frontier-model landscape, where does China’s AI competitiveness actually lie?

Subscribe now

A new RAND Corporation report, AI Development in the United States and China: Evidence from a New Dataset of AI Developer Firms, released Aug. 17, offers a data-driven answer. Drawing on a comparison of 1,181 AI developers across both countries, the study finds that the gap in foundational models and mainstream technical architectures is narrower than commonly assumed. Transformer-based architectures predominate in both nations, and foundational-model developers account for only about one-fifth of all AI firms in each country. The majority of companies, in both cases, are building products around specific tasks and industry verticals.

The real divergence, the report finds, lies in where AI is ultimately deployed. American companies are more concentrated in software, healthcare, scientific research, and cybersecurity. Chinese firms, by contrast, are more deeply embedded in robotics, autonomous driving, manufacturing, transportation, and energy. In essence, the United States is focused on making AI smarter, while China is simultaneously working to give AI “hands and feet” — sending it into factories, vehicles, and the physical world.

Could this divergence shape the next phase of the AI competition? The RAND report cautions U.S. policymakers that if their assessments of the U.S.-China AI race remain fixated on performance rankings of frontier models like GPT, Gemini, Claude, and DeepSeek, they risk missing another contest unfolding in parallel — one defined by who commands richer real-world environments, and who can feed AI with more physical-world data, engineering experience, and commercial demand.

The report itself notes that China’s corporate footprint in robotics, autonomous driving, and industrial automation provides a substantial industrial foundation for AI’s entry into the physical realm — a factor that may prove as consequential as any benchmark leaderboard.

Here comes the full text of the research paper:

Chapter 1 Introduction

The emerging U.S. and Chinese commercial artificial intelligence (AI) developer sectors remain poorly characterized. Significant churn in the sector driven by entry, exit, and reorientation has resulted in the persistence of major gaps in the understanding of industry composition in both countries.

In this data-poor environment, much of what is claimed about the commercial AI ecosystems in the United States and China is derived from the activities of a few high-profile developers and policy priorities articulated in national planning documents. For example, in the United States, leading AI developers—such as OpenAI, Google, Anthropic, Meta, and xAI—have collectively invested tens of billions of dollars in the scaling paradigm, in which performance improvements are achieved by training increasingly large language models (LLMs) on vast datasets using massive computational resources. This concentration of effort around transformer- based architectures has led some observers to argue that U.S. AI development is insufficiently diversified across alternative model architectures and approaches.

China’s policy initiatives since 2025 have emphasized diversification in AI research and application domains. In March 2025, “embodied AI” was formally designated a national development priority, and in 2026, China released the Humanoid Robot and Embodied Intelligence Standard System, establishing a nationwide framework for humanoid robot design and embodied intelligence development.

China’s inclusion of neuromorphic computing in the Made in China 2025 strategy has raised questions about its potential to gain an advantage in non-transformer-based architectures.

These initiatives, combined with evidence of Chinese investment in robotics, have prompted concerns that the United States might be lagging behind China in embodied AI and the integration of AI into industrial and robotic systems.

The veracity of these claims is consequential. If true, this imbalance could leave the United States poorly positioned to be the first to achieve artificial general intelligence (AGI) should the scaling paradigm encounter significant limits to performance improvement.

However, given the scale and dynamism of the AI industry, focusing solely on a handful of major firms risks providing a distorted view of the broader commercial AI developer landscape in each country. Moreover, official plans might not accurately reflect the actual distribution of technical capabilities or commercial activity for the United States or China.

Assessing these claims requires moving beyond policy documents, anecdotal insider testimonies, and trends in high-profile firms to systematically characterize each country’s AI developer ecosystem.

Our study undertook that task: We constructed a broad empirical foundation against which claims about concentration, diversification, comparative advantage, and other topics can be evaluated.

The remainder of this report is organized as follows. Chapter 2 describes the process by which we built the dataset—the identification of candidate organizations, the application of inclusion criteria, and the population of firm-level variables. Chapter 3 presents basic descriptive statistics of the U.S. and Chinese AI developer industries. Chapter 4 examines the commercial orientation of firms in each country, covering commercial positioning, model provenance, application domain, and open-weight status. Chapter 5 analyzes the technical approaches taken by firms in each country across five dimensions: model architecture, learning paradigm, data modality, physical form factor, and core functionality. We conclude in Chapter 6 with a synthesis of findings and a discussion of limitations and directions for future work.

Chapter 2 Building the Dataset

As we will describe in more detail later in this chapter, we defined our target population in this study as the set of firms in each country that are directly involved in the development of AI technology (i.e., AI developers). Specifically, we aimed to include organizations that develop, train, or release AI models or AI-enabled systems in which the core AI capabilities are internally developed.

We excluded (1) firms that primarily provide downstream deployment or integration services for models developed by others and (2) organizations whose principal activity is supplying AI infrastructure or components (e.g., chips, cloud computing, or data) without developing AI models themselves.

Our focus on AI developers is driven both by tractability and by the assumption that they represent a primary locus of private-sector AI innovation, although we fully acknowledge that governments, universities, infrastructure providers, and downstream integrators also shape national AI ecosystems.

Throughout this report, we use the phrase AI developer ecosystem as shorthand to refer to the ecosystem of commercial AI developer firms. This term refers exclusively to the population of profit- seeking companies meeting the AI developer criteria described here; it does not encompass universities, government agencies, research laboratories, or other non-firm actors, even for cases in which such actors play important roles in national AI development. Our data collection process proceeded in three phases. First, we constructed a broad set of candidate organizations that exhibit evidence of AI development activity.

Second, we filtered this list according to our inclusion criteria, retaining only those organizations that are firms, are headquartered in the United States or China, and qualify as AI developers.

Third, for each firm in the resulting target population, we populated a set of variables using LLM search agents, complemented by several rounds of human validation. We describe each of these phases in greater detail in the following sections.

Phase 1. Building a Set of Candidate Organizations

Consistent with our definition of the target population—firms that directly develop, train, or release AI models or AI-enabled systems—we identified candidate AI-developing firms using complementary indicators of AI development activity.

In the absence of a single authoritative registry of such firms, we drew from several data sources that capture distinct and observable manifestations of direct AI development.

Specifically, we constructed our list of candidate organizations using four sources: organizations employing individuals in AI development roles, organizations publishing AI research, organizations holding AI patents, and organizations ranking on AI leaderboards.

Firms Employing Individuals in AI Development Roles

Employing personnel in AI development roles provides strong evidence that a firm is engaged in the development of AI technology. Although some employees in AI-related roles might support out- of-scope activities—such as applying preexisting models to internal business processes—we filtered out such false positives during Phase 2 of our process. The data source we leveraged for this category was Enrich Layer, which provides longitudinal career trajectory information on more than 790 million professionals. We defined a set of AI development–relevant occupational roles (e.g., machine learning [ML] engineer), queried Enrich Layer for individuals holding these roles, and extracted the firms employing them in the United States and China. We ran two queries: one in January 2026 and one in February 2026. Using this approach, we captured firms actively investing in AI-specific human capital, a necessary condition for sustained in-house AI development.

Firms Publishing AI Research

Although not all firms that are engaged in AI development publish academic research, the publication of original research in AI is a clear indicator of direct involvement in advancing AI capabilities. Firms that publish AI research are, by definition, generating new technical knowledge rather than merely adopting existing models or tools. Our data source for this category was SciVal, a repository of scientific publication metadata. We queried SciVal for corporate entities that published articles classified under the “Artificial Intelligence” research category during the 2024–2025 period. Any firm with at least one AI publication was included in our candidate dataset.

Firms Holding AI Patents

In addition to publishing research, firms might encode their progress in building AI technology through patents. Patent assignees represent organizations that have invested in the development of novel, legally protectable, AI-related inventions. As a result, patenting activity provides a complementary and more application-oriented signal of AI development than either employment or publication data alone. Including patent assignees in our population helped capture firms that conduct proprietary or product-focused AI research and development (R&D) that might not be visible through academic publishing, thereby reducing bias toward research-intensive firms and improving coverage of the broader commercial AI ecosystem. The data source we used for this category was the Derwent Innovation Index patent data. Our focus is the contemporary AI commercial ecosystem, so we limited results to the 2024–2025 period. To build a set of AI patent assignees, we leveraged a methodology developed in 2025 by the Organisation for Economic Co-operation and Development (OECD) to identify AI patents.6 Specifically, we used a combination of patent classification codes and AI-related keywords to produce a set of AI patents. From this set, we generated a list of corporate assignees from the United States and China.

Firms Ranking on AI Leaderboards

Submitting models to AI benchmarks or leaderboards provides direct evidence that a firm has developed, trained, and released an AI model for external evaluation. Participation in such benchmarks requires internal model development capabilities and is therefore closely aligned with our target-population definition. We drew on two leaderboards: The first was a performance leaderboard aggregating benchmark results across providers, and the second was a ranking of Chinese open-model developers.

Phase 2. Applying Inclusion Criteria

After removing duplicate organizations, Phase 1 yielded a set of 26,629 candidate entities. In Phase 2, we refined this list to ensure that the dataset focuses on the target population of corporate AI developer firms headquartered in either the United States or China.

To achieve this, we applied three inclusion criteria:

• The organization is a company.

• The organization is headquartered in the United States or China.

• The organization qualifies as an AI developer.

We applied each criterion using a web search–enabled LLM-assisted labeling process, which was subsequently refined and validated through manual review. The first criterion (i.e., the organization is a company) ensured adherence to the study’s focus on the commercial AI sector. A positive classification required that the focal organization be a profit-seeking company. Universities, government research laboratories, and individual researchers were therefore excluded from the sample. The second criterion excludes organizations not headquartered in the United States or China. Country of origin was determined by the geographic location of the firm’s global headquarters.

The final criterion ensured that the resulting sample contained only organizations meeting our definition of AI developers. In this report, an AI developer is an organization actively engaged in developing new AI technology—creating, training, or releasing its own models or AI-enabled systems—rather than primarily applying or enabling models developed by others.

A firm qualifies as an AI developer if it meets at least one of two criteria:

• It develops or trains AI models or systems that advance the state of the art in its domain, including foundation models, embodied AI or robotics control systems, neuromorphic or cognition-inspired architectures, novel training paradigms, or other technically novel AI contributions documented through model releases, peer-reviewed publications, patents, or detailed technical documentation.

• It demonstrates sustained institutional commitment to AI R&D through dedicated AI research divisions, recurring model releases or publications, or ongoing investment in new AI development directions. Examples of firms that qualify are a company that trains its own LLM from scratch and releases it via application programming interface (API) (e.g., DeepSeek), a robotics firm that develops and publishes its own locomotion control models (e.g., Unitree Robotics), and a medical AI company that trains proprietary diagnostic models on its own clinical dataset (e.g., Tempus AI).

Our definition excludes firms that, however sophisticated their AI use, do not themselves develop new models or paradigms. Banks deploying AI for fraud detection, cloud providers offering APIs built on others’ models, systems integrators applying existing AI tools to customer workflows, and companies whose AI activity is limited to fine-tuning, prompt engineering, or retrieval-augmented generation using third-party models are all excluded. In all ambiguous cases, the default classification was “no”: Absent direct, concrete evidence of in-house AI development (such as a named model release, a technical publication describing in-house training, or a documented AI research division), the firm was excluded. Speculative or inferential language in the classifier’s reasoning (e.g., “likely,” “appears to,” “may”) automatically triggered a “no” classification regardless of other evidence.

Phase 3. Populating the Dataset

Once the target population was finalized, we populated a standardized set of variables for each firm using a system of LLM-based search agents. Each agent was scaffolded with a web search tool— primarily Gemini 2.5 Pro with online search—and tasked with conducting background research on the firm and extracting the relevant information. Variables collected were basic descriptive characteristics (such as year of founding, publicly traded status, and military AI relationship), commercial orientation details (such as commercial positioning mode and model provenance), and technical variables (such as model architecture, learning paradigm, data modality, and physical form factor).

We refined our prompt design by sampling outputs from approximately 100 randomly selected firm-variable pairs and inspecting the results manually for fidelity. During this inspection, we also examined the reasoning traces underlying each classification, which helped us pinpoint the sources of LLM disagreement and adjust prompts accordingly.

For certain firm variables, we supplemented our web search with additional data sources. When classifying the more-technical attributes of a firm—such as its dominant AI architectures or the learning paradigm underlying its products and services—we granted the search agent access to Perplexity and OpenAlex (a scientific publication repository), enabling it to more readily analyze the firm’s open-source technical publications.

After all variables had been extracted for the complete firm set, we performed a final human evaluation on a holdout random sample of 25 firms, separate from any firm-variable pairs used during prompt refinement. For each audited firm-variable classification, a human rater categorized their agreement with the agent’s classification as Strongly Agree, Agree, Disagree, or Strongly Disagree. We define human-agent agreement as the share of audited classifications receiving either Agree or Strongly Agree.

This evaluation serves two related but distinct purposes. First, it provides a variable-level estimate of human-agent agreement. For example, an observed agreement rate of 90 percent for n = 25 implies an approximate 95-percent margin of error of 11.8 percentage points under the standard binomial approximation for the human-agent agreement measurement.

Second, the audit helps bound the possible measurement error in the binary fractions estimated from agent data annotations that we report in later chapters. For a binary classification variable, the difference between the agent-reported fraction and the corresponding human-reference fraction is bounded above by the human-agent disagreement rate, assuming that the audit is representative and the human rating is treated as the reference standard. Thus, an observed human-agent agreement rate of 90 percent implies a disagreement rate of 10 percent and a conservative measurement-error bound of approximately plus or minus 10 percentage points for the corresponding AI-coded fraction. This bound is conservative because it assumes that all classification errors move the estimated fraction in the same direction; if false positives and false negatives partially offset one another, the resulting bias in the reported fraction would be smaller.

In the audit, nearly all variables retained for analysis showed high (more than 90 percent) observed agreement between human raters and the classification agent over the evaluation dataset. Variables with low human-agent agreement were excluded from the analysis reported in this study. For example, we initially considered tracking the country of education for founders and chief executive officers of firms in our dataset. However, both the AI search agents and internal team members had difficulty obtaining consistent classifications for this variable because of conflicting early-career narratives, ambiguous corporate hierarchies, and rapid movement of leaders across firms. As a result, we excluded this variable, along with others that faced similar reliability challenges, from the analysis.

Scope and Data Limitations

A Firm-Level View of a Broader AI Ecosystem

By excluding universities, government agencies, government research laboratories, individual researchers, and other non-firm entities from our study, we produced a sample that, although analytically useful, did not capture the full AI ecosystem in either country. The excluded sectors play important roles in AI development, and those roles might differ substantially between the United States and China. The findings in this report should therefore not be interpreted as a complete census of all advanced AI activity in either country.

We nevertheless believe that this firm-centric subset is a valuable object of analysis for three reasons. First, commercial firms are central actors in the development, deployment, and diffusion of contemporary AI systems. They translate research into products, platforms, infrastructure, and embodied systems, and they often determine which technical approaches reach users, customers, and downstream developers. Second, focusing on firms standardizes the analysis and enables systematic cross-national comparison using such observable indicators as employment, patents, publications, model releases, commercial offerings, and public technical documentation. Third, many policy- relevant claims about U.S. and Chinese AI competition concern the industrial structure of each country’s AI sector, focusing on such questions as whether firms are concentrated in foundation models, whether they are building embodied AI systems, whether they are publishing open-weight models, and how innovation activity is distributed across technical approaches. A firm-level dataset is well suited to evaluating these claims.

At the same time, using commercial AI firms to draw broader conclusions about national AI ecosystems requires caution. Such suppositions assume that the commercial AI sector captures a meaningful portion of each country’s advanced AI activity; that the effects of national AI strategies, including Chinese policies emphasizing embodied AI and technological diversification, are at least partly visible in commercial firms; and that AI developer firms are a useful lens through which to observe broader ecosystem orientation. These assumptions likely hold more strongly for questions about commercial deployment, product orientation, and industrial diffusion than for questions about early-stage basic research, long-horizon experimental architectures, or government-directed research programs. For example, some of the most speculative work on neuromorphic computing, brain- inspired architectures, or other non-transformer approaches might occur primarily in universities or government-funded research centers rather than in commercial firms. To the extent that this is true, our firm-level approach could understate forms of technical diversification occurring outside the commercial sector.

Coverage Limits in Public Data Sources

The process used to construct both the candidate firm list and the firm-variable dataset contains uneven levels of coverage across the United States and China. For example, when identifying firms, we use multiple data sources to reduce dependence on any single indicator; however, each source has limitations. Employment-based discovery relies on Enrich Layer’s longitudinal career trajectory data, which draws on professional profile information that might have uneven coverage across countries and sectors. Professional profile penetration in China is substantially lower than in the United States, meaning this step likely underidentifies Chinese individuals in AI-development roles and the firms employing them. This effect might be especially pronounced for individuals who work at firms with ties to the People’s Liberation Army, Ministry of State Security, Ministry of Public Security, or other Chinese security services and who are unlikely to maintain public professional profiles. This undercount is partially offset by our patent-based extraction step: The Chinese government provides strong incentives to patent, and the number of Chinese firms surfaced through patent data exceeded the number of U.S. firms in our extraction, which likely counteracts some of the employment-data undercount of Chinese firms. Publication-based discovery through SciVal is similarly valuable but might miss firms whose research appears in venues, languages, or databases less visible to the indexing systems we use. This effect is likely partially offset by strong prestige and career incentives for Chinese AI researchers to publish in top English-language venues. The issue applies to SciVal as well.

Similarly, when populating firm-level variables (e.g., model architecture, learning paradigm) we rely on academic publications, white papers, and online search agents. Coverage of Chinese firms in these data sources is sparser than coverage of U.S. firms. This issue affects how cross-national comparisons should be read. When one country has a substantially higher unknown rate, differences in named categories might reflect both real ecosystem differences and differences in public disclosure. For example, if a larger share of Chinese firms lacks public documentation about model architecture or learning paradigm, then lower observed shares for named technical categories might partly reflect data availability rather than true absence. Later chapters therefore discuss unknown rates directly when interpreting technical comparisons.

The estimates in this report should be interpreted with attention to both missingness and the size of observed differences. Large cross-national gaps, such as differences in software-only versus embodied form factors, are more likely to be robust to plausible coverage and annotation error. Smaller differences, especially those of only a few percentage points, should be treated as descriptive and tentative. Finally, all findings in this report reflect publicly available data collected in January and February 2026. Given the pace of change in the AI developer industry—firms enter, exit, pivot, and reorient rapidly—some classifications might no longer reflect a firm’s activities, and the cross-national comparisons presented here should be understood as a snapshot of each ecosystem at that point in time rather than a durable characterization.

Chapter 3 Descriptive Statistics of U.S. and Chinese AI Developer Industries

This chapter presents basic descriptive information about the AI developer firms in our dataset. We begin by presenting summary statistics for five key firm-level characteristics—headquarters country, year founded, publicly traded status, military AI relationship, and diversified firm status— broken out by country. We then examine the geographic distribution of firms in each country and the temporal pattern of firm founding. Table 3.1 presents summary statistics for the five firm-level variables of interest broken out by headquarters country. Categorical variables are reported as counts and column percentages; year of founding is reported as the median with interquartile range (IQR).

We assign each firm to a national ecosystem using its headquartered country, defined as the country in which the firm reports its primary corporate headquarters. Of the 1,181 firms in the consolidated dataset, we identify 743 as headquartered in the United States and 438 in China.

For each firm, we record the year founded, defined as the calendar year of incorporation as reported in public filings or the firm’s own communications. The median founding year is 2016 for U.S. firms (IQR 2010–2019) and 2014 for Chinese firms (IQR 2003–2017).

We further classify firms by their public listing status. The publicly traded indicator is labeled “yes” if the firm is listed on any stock exchange, distinguishing publicly listed firms from privately held firms. In our sample, 13.6 percent of U.S. firms and 33.8 percent of Chinese firms are publicly traded; the remainder are privately held or of unknown status.

The military AI relationship indicator is labeled “yes” if publicly available sources directly document a commercial or R&D relationship between the firm and its domestic military—the U.S. Department of War or the People’s Liberation Army—specifically related to AI. Procurement contracts, contract awards, or documented defense R&D partnerships constitute qualifying evidence.14 We find evidence of such a relationship for 30.0 percent of U.S. firms and 16.2 percent of Chinese firms, though the Chinese figure is likely an undercount given the limited public disclosure of Chinese defense-industrial AI activity already noted.

Finally, the diversified firm indicator measures whether a firm belongs to a larger parent organization that produces a broad variety of non-AI products or services rather than focusing exclusively on AI development. Examples include Meta and Google in the United States and Alibaba and Tencent in China. By this measure, 28.8 percent of U.S. firms and 56.6 percent of Chinese firms in the dataset are diversified.

Geographic Distribution

U.S. firms are grouped heavily on the coasts in major technology clusters. The five states with the largest number of firms are California (350), Massachusetts (75), New York (70), Texas (36), and Pennsylvania (25). Chinese firms are concentrated in the eastern and southern provinces; the five province-level regions with the largest counts are Beijing (159), Guangdong (90), Shanghai (58), Zhejiang (29), and Jiangsu (26). Figures 3.1 and 3.2 display these distributions on choropleth maps shaded by the number of firms per state or province.

Year of Founding

The median year of founding is 2016 for U.S. firms and 2014 for Chinese firms, with the largest cohorts in both countries established between 2014 and 2018. Figure 3.3 displays the count of firms founded in each year from 2000 to 2025; 188 firms founded before 2000 are omitted from the figure but retained in all other analyses. This clustering is consistent with the wave of commercial AI investment and firm formation that followed breakthroughs in deep learning in the early to mid- 2010s. Figure 3.3. AI Developer Firms, by Year of Founding (2000–2025)

Chapter 4 Commercial Orientation of U.S. and Chinese AI Developer Industries

In this chapter, we examine the commercial orientation of U.S. and Chinese AI developer firms. We present our results for four firm-level dimensions: commercial positioning, model provenance, application domain, and open-weight status. We then use an additional data source, OpenRouter, to test whether the firm-level picture is borne out in actual developer adoption.

Four Dimensions of Commercial Orientation

Figure 4.1 summarizes the four dimensions used in this chapter and the categories that fall under each. The categories in a dimension are nonexclusive. For example, a firm can be tagged as a foundation developer and as a model adapter if it legitimately engages in both activities. Firm-level percentages in a dimension are therefore allowed to sum to more than 100. Across dimensions, the categorizations are independent of one another.

Commercial Positioning

Our commercial positioning dimension is adapted from the layered service-model taxonomy established for cloud computing and its subsequent application to AI-specific business models.

It distinguishes among four primary commercial positioning modes:

• model-as-a-service: selling access to trained models directly, typically through an API

• inference-as-a-service: selling optimized inference for third-party models

• applied AI product: selling an end-user or vertical product whose value to the customer is delivered through an AI-enabled application

• internal use: developing AI capability primarily for use inside the parent organization without commercializing the model itself.

The dominant commercial positioning mode in both countries is the development of applied AI products. Roughly 90 percent of U.S. firms and 75 percent of Chinese firms fall into this category. The model-as-a-service category—populated by such frontier API providers as OpenAI, Anthropic, Alibaba, and Zhipu—accounts for roughly 10 percent of firms in each country. The popular framing of the commercial AI race as a contest among a set of frontier API providers therefore describes a real but quantitatively small (in terms of number of firms) slice of the AI developer ecosystem in both countries. The majority of firms are building applications rather than selling model access. Figure 4.2 presents the full distribution across all four commercial positioning categories.

The largest asymmetry is in the internal use category. About 16 percent of Chinese firms occupy this category, compared with 10 percent of the U.S. sample. In China, many of the firms in this category are large state-owned enterprises, utilities, regulated banks, and telecommunications firms. The pattern is consistent with the observation that a larger share of Chinese firms has a principal application domain of manufacturing, transportation, or energy.

Model Provenance

The commercial positioning dimension asks how a firm sells AI, whereas the model provenance dimension asks what a firm does with models. We define four categories of model provenance:

• foundation model developer: firms that pretrain general-purpose models at scale from scratch

• task-specific model developer: firms that train domain-specific or task-specific models from scratch, typically on proprietary data

• model adapter: firms that fine-tune or otherwise adapt third-party base models with custom weights

• model integrator: firms that build products on top of third-party model APIs without modifying weights.

Although foundation model development absorbs the largest share of media attention, only 19 percent of U.S. firms and around 19 percent of Chinese firms in our dataset are tagged as foundation developers—although the high rate of Chinese firm data missingness (28 percent) might be masking subtler differences between the two countries across this dimension. According to the available data, it is more common in both countries for a firm to train narrow, task-specific models; to adapt existing foundation models with proprietary data; or to integrate third-party model APIs into vertical products. Figure 4.3 shows the distribution across all four model-provenance categories.

Task-specific model development is more prevalent in the U.S. sample (52 percent) than in the Chinese sample (38 percent). Some illustrative examples of the common profiles we observed in our sample for the United States are Abridge, which trains custom speech- recognition models from scratch for medical conversations; Recursion Pharmaceuticals, which trains task-specific cell-imaging and drug-response models for compound screening; and Tempus AI, which trains oncology-specific models on its proprietary clinical dataset.

In China, on the other hand, Unitree Robotics trains task-specific locomotion and manipulation controllers for its humanoid and quadruped robots, iFlytek trains custom Mandarin speech models for industrial voice interfaces, and BYD trains autonomous-driving perception and planning models for its vehicle line.

Again, it is worth noting that interpreting this gap requires caution. The “unknown” share of Chinese firms is consistently higher than the U.S. share across positively assigned dimensions; Chinese counts should therefore be treated as lower bounds. Given the elevated Chinese unknown rate on this dimension (28 percent), we cannot be confident that the observed gap in task-specific model development reflects a real underlying difference rather than an artifact of higher Chinese missingness.

Application Domain

The application domain dimension records the verticals that each firm targets. A firm can serve more than one domain, and our classifier assigned a primary domain alongside any clearly evidenced secondary domains.

Application domain is where the clearest structural difference between the two ecosystems appears. Chinese firms are concentrated in physical-economy verticals: manufacturing (30 percent of Chinese firms versus 15 percent of U.S. firms), transportation (26 percent versus 15 percent), and energy (12 percent versus 6 percent). In contrast, U.S. firms are concentrated in knowledge-intensive verticals: health care and life sciences (28 percent versus 17 percent), scientific research (10 percent versus 5 percent), and cybersecurity (7 percent versus 3 percent). Figure 4.4 plots the share of U.S. and Chinese firms tagged with each domain.

The most pronounced gap is in defense; 13 percent of U.S. firms are tagged versus 4 percent of Chinese firms. We are cautious about reading this gap as a real industrial difference; Chinese defense- industrial AI work is often conducted in state-affiliated enterprises whose public disclosure is limited.

Open-Weight Status

The fourth commercial dimension we consider in this chapter is whether a firm releases any open- weight models. Open-weight release is a strategic choice that crosses the commercial positioning dimension. For example, a model-as-a-service firm, such as DeepSeek, can release its weights publicly while still selling a managed API.

Similarly, an applied-AI-product firm, such as Anthropic, can build its commercial offering entirely around proprietary, closed-weight models. Figure 4.5 compares the share of firms in each country that publish at least one open-weight model.

Assessing the U.S. and Chinese ecosystems by counting firms suggests a similar concentration: Approximately 15 percent of U.S. firms publish at least one open-weight model, versus 17 percent of their Chinese counterparts. The remainder of firms in each country are proprietary only. Some keep their model usage internal (applied product and internal use firms); others release access through closed APIs.

However, the argument we develop in the next section is that mere publication of open-weight models tells only part of the story. Model usage also matters. Whichever set of models developers actually build on is the set that shapes the products, conventions, and de facto standards around which the global AI ecosystem will consolidate.

OpenRouter Data: A Window into Developer Adoption

OpenRouter is a developer-facing gateway that routes API requests across roughly 300 frontier and open-weight models from dozens of model providers and inference vendors. It publishes daily token-volume statistics by model and by developer. We use OpenRouter as a complement to the firm- level taxonomy, not as a substitute for it.

The data capture developer-mediated traffic—requests that third-party developers route through OpenRouter as part of their own software—and therefore exclude consumer chatbot usage and direct use of provider APIs (e.g., requests sent straight to OpenAI, Anthropic, or Google rather than through OpenRouter). The relevant property for our purposes is that OpenRouter is the principal way to access many open-weight Chinese models from inside the United States, which makes the data well suited to measure how open releases are used in this specific developer population.

Three caveats apply to the OpenRouter data and should be borne in mind throughout this section. First, OpenRouter represents an unknown but likely small fraction of total global developer API traffic; the majority of developer calls to models such as GPT-4o, Claude, and Gemini are made directly to provider APIs and are not captured here.

Second, because OpenRouter structurally overrepresents open-weight models relative to the broader market—and because Chinese frontier models are predominantly open-weight, whereas leading U.S. frontier models are predominantly proprietary—the platform’s user base is not a neutral sample of global developer preferences and might be most relevant for describing relative trends rather than absolute ones.

This interpretation is consistent with a 2026 RAND analysis of global LLM adoption patterns using broader web traffic data, which found Chinese models rising from roughly 3 percent to 13 percent of global market share over a similar period, a meaningful gain but well short of the parity we observe on OpenRouter (discussed next).

Third, usage patterns on OpenRouter might exclude enterprise or consumer AI engagement, which often happen through internal and/or customized APIs that are not captured on the OpenRouter platform. With these caveats in mind, we interpret the OpenRouter data as a directional signal about the competitive trajectory of Chinese open-weight models among cost-sensitive developers rather than as a definitive measure of global market share.

Examining the data reveals that usage of Chinese models has caught up to that of the U.S. models. In late 2024, U.S. models accounted for roughly 85–90 percent of the top-seven token volume on OpenRouter, with Chinese models near zero. By April 2026, the two countries were at parity on the platform. The Chinese share rose to roughly 25 percent through the first half of 2025 on the strength of DeepSeek-V3 and the Qwen series, retracted as new U.S. frontier releases pulled developers back, and then re-accelerated sharply through the first quarter of 2026. Figure 4.6 shows the trajectory and marks several pivotal model releases.

At the endpoint of the chart, on April 30, 2026, the nine providers accounting for the most OpenRouter traffic are led by Anthropic, Google, and OpenAI from the United States, followed by Tencent, Moonshot AI, DeepSeek, Alibaba (Qwen), MiniMax, and Zhipu AI from China—so, six of the nine are Chinese. Among the Chinese providers, every flagship model is released open-weight. Among the three U.S. providers, only Google’s Gemma model line is open-weight (and only at smaller sizes; Gemini frontier-tier models remain proprietary).

Why is the Chinese open-weight cohort displacing closed-API providers in this slice of the market? Part of the answer likely depends on the relationship between price and performance. Figure 4.7 plots the leading U.S. and Chinese frontier models against each other on two axes: an external intelligence-benchmark score and the blended cost per million tokens. The Chinese open-weight models cluster in a narrow band around $1–2 per million tokens at intelligence scores of roughly 48– 54. The leading U.S. proprietary models span roughly $1–12 per million tokens at intelligence scores of roughly 48–60. At the top of the intelligence axis, the U.S. proprietary models are still ahead; for the bulk of developer-facing workloads, the Chinese open-weight cohort is delivering comparable measured intelligence at roughly one-quarter to one-fifth of the price.

Figure 4.7. Frontier Models: Blended Cost per Million Tokens Versus External Intelligence- Benchmark Score

Chapter 5 Technical Focus of U.S. and Chinese AI Developer Industries

In this chapter, we examine five technical dimensions of the same firm population: model architecture, learning paradigm, data modality, physical form factor, and core functionality. Finer- grain details on these dimensions are presented in Appendix A, the AI-driven process executed to classify firms along them is discussed in Appendix B, and core prompts are presented in Appendix C.

Five Technical Dimensions

Figure 5.1 summarizes the five technical dimensions used in this chapter and the categories under each. As discussed in Chapter 4, categories in a dimension are non-exclusive, and categorizations across dimensions are independent.

Figure 5.1. Technical Profile Dimensions and Categories

Two data coverage notes, common to much of the analysis in the report, apply to the technical dimensions. First, technical disclosure is uneven across firms. Many firms describe their products but not their underlying model architecture or training methodology. The “unknown” rate for model architecture is 56 percent in the U.S. sample and 68 percent in the Chinese sample; for learning paradigm, the corresponding rates are 35 percent and 60 percent. Second, even when sufficiently granular information is available to classify firms across these variables, it is typically drawn from technical blog posts, white papers, and research publications. Therefore, our data are more likely to capture firms that publicly document their technical activities, creating a potential visibility bias against firms that disclose less.

Model Architecture

The model architecture dimension records the neural-network family (or families) underpinning a firm’s AI work in the following categories:

• transformer • convolutional neural network (CNN)

• recurrent neural network, graph neural network

• diffusion model, generative adversarial network

• neurosymbolic

• post-transformer (state-space and related)

• classical (non-neural) ML

• unknown

As already discussed, the category of unknown is the largest single share in both countries because many firms do not disclose architecture-level detail publicly.

The structure of our taxonomy was motivated by merging AI/ML technical taxonomies established by such conferences as NeurIPS and ICML, the arXiv preprint server, and the Stanford AI Index.17 We combined, deduplicated, and blended these external taxonomies using insight from internal subject-matter experts; further guidance was provided by exploring the technical details available across our dataset.

Among the firms for which architecture could be confidently classified, transformer is the most prevalent category in both countries, tagged on 28 percent of U.S. firms and 27 percent of Chinese firms. Convolutional networks are second (13 percent and 11 percent, respectively), consistent with the continued prevalence of CNNs in vision and perception pipelines. Diffusion models appear at roughly 7 percent of U.S. firms and 7 percent of Chinese firms, principally in image and video generation. Overall, the architecture shares are very similar across the two countries. Figure 5.2 presents the share of U.S. and Chinese firms tagged with each architecture category.

The learning paradigm dimension records how a firm trains its models. Following the taxonomy discovery process mentioned in the previous section, seven categories are distinguished:

• supervised learning

• self-supervised learning (e.g., next-token prediction)

• reinforcement learning (RL) (e.g., environment RL for robotics and preference-based RL for language models)

• imitation learning • unsupervised or classical methods

• continual learning

• unknown

Figure 5.3 shows the distribution of firms across all learning-paradigm categories for each country.

Supervised learning is the most common paradigm in the U.S. sample (56 percent) and is substantially less common in the Chinese sample (35 percent). Self-supervised learning, the paradigm underlying LLM pretraining, is tagged on 20 percent of U.S. firms and 16 percent of Chinese firms, consistent with the similar foundation-developer cohort sizes described in Chapter 4. However, it is important to recognize for both of these claims that the elevated Chinese unknown rate (60 percent versus 35 percent) should be borne in mind.

Figure 5.3. Distribution of Firms, by Learning Paradigm

RL is an interesting category for comparison because its application is different in the two ecosystems. RL is tagged on 19 percent of U.S. firms and 19 percent of Chinese firms, but the firms apply this learning paradigm in distinct ways. After a qualitative read of firm-level information, on the U.S. side, the concept is most often associated with preference-based RL for language models (RL from human feedback [RLHF], direct preference optimization, group relative policy optimization- style reasoning training at OpenAI, Anthropic, and similar labs). On the Chinese side, the concept is more often associated with environment RL for robotics, autonomous-vehicle control, and industrial automation—as reflected in the joint distribution with the application domain taxonomy.

Data Modality

The data-modality dimension records the kinds of data a firm’s models consume. Categories are the following:

• text

• vision (images)

• video

• audio or speech

• three-dimensional (3D) or spatial data

• sensor

• IoT signals

• tabular structured data

• time series

• genomic and molecular data

• unknown.

Vision and text are the two most prevalent modalities in both ecosystems, and they have similar shares: Vision is tagged on 60 percent of U.S. firms and 67 percent of Chinese firms, and text is tagged on 58 percent and 51 percent, respectively. Video and audio are similar across the two countries as well. Figure 5.4 presents the share of U.S. and Chinese firms tagged with each data modality. Although the rate of Chinese firm data missingness across this variable is nontrivial at 17 percent, the degree of missingness is smaller than that of the other variables in this chapter, suggesting comparative assessments are more robust here than for other variables.

The data-modality dimension provides additional evidence of China’s strength in embodied AI. Chinese firms have large prevalence percentages in the 3D or spatial and the sensor or IoT modalities most directly associated with embodied systems. For example, 3D or spatial data is tagged on 36 percent of Chinese firms versus 27 percent of U.S. firms, and sensor or IoT is tagged on 29 percent versus 25 percent. On the other hand, for the tabular or structured, time-series, and genomic or molecular data modalities (which are more closely related to knowledge-work), U.S. firms exhibit slightly larger prevalence percentages than Chinese firms, although most differences are in the noise floor set by data missingness.

Physical Form Factor

The physical form factor dimension records what physical instantiation, if any, a firm’s AI is delivered in for the following categories:

• software-only (no embodiment)

• humanoid robot

• ground robot or quadruped

• autonomous vehicle

• aerial drone

• deployed sensor network

• wearable or bio-device

• other form factor

• unknown.

Figure 5.5 shows the distribution of firms across these form-factor categories for each country.

Physical form factor is the dimension on which the two ecosystems diverge most sharply. Sixty- one percent of U.S. firms in our dataset are software-only versus only 26 percent of Chinese firms. The Chinese unknown share on this dimension is 22 percent, and because positively assigned counts of Chinese firms should be treated as lower bounds (Appendix B), the true software-only gap is likely real in direction but could be considerably smaller in magnitude than the observed figures suggest.

The measured distribution of embodiment modes reflects that humanoid-robot firms are more than five times as prevalent in the Chinese sample (12 percent) than in the U.S. sample (2 percent). Similarly, ground robots and quadrupeds are tagged on 24 percent of Chinese firms versus 16 percent of U.S. firms, and autonomous vehicle firms are tagged at 14 percent versus 7 percent.

This difference is consistent with the application-domain asymmetry discussed in Chapter 4. Chinese overrepresentation in manufacturing, transportation, and energy verticals is mediated through embodied form factors—humanoid and quadruped robots for manufacturing inspection and material handling, autonomous vehicles for transportation, and sensor-network deployments for grid and energy monitoring. The U.S. concentration in health care, scientific research, cybersecurity, and developer tooling, by contrast, is overwhelmingly software-only.

Core Functionality

The core functionality dimension records what an AI system is being used to do. The categories are not mutually exclusive—a product can perform both perception and planning, or both language understanding and code synthesis—and the classifier assigns all categories that are clearly evidenced. The category list is as follows:

• perception and sensing

• language understanding

• planning and autonomy

• predictive analytics

• explicit reasoning

• retrieval and search

• code synthesis

• creative generation

• recommendation/personalization • multi-agent/distributed

• unknown/unclassified. Figure 5.6 presents the share of U.S. and Chinese firms tagged with each core functionality category.

Most categories show only modest U.S.–China differences. Perception and sensing, language understanding, and planning and autonomy are the three most prevalent functionalities in both ecosystems, each tagged on roughly one-half or more of firms. The largest apparent U.S.–China gaps appear in predictive analytics (44 percent of U.S. firms versus 27 percent of Chinese firms) and retrieval and search (17 percent versus 8 percent), both of which lean toward knowledge-work use cases consistent with the U.S. vertical mix. However, because positively assigned counts of Chinese firms should be treated as lower bounds (Appendix B), the true magnitude of both gaps might be considerably smaller than shown. By contrast, the China-leading gap on planning and autonomy appears more robust and is consistent with China’s national emphasis on robotics and embodied AI.

Chapter 6 Conclusion

This report provides a systematic empirical foundation for analyzing the AI developer firm ecosystems in the United States and China. By constructing and analyzing a dataset of more than 1,100 AI developer firms across both countries, we aimed to ground claims about ecosystem structure, concentration, and comparative advantage in observable firm-level data rather than inference from national plans, expert intuition, or the activities of a handful of prominent firms.

One of the central motivating concerns in the debate over U.S.-China AI competition is whether the United States is insufficiently diversified in its approach to AI development relative to China’s approach. China’s national AI strategies have emphasized diversification through such actions as a formal designation of embodied AI as a national development priority and promotion of neuromorphic computing and other post-transformer architectures via government-funded universities and research centers. Our data shed light on how this agenda is—and is not—reflected in the AI developer firm ecosystem. In our dataset, we find no substantial evidence that

Chinese AI developer firms are pursuing neuromorphic or post-transformer architectures to any greater degree than their U.S. counterparts. Rather, in our study sample, both ecosystems are transformer-led and have nearly identical distribution of firms across other architectures we explored. Chinese AI developer firms diverge most clearly from their U.S. counterparts in the physical instantiation of AI. Only 26 percent of Chinese firms are software-only, compared with 61 percent of U.S. firms. Chinese firms are far more likely to deliver AI through humanoid robots, ground robots and quadrupeds, and autonomous vehicles, and they are overrepresented in physical-economy verticals, such as manufacturing, transportation, and energy. Chinese firms’ RL activity is more often based in a physical environment and centers on application to robotics locomotion, autonomous- vehicle control, and industrial automation—a contrast to the preference-based RL that is dominant in the United States.

Taken together, our findings offer a more empirically grounded assessment of the U.S.-China AI competition than has previously been available. The strongest version of the concentration concern— that the United States pursues a single architectural approach while China pursues many—is not well supported at the level of model architecture by our dataset. However, we do observe that the U.S. AI developer firm ecosystem is more oriented toward software-delivered, language-model-centric AI, whereas the Chinese AI developer firm ecosystem has made structurally larger bets on embodied AI, physical-world applications, and the integration of AI into manufacturing, transportation, and industrial systems.

If the path to transformative AI capabilities ultimately requires codevelopment with robotics, world models, and physical-environment learning, China’s heavier investment in physical form factors in the AI developer firm ecosystem represents a meaningful hedge that the U.S. ecosystem lacks. Without systematic cross-national firm-level data, this distinction would be difficult to draw or support with empirical rigor.

As AI development accelerates and geopolitical competition intensifies, empirical grounding becomes more critical. The findings presented here suggest that U.S.-China AI competition is less a race between monolithic national strategies than a complex interplay of diverse, dynamic ecosystems pursuing overlapping yet distinct pathways. Understanding these pathways requires systematic firm- level analysis. This report provides that foundation, but the data will require continual updating as both ecosystems evolve.

Share

Leave a comment

Subscribe now

The customer gave Lucky an ultimatum: refund the deposit within three days, or she would return with a machete and his bully men.

It was May 2026. Lucky (alias) had sold three shipping containers of solar panels to clients in Lagos, Nigeria’s sprawling port city. But the cargo had languished in customs for over a month. By the time it cleared, the market was already flooded with cheaper products, their prices undercutting her own costs.

The client demanded their deposit back. The company said it had no money. So Lucky paid it herself, out of her own pocket.

This was her second year selling solar power in Africa’s most populous nation. Before this, she had been a nurse back in China. In the summer of 2024, she flew to Lagos to visit family and, on the plane, helped a woman who spoke no English fill out her landing card. The woman later invited Lucky to come to Nigeria and sell solar panels. She accepted.

Lucky recalls looking around the cabin on her return flight to Nigeria. She estimates that 60% of the passengers were there for the solar business.

Chinese solar giants have long since outgrown their home market. Yet the overseas battlegrounds—Europe, India, Southeast Asia—keep redrawing themselves, and yesterday's frontier is today's crowded field.

Now, merchants have set their sights on Africa, drawn by its large populations, abundant and consistent sunshine, and aging, unreliable power grids. Nigeria, with over 200 million people, a decrepit grid, and plentiful sun, has been hailed as the next big frontier.

A year later, Lucky has worked for four companies, closed no significant deals, and found herself entangled in a web of legal disputes. She has been stopped outside markets by local unions demanding a “toll” to enter. Inside, she has seen mislabeled goods and heard rumors of cheap stock used to launder money. In the end, her own shipment rotted in a warehouse. She has decided to go home and return to nursing.

Yet, at the Lagos ports, new containers are still arriving. And on the planes, the seats are still filled with solar entrepreneurs. This land of opportunity is still hot—but it is not a place where everyone can stand their ground.

The Flight

Summer 2024. Lagos, Nigeria’s largest city.

When the cabin door opened, the heat slammed Lucky like a solid wall. She had helped the woman in the neighboring seat fill out her arrival card. The woman, in her forties and dressed in casual business attire—she looked a bit like an accountant—told Lucky she was in the solar business.

Lucky knew nothing about solar at the time. A native of Guangdong province born in 1992, she had been a nurse. She was on her way to visit her husband, a Nigerian, and their two children, who lived in Lagos. The two women exchanged contact information after clearing immigration.

That November, the woman reached out. She told Lucky she planned to start a solar and storage trading business in Nigeria and needed a salesperson. She saw potential in Lucky’s language skills and sociability. Lucky hesitated. The woman insisted, painting a picture of a bright future: Africa’s power shortages made solar an essential commodity. The chance to be with her family and broaden her horizons abroad was too tempting. Lucky quit her stable nursing job in China and flew back to Lagos in January 2025. Once she truly entered the industry, she realized that chance encounter on the plane was no coincidence.

Looking back, Lucky recalled the passengers in the rows around her. “Sixty percent were in the solar business,” she said. Factory representatives, traders, first-time explorers, and those who had already leased warehouses and were waiting for their containers to arrive.

From November to April each year, the dry season in Nigeria brings peak sunshine—and peak solar season. She remembers the flights from China’s Pearl River Delta being packed, wave after wave of solar hopefuls.

The “Next South Africa”

Why were they all converging here? The female boss, the other passengers, the factory owners in the Pearl River Delta, the traders in Lagos markets—they all shared the same vision.

In the second half of 2022, South Africa experienced a rare and prolonged period of load-shedding. Aging transmission and distribution equipment, concentrated maintenance, and a confluence of factors caused power outages to spike. Even in major cities, there were at least four hours of blackouts daily for several months. It was not a short-term blip.

At the time, Li Jiahui (alias) worked for a leading Chinese inverter manufacturer, covering the South African market. He saw, amid the crisis, a simple economic calculation taking hold among residents and business owners. Diesel generators were cheap to buy, but fuel prices were climbing. A solar-plus-storage system, costing close to 20,000 RMB (2,976 U.S. dollars), was expensive. But once installed, the electricity was free. Twenty thousand RMB against “infinite” diesel expenses—the wealthy did the math first.

Demand surged. Companies that reacted quickly captured the wave, designing small-scale solar-storage products for the local market. Only a few manufacturers could meet the demand. Scarcity drove sales; stock would sell out as soon as it arrived. Li saw customers driving their own trucks to the docks to grab units as soon as they landed.

His company, however, had been in South Africa for six years, focused on large-scale ground-mount and utility projects. Their products were high-spec and high-cost, ill-suited for off-grid scenarios with frequent outages. They, and other industry giants, missed the boat.

The boom lasted just over half a year. By the end of 2023, South Africa’s grid had recovered, and outage durations plummeted. The market was left with a glut of solar and storage inventory, prices halved, and many manufacturers suffered heavy losses. “A straight nosedive,” Li described it. South Africa went from one extreme to the other.

But “South Africa” had been seared into the Chinese solar industry’s consciousness. It taught everyone a lesson: in Africa, blackouts are business opportunities. The worse the outages, the bigger the prize.

Zhang Jingkun, who has been in the China-Africa solar-storage trade for years, sees another attraction: low barriers to entry. While Western Europe, India, and Southeast Asia are piling up trade barriers and certification requirements, Africa has no EU-style technical standards. As long as products meet Chinese national standards, they can clear customs, and shipping routes avoid conflict zones.

With the allure of South Africa’s success, the low threshold, and intensifying competition at home, Nigeria—with its chronically unstable power supply, population exceeding 200 million, and status as Africa’s largest economy—became the industry’s promised “next South Africa.”

Before she arrived, Lucky’s mental image of the country was a common one: “Like China in the 1990s. Opportunities everywhere. Hard work will be rewarded.”

The Lamp

Julian (alias) spent eight months in a small central Nigerian city, working for a Chinese mining company.

The first thing he complains about when discussing electricity: “I blew up two chargers.”

Not the cables—unstable voltage, a sudden surge, frying the chargers. He had a spare, and it got fried too.

Once, Julian visited a local state-owned enterprise and found the overhead lights flickering painfully. “They couldn’t be bothered to fix them. They’d just break again anyway,” he said.

Julian lives in what is considered a wealthy neighborhood. Blackouts last two to three hours a day, though power is mostly available. He pays his electricity bill every week or two—around 20,000 naira (about 100 RMB) per payment. That adds up to 300 or 400 RMB per month.

“About the same as a local’s monthly salary.”

Electricity is more expensive in affluent areas but supply is more stable. Down the ladder, in middle-class zones, outages stretch to six or seven hours daily. In poorer areas, power is available for just a few hours a day, at a cheaper rate, but on an inflexible schedule. “Maybe two or three hours in the morning, two or three in the evening.” Beyond the cities, in rural areas, there is no electricity at all.

Julian keeps a gasoline generator at home. Once, when a vehicle knocked down a nearby utility pole, the neighborhood went without power for two weeks. He fired up the generator to save the contents of his refrigerator. His factory’s energy costs are even higher; they run a diesel generator, spending about 60,000 naira (300 RMB) per week on fuel.

This is the daily reality of Nigerian power. Unstable voltage, sudden blackouts, no warnings. The wealthy and factories rely on generators, the poor adapt by using less electricity.

Solar seems like the answer. But the question is: can locals afford it?

One of Julian’s Nigerian colleagues recently spent 3 million naira (about 10,000 RMB) on a solar-storage system for his home. It provides stable power. Julian finds the price “not cheap, but not expensive either—about what it would cost to install a system in China.” If he were living there long-term, owned a house, and had the money, he would certainly install solar.

Prices for home storage systems have nearly halved over the past two years, from around 20,000 RMB to about 10,000 RMB. But what does that mean for the average Nigerian? Julian calculates: a person earning 300,000 naira a month—a good salary—would have to work for ten months, spending nothing, to afford a system. Most people earn 70,000 to 100,000 naira a month, roughly 300 to 500 RMB.

“A system costs ten, twenty, thirty times their monthly income. If you can barely afford to eat, you can’t afford a system,” Julian said. Those who can afford solar are ultimately a minority. Lucky found that her direct clients—wholesalers, retailers, installers—served the same demographic: people who live in wealthy neighborhoods, own factories, or build new homes.

The market exists, but it’s narrower than imagined. Among a population of 200 million, the number who can really afford this technology is likely a very small fraction. And those few are scattered across a country with crumbling infrastructure and a fiendishly complex distribution network.

The first task for every Chinese boss flying in is to find them.

The Ledger

Before finding customers, you have to run the numbers.

The female boss who hired Lucky rented a house, a warehouse, and bought a car in Lagos. Rent is paid annually, the warehouse for one or two years. The car had to be bought outright—over 100,000 RMB. Adding containers, freight, and customs clearance, Lucky calculated: “You land and 1 to 2 million RMB is already spent.”

And that’s before selling a single unit.

The boss set Lucky a sales target: 50 million RMB. Lucky asked if that was in RMB or naira. The boss said RMB.

“If I could do 50 million in sales, would I be out here working for someone else? I’d be my own boss,” Lucky thought.

She couldn’t figure out how the number was set. Later, she realized many Chinese bosses in Nigeria share the same mindset. “They think as soon as they land, their products will have an edge. They believe one, two, ten, or twenty containers will sell out completely.”

However, selling solar-storage in Nigeria is not easy.

First, customs clearance. Li Jiahui’s former employer seriously studied the Nigerian market. They found the import-export customs system chaotic, with official duties at 30% to 40%. Much trade shifted to “gray clearance,” and the actual installed volume and import figures are often not reflected in official statistics.

Then there’s the transaction model. Local specialized distributors have limited capital. They cannot afford to buy a whole container, cannot offer credit terms, and only buy spot goods. This means that for manufacturers to enter the market, they need to hold inventory locally. And whether that inventory sells becomes the boss’s first major headache.

Currency is another challenge. Li noted that in 2023, Nigeria was experiencing significant currency devaluation. This eroded local purchasing power and made it harder for Chinese traders to repatriate their revenues.

“Many bosses misjudged the state of Nigeria’s solar market and the level of competition,” Lucky reflected. She sometimes wonders whether Chinese trading companies, upon arriving in Africa, can realistically expect short-term profits.

The typical scenario: goods arrive and don’t sell. Or they sell, but money doesn’t come back. If the money doesn’t come back, the next order can’t be placed. Warehouse rent accrues, salaries must be paid, fuel must be bought. The initial investment tightens like a noose.

“Living costs here are very high, two to three times what they are in China. The boss thinks that by hiring a Chinese person, who costs as much as ten locals, you have to do the work of ten people.”

That anxiety translates directly into sales KPIs.

And the market where those salespeople are headed, on the other side of Lagos, has its own rules.

One Street, One Union

There’s a place in Lagos called Alaba International Market.

Lucky says it has 5,000 shops, many specializing in solar. It’s one of West Africa’s key solar distribution hubs. When she tried to go in and meet clients, the union officials came.

They told her that for a Chinese person to work the market, she had to pay an entry fee: about 500 to 600 RMB.

Lucky was initially bewildered. She soon understood the unwritten rule: Chinese nationals are not allowed to directly work the market. If you want to go in, you pay. Each street has its own union. You pay the union for Street A, then again for Street B. Otherwise, Union men will stop you and take you to a back room to fine you.

The first time she was blocked, Lucky thought it was extortion. But she realized it was part of the local ecosystem.

“He’s protecting the local secondary market,” Lucky realized. The local solar market has broadly two tiers: a few large-scale wholesalers who take several containers at a time and sell easily, and the many small retailers who typically buy smaller quantities, often pooling together to fill a container.

Lucky’s Chinese trading company had purchase prices similar to those of the large local wholesalers. Selling directly in the market would be seen as disrupting the local business environment. Hence, most market unions impose an entry fee on Chinese sellers. Local regulations also prohibit Chinese companies from opening their own sales shops in these markets.

But there are loopholes. Lucky observed that most Chinese bosses rent a warehouse nearby, stock it, and then hire a large number of local staff to work the markets. They hire ten, twenty, or more at low wages, let them run around, then gradually screen them. “The good ones are kept, maybe one or two become team leaders.” Chinese staff are typically kept for management, finance, and warehousing roles.

Li has reached a similar conclusion. From the manufacturer’s perspective, selling in Nigeria requires significant upfront investment. The most practical method is to partner with local entities, rent a warehouse, and hold inventory.

Zhang Jingkun, who has worked with African clients for years, sums it up in six words: “There is demand, but you need connections.”

Lucky’s eventual strategy essentially fit that logic. She stopped running the markets herself. She hired locals to do it. They could visit ten or twenty shops a day, collect client info, and add them to WhatsApp groups. She would post product details in the groups, and interested clients would message her privately. Once she got a lead, she’d go directly to their shop.

“The union won’t catch you then. You say you’re visiting a friend, not working. You walk straight from the entrance to that shop, talk to two or three people, and leave.”

This was her first lesson in Nigeria: the market has doors, but they are not wide open. You have to learn to go around.

The Fifty Yuan Gap

Having dodged the unions, Lucky finally stood before her clients. And then she hit a harder wall.

Nigeria’s solar market is crowded with several groups: large local wholesalers, small retailers; Chinese trading companies like Lucky’s; large Chinese distributors who rent big warehouses, sell spot goods at volume; and original equipment manufacturers (OEMs) with the lowest ex-factory prices, though most don’t sell directly.

Multiple sources estimate that at least a hundred Chinese solar sales firms are competing locally, along with thousands of solar retailers.

Lucky’s company was caught in an awkward middle position. Like the large local distributors, they bought from Chinese OEMs, so they couldn’t compete on price. When dealing with retailers, they faced rivals both local and Chinese who moved more volume with thinner margins and faster turnover. “When their stock arrives, people from all over come to buy it.”

She had one long-time client, a regular since her first deal. He followed her when she switched companies. He’d reply to her product messages. They were friendly.

But Lucky knew the “loyalty” had an exact price.

Her company’s purchase price was higher—by one yuan per watt. On a 500-watt panel, that was a 50-yuan premium. “Same materials, same quality. Why would they buy from you?”

Li finds this normal. Distributors are not end-users. “They chase maximum profit and margins. Product differentiation is subtle. Competition comes down to price.”

Lucky was selling to middlemen. “Ninety percent of the shops in the market are middlemen, making a spread.”

But the spread is shrinking. Zhang explains: a system once sold for 10,000 RMB with a 5,000 RMB cost; now it still sells for 10,000 RMB, but the cost is 8,000 RMB. “They won’t raise the price to 10,500 because then people won’t buy. The race to the bottom leaves the factories with no profit.”

Lucky asked her clients what criteria they used. They said: first, after-sales—can they fix or replace defective items; second, price.

She couldn’t guarantee after-sales—the company itself might not be there next month. And she had no price advantage.

“The era of clients begging us to sell is over. Now we beg them to buy.”

The Assembly Factories

Connected to the Nigerian market is another world: the solar assembly factories of the Pearl River Delta in south China.

In the summer of 2026, Zhang spent ten days in Guangdong, visiting factories in Huizhou (惠州), Dongguan (东莞), Shenzhen and Zhongshan (中山). Every industrial area was full of them—not one or two, but “a great many.”

Most are assembly plants.

“You don’t need to develop your own cells. CATL and BYD make them, and their prices are so transparent the industry can recite them. You just buy the cells, the casing, the inverter and put them together,” Zhang said.

Li has watched this unfold from Africa. “There’s not much technical barrier in this industry, no solid patent moats, no strong industry constraints.” This means that in the eyes of an African client, the difference between a small factory in Dongguan and a large corporation is not as pronounced as on the price sheet.

Competition boils down to one factor: who is cheaper.

The domestic industry is also mired in a price war. According to the China Photovoltaic Industry Association, from January to October 2024, the price of silicon wafers fell over 45%, and cells and modules dropped over 25%. The industry has experienced repeated cycles of low-price competition.

In response, the government has implemented measures to curb capacity and anti-dumping practices, like introducing cost baselines for components and revising national standards for solar modules.

Lucky recently heard that under the new national standards, modules with a power output below 630 watts will no longer be allowed to be sold domestically. “If they can’t be sold in China, they have to go abroad,” she said.

Lucky has seen the most concrete forms of this price war on the streets of Lagos. The simplest is mislabeling: “A 500-watt panel is labeled as 650 watts. If you sell your genuine 500-watt for ten yuan, and your competitor sells a fake 650-watt for the same price, of course the customer buys the fake.” She adds, bluntly: “Honest business here can be a disadvantage.”

There are also used batteries, wrapped in attractive packaging. “The inside is second-hand. There’s a lot of dirty tricks in this industry,” Lucky said.

Her clients have been burned. One customer told her he’d ordered a certain model, paid in full, and received a container where only the top panels were good; the rest were junk. The supplier stopped responding.

The consequence: local clients eye every new salesperson with suspicion. “They’ve been around, seen too many people come and go.”

Li reveals that sellers of substandard products exploit China’s supply chain advantage in an extreme way—not by making good products cheaper, but by making bad products look decent and selling them for less. When after-sales issues surface en masse in one country, they’re gone, headed for the next.

“A textbook speculative strategy.”

From the assembly lines in the Delta, to the spot markets in Lagos, to Lucky’s price sheet that was fifty yuan higher—it’s a complete chain. Every link squeezes the next. The pressure ends up on the person at the very bottom.

Three Containers

In March, the outbreak of the US-Iran-Israel war and the blockade of the Strait of Hormuz sent global oil prices soaring. In Nigeria, the cost of running gasoline and diesel generators skyrocketed. Solar products experienced a boom reminiscent of the South African surge three years earlier.

“From March and April onwards, virtually all solar panels in Lagos were sold out,” Lucky said.

BloombergNEF estimated at the time that with higher oil prices, a solar-storage system for a household with a backup generator in Nigeria could break even in as little as 16 months. The agency also forecast that Nigeria would compete with South Africa for the position of sub-Saharan Africa’s largest solar market in 2026.

Her company gave Lucky a promise: a container of panels was due at the end of May. She confidently lined up several big clients and secured orders for three containers, collecting deposits. It was the closest she had come to a “win.”

Then the market showed its complex side. In May, the goods landed at the Lagos port but were held up in customs repeatedly, stuck for over a month. Competitors had cleared the same products weeks earlier and flooded the market.

Lucky called the freight forwarder daily. They said, “Soon.” She kept explaining to clients, “Please wait.” When the goods finally cleared in June, the boom was over. The market was flooded with cheap panels, selling for 20% less than her cost.

Clients demanded their deposits back. Because they were past the promised delivery date, she had to refund them. The containers sat in the warehouse, eventually sold at a loss.

She suspects the clearance issues were due to her company using an unreliable agent, or maybe rival firms had better government connections. She’s also heard another theory: Lagos hosts many international telecom fraud companies. Some goods sold in the market are priced far below normal because they were never meant to turn a profit.

“They launder 5 million USD of illegal money, buy 5 million USD worth of goods, sell them and get 4 million back.” She can’t confirm it. What she knows is that during that crucial window, her cargo—and her hopes—were stuck in a place she couldn’t reach.

It was the closest she came to a “big deal.” And it was her last.

Four Companies

Lucky knows some have succeeded.

Mr. Wang, a distributor for a major Chinese solar firm, has been in Nigeria for four years. He lost money the first two, broke even the third, and exploded in the fourth. This year, he plans to move 100 containers—eight or nine a month. He sells spot goods at stable prices and has built a solid reputation. “If I have stock, I sell. If I don’t, I refund the deposit. No problems.”

Lucky admires him. Admires that he made it through, and especially that he had the backing—a stable company—to see it through.

She hears he’s moved his wife and children over, bought land and a house in a wealthy Lagos district. But Lucky was not so fortunate. In a year and a half as a salesperson, she has worked for four companies.

The first was the woman from the plane. After two months, the boss subjected her to psychological pressure and made her pay for her own business trips, only paying her half a month’s salary. Lucky is still in a legal battle with her.

She later heard that boss hired a bogus salesperson who made off with one or two million yuan worth of inventory.

The second company folded, unable to pay its rent.

At the third, the boss wanted one person to do the work of ten and demanded Lucky sign a non-compete clause without compensation. She refused and left.

The fourth was the one with the customs issue. When the company couldn’t repay the clients’ deposits, the clients’ anger turned on Lucky. They threatened her with machetes.

She says her immediate supervisor took kickbacks. He’d tell her to quote 530 yuan on a product the company sold for 500, pocketing the 30 yuan difference. Commissions promised to her went to him. He also gambled, losing millions of yuan in a single night. He borrowed money from Lucky, promising to repay in two weeks. He never did.

When clients demanded refunds, he said the company had no money and told Lucky to find a way. She had to use her own savings to cover the company’s debt.

Finally, he gave her an ultimatum: cooperate, or he’d get her fired. She refused. She was dismissed.

Four companies, four different ways to fail. Lucky says she never wants to work in this industry again.

“Keke”

After being let go, Lucky decided to leave Nigeria.

She vividly remembers her first day in Lagos, early 2018. Her husband drove her around, and they were stopped by police eleven times. They suspected her of drug smuggling, then of trafficking. After some haggling, they’d mutter the Chinese word for “tip”(小费). A small payoff and they’d let her go.

She later learned that low-paid police officers often supplement their income this way.

Back then, she never imagined she would one day work there. When she eventually did, and the drudgery of pounding the market began, she learned to ride the “Keke.”

These are yellow, three-wheeled contraptions, imported second-hand from abroad, ubiquitous on Lagos streets. They’re cheap—about one RMB per ride.

The company car had broken down, no replacements were provided, and transport costs were not reimbursed. Lucky had to improvise: Keke to the corner, walk, catch a minibus, walk again, take another Keke or motorcycle, and finally reach the market. A single trip meant three changes.

“Every day, going back and forth. And I thought, ‘OK, I’ve really become a local.’”

She started moving like a local. The police stops became rare. When stopped, she could laugh and say, “You are my brother,” and they’d usually let her go.

Later, she realized she was never truly accepted. “Nigeria has broken my heart.” She plans to leave this year and return to nursing.

Her friends think it’s a shame—she speaks English, has sales experience, and has spent almost two years in Nigeria. But Lucky says this experience is fate’s way of telling her: “You’re not cut out for sales. Fate is protecting you, warning you.”

She once believed Nigeria was China in the 1990s, where effort equals reward. She now sees the opportunities are real, but only for those with capital, patience, and connections to survive. She is not in that league.

The Hot Earth

In a fast-changing industry, the future is hard to predict.

On April 1, 2026, China canceled the VAT export rebate on 249 solar products, including modules, wafers, and cells. Zhang Jingkun noticed many African clients accelerated their purchases, fearing rising prices ahead.

Li Jiahui, who has weathered the ups and downs of South Africa and the wild growth of Nigeria, now heads the North African market for Fox ESS (麦田能源). He calls the past two years the “darkest” for the industry.

Price wars raged from domestic to overseas, from solar to storage, with bad products driving out good. Yet he believes low prices are not a long-term winning strategy. “The market will eventually favor high-reliability products. The producers of low-end goods will be gradually eliminated. This cycle is inevitable.”

He senses the African solar market is slowly growing, and most manufacturers are increasing their investment. His former company, a major player, has launched lower-cost solar-storage products for several African countries, including Nigeria. More financing options are available, and channel players are placing greater importance on customer engagement, both online and offline.

“Deep cultivation, long-term operation, solid products and consistent investment will surely lead to success,” Li says. Many leading manufacturers are adapting to lower prices, pushing volumes, and many remain confident in Africa’s long-term potential.

“Africa is a hot land, but for now, don’t expect too much,” Li cautions.

Lucky agrees. Nigeria’s solar demand gap is immense. She thinks it will only truly open up when local wages rise and people have more disposable income. But when those gaps will be filled—”That will take a long time.”

Before she leaves, Lucky remembers the flight that changed her life. In the summer of 2024, she helped a woman who spoke no English fill out a form. That flight, she estimates, was 60% full of people chasing the same dream. They had all heard the South Africa story. They all believed Nigeria would be next.

Now she is leaving. But the planes are still landing.

And in Lagos, the hum of generators drowns out the crash of waves against the harbor.

Share

Leave a comment

This weekend, the People's Daily, the official newspaper of the Communist Party of China, published two signed articles under the byline "Zhong Cai Wen (钟才文)." This is a collective pseudonym used by People's Daily to publish heavyweight economic commentaries at critical junctures.

"Zhong Cai Wen," a homophonic abbreviation of "articles by the Central Financial and Economic Affairs Commission"—is widely regarded in financial circles as the authoritative voice of China's top economic decision-makers. This series serves as both an important window for observing how China's leadership views the country's economic situation, or a channel for conveying confidence to the world that there is no need for concern about China's economy.

I've translated these two articles: 钟才文：中国是世界经济增长的积极贡献者和强大稳定锚 and 钟才文：中国经济的韧性与活力. While the overarching narrative of China's economic resilience remains largely unchanged, it's worth noting that the evidence used to support these arguments has shifted.

China is a positive contributor to global economic growth and a powerful stabilizing anchor

China is the primary engine of the global economy, making vital contributions to worldwide development and prosperity. For years, China has contributed around 30% to global economic growth, firmly establishing itself as the undisputed “No. 1 engine.” In the first half of this year, the Chinese economy grew by 4.7%, meeting the full-year target and outpacing other major economies—truly a case of “scenery remaining beautiful here.” Domestic demand played a leading role, accounting for over 80% of economic growth, with consumption contributing nearly 50%. The recently released 15th Five-Year Plan for Expanding Consumption marks the first national-level special plan in this sector, signaling that China’s consumption potential will be further unleashed. Meanwhile, the Consumer Price Index (CPI) rose by 1% year-on-year in the first half of the year. This reasonable recovery in prices aligns with macroeconomic goals and stands in stark contrast to the high inflation and declining purchasing power seen in many other countries driven by surging global energy and raw material costs. China’s economy is characterized by a solid foundation, distinct advantages, strong resilience, and immense potential. The fundamentals supporting its long-term growth remain unchanged, and it will continue to be the most powerful, stable, and reliable source of momentum for the global economy.

China serves as the stabilizing anchor of global supply chains, providing strong support for the smooth circulation of the world economy. Whether facing the shocks of the pandemic or geopolitical conflicts, China has always been a steadfast defender of supply chain resilience and stability. Since late February this year, the US-Israel-Iran conflict has hindered navigation through the Strait of Hormuz, causing tight supplies of international crude oil and key raw materials. However, thanks to the accelerated green transition in recent years, nearly 4 out of every 10 kilowatt-hours of electricity consumed in China now comes from green sources, and the retail penetration rate of new energy vehicles (NEVs) continues to rise, reaching 54.1% in the first half of the year. Leveraging its strong new energy advantages, China has actively responded by reducing crude oil imports, greatly alleviating the pressure of rising international oil prices. Meanwhile, China’s industrial sector has maintained efficient and stable production, offsetting supply gaps caused by production halts and logistics disruptions in certain regions. As The Wall Street Journal noted, China has supported the world economy by reducing its oil imports. Le Figaro observed that China is saving the world economy for the second time since the 2008 financial crisis. Relying on technological innovation and robust market competition, China has built the world’s largest and most complete new energy supply chain system. Over the past decade, it has driven down the global average levelized cost of wind and solar power by over 60% and 80%, respectively. China’s advanced technologies in solar, wind, and energy storage, along with high-quality products like electric vehicles and e-motorcycles, have turned the dream of energy transition into reality for many developing nations. Chinese enterprises have illuminated the nights in sub-Saharan Africa through distributed solar and microgrid projects. Moving forward, China will further guide the rational and orderly cross-border layout of supply chains, aligning with the industrialization aspirations of the Global South to help more countries develop their economies, create jobs, and improve livelihoods.

China is a wellspring of global technological innovation, injecting powerful “innovative momentum” into the world economy. Innovation is a crucial endogenous driver of economic development. Through its own innovative philosophy and successful practices, China is becoming a premier hub for large-scale global innovation. China’s innovation capacity has grown significantly. According to the World Intellectual Property Organization (WIPO), China’s Global Innovation Index ranking climbed from 34th in 2012 to 10th in 2025, and it has held the top spot for the number of global top 100 science and technology clusters for three consecutive years. This year, Chinese innovations have continued to capture global attention: the humanoid robot “Lightning” broke the human world record in a half-marathon, the semiconductor “Tao’s Law” made its debut, China’s aerospace sector achieved consecutive major breakthroughs—including the world’s first dual-mode (sea and land) rocket recovery—and the world’s first open-source large model with nearly 3 trillion parameters, Kimi K3, was released. Chinese innovations are accelerating their global reach. In the first half of this year, 8 out of every 10 humanoid and quadruped robots sold globally were made in China, and the total value of outbound licensing deals for Chinese innovative drugs reached approximately $110 billion. An increasing number of foreign companies note that China has grown into a “super training ground” for global tech innovation and a “hardcore gym” for competitive capability—developing in China helps them build their innovative “muscles” and forge their competitive “sinews.” China is increasingly becoming an enabler of global innovation. In July, the World AI Cooperation Organization was officially established in Shanghai, a major move by China to answer the call of the Global South and unite the international community in promoting AI development and governance. Foreign media commented that China’s advocacy of a "people-centered, AI for good" philosophy, combined with the open-source nature of its AI technology, better aligns with the trends of technological innovation and the needs of the Global South, helping to bridge the global AI divide and promote equitable benefits for all.

China is a practitioner of mutually beneficial cooperation, creating vast opportunities for collaborative development worldwide. The best rebuttal to distortions like “China Shock 2.0” or the “China squeeze theory” is the facts themselves. Chinese manufacturing enhances global well-being. As the world’s largest manufacturing power, China consistently provides the world with high-quality, cost-effective, stable, and reliable products. During the recent heatwave sweeping across Europe, Chinese air conditioners brought much-needed relief, becoming highly sought after. Behind this “cooling power” lies Chinese enterprises’ keen insight into overseas market demands, the swift response of the entire industrial chain, leading energy-saving and eco-friendly technologies, and rapid delivery facilitated by the China-Europe Railway Express. The dividends of Chinese manufacturing benefit consumers and partners worldwide. The Chinese market offers new opportunities for global development. In today’s world, the scarcest resource is the market. With annual imports exceeding 20 trillion yuan, China is a major export destination for nearly 80 countries. While some countries turn to protectionism, China remains committed to expanding openness. Through measures like lowering import tariffs, hosting the CIIE (China International Import Expo), and launching the “Export to China” series of events, China provides a vast market space for global goods and services. On May 1 this year, China implemented zero-tariff treatment for all 53 African countries with which it has diplomatic relations, marking another milestone in China-Africa economic and trade cooperation. Facts prove that the “shock” and “squeeze” theories are fallacies based on a distortion of reality. China’s high-quality economic development brings new and greater opportunities to the world.

China’s Economic Resilience and Transformation

Since the beginning of this year, under the strong leadership of the Party Central Committee with Xi Jinping the core, all regions and departments have taken proactive actions and risen to challenges. By effectively addressing various external shocks and internal difficulties, and solidly promoting high-quality development, China's economic operations have demonstrated a development trend characterized by new growth drivers and an optimized structure, yielding remarkable achievements.

In the first half of the year, macroeconomic indicators remained generally stable. GDP grew by 4.7%, continuing to rank among the top of major global economies. Prices recovered reasonably, employment remained generally stable, and the balance of payments was basically in equilibrium. Meanwhile, profits of industrial enterprises above designated size grew by 18.7%, and general public budget revenue nationwide increased by 4.7%, indicating improved economic performance for some market entities.

Currently, the Chinese economy is in a critical period of accelerating the shift from old to new growth drivers and transforming its development model. Observing the Chinese economy requires looking beyond mere speed; one should not fixate on the highs and lows of individual quarterly or monthly data points. Instead, greater emphasis must be placed on the quality of economic operations—focusing on the substance of technological innovation, industrial upgrading, and the sustainability of economic development, as well as the vitality of enterprises as market entities. In fact, viewed through the lens of quality and efficiency, and from a longer-term perspective, profound changes are taking place in China’s economic development, demonstrating strong resilience and vitality.

China’s technological innovation is being successively translated into industrial competitiveness. Since the beginning of this year, many impressive achievements have been made in sci-tech innovation. For instance, new breakthroughs have been achieved in tackling key core technologies such as integrated circuits, innovative drug R&D is at the forefront globally, and China has successfully achieved both sea-based and land-based rocket recoveries. The supporting and leading role of new growth drivers has strengthened. Value added in equipment manufacturing and high-tech manufacturing grew by 9.3% and 13.3% respectively in the first half of the year, maintaining a sound development trajectory and holding the promise of expanding new development space through technological empowerment. R&D and application in robotics and AI are gaining strong momentum. Industrial, surgical, cleaning, and bionic robots are becoming representative products of “Intelligent Manufacturing in China.” AI technology is driving the orderly transformation of manufacturing toward intelligence, digitalization, and green development, while accelerating the renovation and upgrading of traditional industries.

China’s hub status in global production and supply chains continues to strengthen. China is deploying innovation chains around industrial chains and reinforcing industrial chains through innovation chains. For a new energy vehicle, a complete set of over 10,000 parts can be assembled within four hours at an OEM plant in the Yangtze River Delta; for a humanoid robot, a “one-hour industrial ecosystem” can be formed through efficient collaboration among various enterprises in the Pearl River Delta. China’s exports of intermediate goods have grown rapidly, providing critical component supplies for global production and supply chains, effectively ensuring their stable operation. China’s complete production and supply chain system not only improves global production efficiency, including that of China itself, but also provides certainty for stable global economic growth.

Chinese enterprises are undergoing a brilliant digital and intelligent transformation. Some companies are leveraging big data to gauge market demand and using AI algorithms to optimize production scheduling. Through flexible manufacturing and customized services, they are precisely meeting the needs of small-batch, personalized markets. Others are seamlessly connecting production lines, logistics centers, and consumer orders through digital production systems, achieving end-to-end “zero-touch” integrated fulfillment. This has significantly shortened delivery times, ensuring that products reach consumers fresher than ever. In the construction sector, some enterprises are utilizing industrial standards, digital precision, and automated efficiency to shift the building of “good houses” from “on-site construction” to “factory manufacturing.” Changes like these have greatly expanded the operational scope of enterprises, breaking through geographical limitations and demonstrating powerful vitality.

Risks and hidden dangers in China’s economic operations are being defused and contained in an orderly manner. Since the beginning of this year, the real estate market has shown a stronger trend of bottoming out and stabilizing. In some Tier 1 and Tier 2 cities, the transaction volume of second-hand homes has increased significantly, surpassing that of new homes, with prices stabilizing and inventory pressure easing somewhat. Local government existing hidden debts are being replaced in an orderly manner, and the number of financing vehicles continues to be reduced. The reform, risk resolution, and quality improvement of small and medium-sized financial institutions are advancing steadily, and the number of high-risk local small and medium-sized financial institutions has declined. Overall, by strictly controlling new risks, properly handling existing ones, and strictly preventing sudden collapses, we have prevented and defused major economic risks, holding the bottom line of preventing systemic risks.

Recently, the China Council for the Promotion of International Trade (CCPIT) organized a business delegation to visit the United States, reflecting the wishes and voices of the business communities in both countries. During face-to-face exchanges, the U.S. business community highly affirmed China’s development advantages in emerging fields such as advanced manufacturing, AI, clean energy, quantum computing, and life sciences, calling for expanded pragmatic cooperation while maintaining necessary competition. Meanwhile, executives from multiple U.S. companies investing in China believe that the Chinese government’s efforts to boost domestic demand, ensure and improve people’s livelihoods, strengthen public services, and enhance local R&D capabilities will bring bright prospects to the Chinese economy and provide new scenarios and spaces for multinational companies to deepen their presence in China. They expressed their intention to increase investment and business operations in China, casting a “vote of confidence” in China through concrete actions.

The Chinese economy is like a vast ocean, capable of withstanding storms. Backed by the solid foundation built over more than 40 years of reform and opening up, a massive consumer market, deep manufacturing roots, good infrastructure, a strong talent pool, and a complete production and supply chain system, the fundamental conditions and basic trends supporting the long-term improvement of China’s economy remain unchanged. To truly understand the Chinese economy, one must fully grasp its various advantages and immense potential, allowing one to “not fear the floating clouds that obscure the view.” In the second half of the year, as major projects outlined in the central and local “15th Five-Year Plan” and national special plans are successively launched, and macroeconomic policies are deployed in a timely and effective manner, China’s economic performance will surely remain within a reasonable range, successfully achieving the annual targets and tasks.

The key takeaways of the two pieces are as follows:

📈 Economic Performance & Global Role

- **Steady growth:** H1 GDP grew by 4.7%, ranking among the top major economies, with consumption contributing nearly 50% to economic growth.

- **Global engine:** China has contributed approximately 30% of global economic growth for many years, serving as an “anchor” stabilizing global supply chains.

- **Price stability:** CPI rose by 1% year-on-year, standing in sharp contrast to high global inflation.

🚀 Innovation-Driven Development & Industrial Upgrading

- **Innovation leap:** China’s ranking in the Global Innovation Index rose from 34th in 2012 to 10th in 2025.

- **Notable breakthroughs:** Achievements in semiconductors, aerospace (sea-land reusable rockets), humanoid robots, and AI large language models.

- **Industrial transformation:** Value-added output of high-tech manufacturing grew by 13.3%, with enterprises accelerating digital and smart upgrades.

🔗 Supply Chains & Green Transition

- **Supply chain hub:** Home to the world’s most complete new energy supply chain, with the Yangtze River Delta and Pearl River Delta forming highly efficient industrial ecosystems.

- **Green advantage:** Green electricity consumption accounts for nearly 40% of total usage, and the retail penetration rate of new energy vehicles reached 54.1%, reducing dependence on crude oil imports.

- **Global contribution:** Over the past decade, China has helped drive down the global cost of wind and solar power generation by more than 60% and 80%, respectively.

🤝 Open Cooperation & Market Opportunities

- **Open stance:** Implemented zero-tariff treatment for 53 African countries with diplomatic ties, hosted the China International Import Expo (CIIE), and recorded annual imports exceeding RMB 20 trillion.

- **Mutual benefit:** Chinese manufacturing (e.g., air conditioners) has helped alleviate heatwaves in Europe, refuting the “China shock theory.”

- **Foreign investor confidence:** U.S. companies are optimistic about opportunities in China’s advanced manufacturing, AI, and other sectors, with plans to increase investment.

🛡️ Risk Mitigation & Long-Term Trends

- **Risks under control:** The real estate sector shows signs of bottoming out and stabilizing; local government debt and risks at small- and medium-sized financial institutions are being resolved in an orderly manner.

- **Positive long-term outlook:** Supported by a mega-scale market, a complete industrial chain, and a demographic dividend of talent, the fundamentals for long-term growth remain unchanged.

- **Policy support:** Major projects under the “15th Five-Year Plan” will be rolled out successively, ensuring the achievement of annual targets.

Share

Published in Beijing Cultural Review (BCR) (2026, Issue 2), Oxford graduate students Huang Leike and Yang Dianyi reframe the “China overcapacity” debate by shifting the focus from China to Southeast Asia. Tracing the post-WWII evolution of industrial policy, they argue that ASEAN is undergoing a “re-convergence” toward state-led re-industrialization, aligning with China’s manufacturing expansion.

Rather than viewing this as zero-sum competition, the authors highlight a structural transition in regional division of labor. ASEAN seeks Chinese technology, talent, and management to upgrade local industries, while Chinese firms are moving beyond simple exports to establish multi-layered partnerships in production, investment, and consumption.

They point to complementary opportunities in sectors like EVs and steel, urging Southeast Asian nations to embrace open cooperation. By integrating capacity transfer with localized manufacturing, both sides can mutually support ASEAN’s re-industrialization and China’s global industrial expansion.

Please find the article translation below.

Walking with the “World Factory”: The Future of ASEAN’s Re-industrialization

Recent debates over China’s export surge have sparked intense global scrutiny. Given their geographic proximity and overlapping industrial bases, ASEAN nations—particularly Vietnam, Thailand, and Malaysia—have voiced growing unease.

Since 2021, several of these countries have imposed trade restrictions on Chinese goods; while modest in scale compared to other Asian economies, these measures signal a notable policy shift.

Some analysts attribute this anxiety to China’s overwhelming competitive edge, driven by its sheer manufacturing scale and capacity. Yet this article contends that export volumes and market share are not the crux of the matter; the real issue lies in the evolving industrial division of labor between China and ASEAN.

To grasp the roots of ASEAN’s concerns, we can look beyond short-term trade friction and examine the eighty-year arc of industrialization across the developing world.

Since the end of World War II, industrial policy in the developing world has broadly followed a three-stage arc: convergence, divergence, and re-convergence.

From the 1950s through the 1980s, guided by structuralist and state-led thinking, most developing nations pursued either Import Substitution Industrialization (ISI) or Export-Oriented Industrialization (EOI).

The Latin American debt crisis of the 1980s and the 1997 Asian Financial Crisis, however, ushered in an era of neoliberalism across Asia, Africa and Latin America. Industrial policy receded, and development paths diverged. In the aftermath, only a handful of emerging economies—notably China—stayed the course with state-led strategies, leveraging policy support to drive manufacturing expansion and technological accumulation.

Today, as neoliberalism loses its grip on the Global South, the state is once again emerging as a central actor in industrial strategy. Chinese manufacturing’s outward expansion and Southeast Asia’s industrial revitalization plans are moving in tandem, creating overlaps in certain segments of the supply chain.

This parallel trajectory has deepened the entanglement of regional industrial structures, prompting ASEAN nations to closely monitor their domestic industrial space and policy coordination.

Yet despite this phase of policy-driven competition, the outlook for China-ASEAN economic and trade cooperation remains bright. The two sides remain highly complementary in labor demographics, resource endowments, capital availability, and geographic positioning, leaving ample room to deepen regional collaboration.

By fine-tuning the division of labor across regional supply chains and establishing institutionalized cooperation in emerging sectors such as green manufacturing, the digital economy, and supply chain resilience, both sides can not only ease short-term competitive pressures but also forge a mutually beneficial, win-win industrial synergy.

Divergence and Convergence: The Stages of Economic Development in Developing Countries Since WWII

Industrialization is the inevitable path for nations to achieve modernization and a key force in driving economic growth, promoting technological progress, creating employment opportunities, and facilitating industrial upgrading.

Historical experience shows that almost all countries that have achieved modernization have relied on industrialization to complete economic structural transformation and technological catch-up.

W. W. Rostow described the key node of the transition from traditional society to a modern economy as the “take-off stage,” emphasizing that this period is usually accompanied by breakthroughs in industrialization and profound changes in production methods, serving as a watershed for countries to escape underdevelopment and achieve sustained growth.

In the late 18th century, European countries led by Britain initiated the First Industrial Revolution, breaking the Malthusian equilibrium that had long constrained human social development and directly promoting a shift in the global development landscape. This created what scholars like Pomeranz called the “Great Divergence” between East and West.

For latecomer countries, industrialization is not only a means to drive economic growth and improve employment rates but is also regarded as an important way to break free from external dependence, strive for economic autonomy and development initiative, carrying profound political and psychological symbolic significance.

In the wave of decolonization after World War II, industrialization was viewed by newly independent nations as a “second battlefield” for independence; only by establishing an autonomous and resilient industrial system could they break free from structural dependence on developed countries, break through the unequal “center-periphery” pattern, and gain a relatively independent status in the global system.

Therefore, almost all latecomer countries that successfully achieved economic transformation after World War II chose a state-led industrialization path, forming the first “convergence” from the 1950s to the 1980s.

In this historical process, the industrial policies of developing countries gradually converged, mainly divided into two paths: one is Import Substitution Industrialization, and the other is Export-Oriented Industrialization.

The core of the import substitution strategy lies in gradually replacing capital-intensive and technology-intensive industrial products that were originally dependent on imports with domestic products, cultivating corresponding industrial foundations, breaking dependence on high-value-added imported products, and achieving leapfrog upgrades.

To promote the rapid start-up of weakly based industrial sectors in the short term, governments often need to set high tariffs, subsidize domestic production, and invest state capital; thus, this development model relies heavily on state intervention and fiscal support.

This model is common in most newly independent developing countries in Latin America and Africa, attempting to protect local primary industries and domestic markets through state intervention, develop and establish their own manufacturing systems.

Correspondingly, Asian countries have more often adopted the export-oriented strategy, basing themselves on comparative advantages, prioritizing the development of labor-intensive industries, acquiring foreign exchange, accumulating experience and technology through participation in international division of labor, and achieving gradual upgrades on this basis.

Through government subsidies, low exchange rates, and industrial support, they shape industries with global competitiveness. Under this model, the government assumes more of an institutional platform function, promoting trade facilitation, improving infrastructure, attracting foreign investment, and strengthening talent cultivation.

With the help of the above industrial policies, a few emerging economies successfully completed industrialization, narrowed the gap with developed countries, and to some extent changed the historical trend of the Great Divergence.

During this period, the average annual GDP growth rate in the entire Southeast Asian region generally remained above 7%, making it one of the fastest-growing regions globally, attracting a large influx of foreign investment and rapidly accumulating foreign exchange reserves.

Between 1950 and 1980, the share of developing countries in the global economy rose from 27% to 32%, and their population share rose from 67% to 74%. Although the growth margin was not large, compared to the Great Divergence period (1820–1950), developing countries in this stage had initially shown the potential for catch-up.

However, good times did not last long. By the 1980s, under the impact of neoliberalism, state-led industrialization rapidly receded, and the industrial policies and paths of developing countries experienced a two-stage “divergence.”

The first stage of divergence occurred between Latin America/Africa and Asia. Although Latin American and African countries adopting import substitution policies achieved capacity expansion in the short term, the long-term lack of a competitive environment led to generally low efficiency and technological stagnation in enterprises; state intervention and subsidy policies also brought huge fiscal deficits and massive external debt.

The oil crisis of the 1970s led to tighter financing and export fluctuations, ultimately triggering a severe debt crisis and a major setback in the industrialization process.

After the crisis broke out, as a condition for providing loans, the IMF and the World Bank required developing countries to abandon state-led industrial policies, promote liberalization and privatization, and implement fiscal austerity. The neoliberal development paradigm spread rapidly among Latin American and African countries.

The second stage of divergence appeared between Southeast Asia and China. Unlike Latin America and Africa, Asian countries in the 1980s not only did not abandon state intervention but strengthened the government’s leading role in industrial upgrading, export trade, and technology introduction.

Singapore’s manufacturing share of GDP rose from 20% in the early 1980s to 25% in the mid-1990s, with the rise of high-tech electronics and precision manufacturing, and total exports growing at an average annual rate of over 10%.

South Korea also developed industries such as steel, automobiles, and electronics through its “Heavy and Chemical Industry Promotion Policy,” with manufacturing’s share of GDP rising from 28% in 1980 to over 33% in 1995.

However, the Asian Financial Crisis of 1997 ended this round of expansion and also marked the second divergence in the industrialization path of developing countries.

Afterwards, Southeast Asian countries implemented structural adjustments under rescue plans led by the IMF and the World Bank; industrial policies were weakened, fiscal and monetary policies tightened, and state intervention capabilities significantly declined.

ASEAN experienced a trend of de-industrialization, with the share of manufacturing output, employment, and exports continuing to decline: Malaysia’s manufacturing share of GDP fell from 31% in 1999 to 23% in 2010, and the employment share fell from 23% to 17% during the same period; similar trends also appeared in Thailand and the Philippines, and the industrialization process in Southeast Asia stagnated for a time.

In sharp contrast to the de-industrialization trend in Southeast Asia, after 1997, China not only maintained the continuity of its industrialization policy but further strengthened its state-led development path. Between 1990 and 2018, China’s manufacturing value-added increased by nearly 18 times, and its share of global manufacturing value-added rose from less than 4% to 21%, jumping to become the world’s largest manufacturing country; China’s per capita GDP also grew to 23.2% of the US level by 2016.

Manufacturing not only contributed the core driving force for China’s capital accumulation and technological transformation but also formed the widely discussed “China Miracle” through industrial chain integration, export expansion, and regional development driving effects.

Re-convergence and Conflict: The Convergence and Conflict of Current China-ASEAN Industrial Policies

In the 21st century, the successful experience of China’s rapid economic rise has significantly changed the global economic landscape and shaken the neoliberal development paradigm that has dominated since the end of the 20th century.

In 2008, the financial crisis erupted, and developed countries also began to reflect on the “de-realization to virtualization” development model, returning to a growth logic dominated by the real economy.

Driven by these two forces, more and more Southeast Asian countries have begun to re-examine their own industrialization paths, setting off a wave of re-industrialization, once again attempting to promote the upgrading of traditional industries and the development of emerging industries through government guidance and policy support, achieving a transformation from processing bases to industrial centers, and enhancing their position in the global value chain.

The development paths of China and Southeast Asian countries have converged again in terms of policy and industrial direction. Currently, ASEAN countries’ re-industrialization strategies focus on two directions: “revitalizing four traditional industries” and “cultivating six emerging industries,” aiming to consolidate existing advantages and achieve breakthroughs in high-value-added industries.

In terms of traditional industries, food and beverage processing, textiles and apparel, automobile manufacturing, and basic chemicals and building materials are regarded as the core pillars of the industrial system, undertaking important functions of absorbing employment and enhancing export competitiveness.

Countries have successively proposed specific goals in fields such as automobiles, electric vehicles, textiles, and steel: Thailand plans to increase the ownership of new energy passenger vehicles to 1.2 million by 2036, creating a regional “Electric Vehicle Corridor,” and positioning the development focus of its textile industry on functional fabrics and high-end design processing; although the steel industries of Indonesia, Vietnam, and other countries still rely on imports, they are accelerating the construction of smelting bases, striving to strengthen industrial security and industrial foundations through local capacity and technological upgrading.

From the policy practices of major countries, Indonesia launched the “Industry 4.0 Roadmap,” focusing on the food and beverage industry; Thailand has implemented the “Thailand 4.0” plan since 2016, establishing five traditional advantageous industries and five future industries, promoting an innovation-driven economy through public investment, R&D incentives, and the introduction of high-tech foreign investment; Malaysia released the “12th Malaysia Plan (2021–2025)” in 2021, with “revitalizing manufacturing” as the core, focusing on the development of electronics and electrical, medical devices, renewable energy, and precision engineering, and promoting digital transformation through the High-End Manufacturing Fund and the “National Policy on Industry 4.0.”

The re-industrialization trend in ASEAN countries has led to a high degree of overlap between their industrial policy layouts and China’s. In the field of traditional manufacturing, China has always maintained strong capacity and technological advantages; in the field of emerging industries, core areas of ASEAN’s “Industry 4.0” strategy such as electric vehicles, semiconductors, the digital economy, medical devices, green energy, and aerospace are also important directions for China’s development of new quality productive forces.

In comparison, China’s manufacturing started earlier and has significant advantages in industrial chain completeness, technological accumulation, and large-scale production; on the other hand, against the backdrop of China-US strategic competition, to avoid trade barriers and expand intermediate markets, Chinese enterprises are accelerating their southward layout, gradually deepening industrial ties with ASEAN through capacity transfer, technology landing, and trade channel reconstruction.

Overall, policy convergence provides an opportunity for both sides to deepen cooperation. For ASEAN, how to strike a balance between introducing Chinese enterprises and maintaining local industrial resilience is an important issue to be addressed in the re-industrialization strategy.

As economic and trade exchanges between China and ASEAN continue to expand, bilateral trade volume has grown rapidly, and the trade structure has also shown new characteristics. First, the scale of ASEAN’s trade deficit with China has gradually increased. In the past three years, China’s exports to ASEAN grew by about 24%, while ASEAN’s exports to China grew relatively slowly; the expansion of trade scale is more reflected in the increase in the share of Chinese goods in the regional market.

More importantly is the change in the economic and trade structure. For a long time, the main body of China-ASEAN bilateral trade was intermediate products, with the proportion basically stable at around 65%, reflecting a high degree of synergy and complementarity between the two sides in the regional industrial chain.

These products provided low-cost input factors for ASEAN’s local export-oriented industrial chains in the past, greatly promoting the development of local manufacturing and export-oriented industries.

Some studies point out that although ASEAN has a significant trade deficit with China, this deficit is largely offset by ASEAN’s continuously expanding trade surplus with other markets (especially the United States), reflecting that the benefits of ASEAN-China economic integration depend to a considerable extent on the demand from final consumer markets such as the United States.

This structure has changed in recent years. From 2023 to 2024, the proportion of intermediate goods exported from China to ASEAN dropped from 63.14% in 2020–2022 to 41.01%. In 2024, China’s exports of manufactured goods to ASEAN reached $76.5 billion, an increase of about 75% compared to 2018, and the proportion in total exports increased from 13% to 18%.

The focus of bilateral trade is gradually shifting from intermediate product collaboration to terminal market supply, and the division of labor in the regional industrial chain is also being reorganized accordingly, gradually shifting to a multi-level interaction pattern covering production, investment, and terminal consumption.

Against the backdrop of ASEAN countries universally promoting “re-industrialization” strategies and actively enhancing their domestic industrial chain positions, China’s mature manufacturing system and export expansion have not only brought new momentum to regional cooperation but also, to some extent, formed competition with local industrial policy development and industrial cultivation.

Prospects: Cooperation Opportunities Between China and ASEAN

Although there are differences in the details of industrial policies between China and ASEAN, their development logic is highly consistent—the demand for Chinese industrial relocation and ASEAN’s capacity to undertake industries constitute the structural foundation for alleviating conflicts within convergence.

Currently, China’s manufacturing industry is in a critical stage of transformation, upgrading, and industrial chain reshaping; the outward transfer of some low- and medium-end links has become an inevitable trend for optimizing resource allocation and achieving high-quality development.

At the same time, ASEAN countries are in an upswing period of re-industrialization and attracting foreign investment, urgently needing to introduce technology, capital, and management experience to promote local industrial upgrading.

The Southeast Asian region is rich in labor resources, with relatively moderate costs and a young population structure, providing a continuous supply of human resources for manufacturing development; geographically, ASEAN countries are located at key nodes of the global industrial chain, connecting Asian, Pacific, and Indian Ocean markets; at the policy level, they generally adopt open and inclusive investment strategies, providing relatively complete infrastructure and tax incentives for foreign enterprises.

From a macro perspective, this complementary demand has already been reflected in the continuous deepening of regional institutional arrangements and bilateral investment and industrial cooperation.

With the completion of the upgrade of China-ASEAN Free Trade Area 3.0 and the continuous deepening of a series of free trade agreements including the Regional Comprehensive Economic Partnership (RCEP), the institutional foundation for regional trade and investment cooperation has been further consolidated, providing a more stable policy environment for industrial chain synergy and factor mobility.

China has become the second-largest source of foreign investment for ASEAN, and investment in ASEAN has maintained rapid growth in recent years. As of July 2025, the cumulative two-way investment scale between the two sides has exceeded $450 billion, with Chinese enterprises establishing more than 7,400 direct investment enterprises in ASEAN, employing over 720,000 foreign employees, and the breadth and depth of cooperation continuing to expand.

By the end of 2023, Chinese enterprises had established more than 30 overseas economic and trade cooperation zones in Malaysia, Cambodia, Thailand, Vietnam, Laos, Indonesia, and Brunei, attracting more than 2,000 settled enterprises, forming a development pattern where industrial agglomeration effects and regional driving effects run in parallel.

Among them, projects such as the China-Vietnam (Shenzhen-Haiphong) Economic and Trade Cooperation Zone and the Thai-Chinese Rayong Industrial Park in Thailand have become typical cases of capacity cooperation and localized development; the construction model of the China-Malaysia and China-Indonesia “Two Countries, Twin Parks” has promoted the deep integration of investment, trade, and industrial synergy, becoming a new highland for China-ASEAN economic and trade cooperation.

At the industrial level, cooperation in new energy industries such as batteries and electric vehicles, as well as traditional industries like steel, reflects a trend in China-ASEAN relations gradually shifting from one-way competition to division of labor and collaboration.

In the battery field, Indonesia possesses some of the world’s richest nickel ore resources, a key raw material for new energy vehicle battery production. In the past, Indonesia mainly exported raw ore with low added value; since 2020, the Indonesian government has promoted the development of domestic smelting and battery material processing technology by banning raw nickel ore exports, striving to enhance its position in the global electric vehicle industrial chain.

China’s battery technology and new energy vehicle industry system are mature, and the two have formed a natural complementarity in industrial layout. In 2022, China’s CATL and Indonesia’s state-owned mining company (ANTAM) jointly built a comprehensive battery project worth $5.9 billion, covering nickel mining, material processing, battery manufacturing, and recycling, becoming a model of win-win cooperation between resource-rich countries and manufacturing countries.

As the largest automobile manufacturing center in Southeast Asia, Thailand proposed an “Electric Vehicle Strategy” as early as 2016, hoping to develop the new energy vehicle industry on the basis of traditional automobile manufacturing. Chinese car companies, with their complete industrial chains and technological advantages, have injected new vitality into the Thai market.

Currently, seven Chinese new energy vehicle companies have set up factories in Thailand, with a total investment of about 70 billion Thai Baht. BYD’s Thai factory has an annual capacity of 150,000 vehicles, covering all major processes of vehicle manufacturing and driving the localization of parts and components; Chinese battery companies such as CATL also plan to build battery production bases in Thailand, promoting the formation of a local new energy vehicle ecosystem.

In the field of traditional industries, cooperation in the steel industry also reflects a logic of mutual benefit. Countries such as Thailand and Malaysia have strong demand for high-quality steel, especially special steel and engineering steel, in infrastructure construction and manufacturing development.

Compared to simple imports, local governments are more inclined to attract foreign enterprises to set up production lines locally to drive employment and industrial supporting facilities.

Against this backdrop, Chinese steel enterprises are actively “going global,” which not only alleviates domestic capacity pressure but also helps ASEAN improve its industrial chain structure.

Guangxi Guixin Steel built a 1.8 million ton special steel production line in Indonesia’s Tsingshan Industrial Park, and Taishan Steel’s first phase of four nickel-iron production lines in Indonesia has also been fully put into operation, both being typical models combining capacity cooperation and localized manufacturing.

In addition to manufacturing cooperation, cooperation between China and ASEAN in the green economy and digital economy is also deepening. Taking the digital economy as an example, in recent years, many ASEAN countries have successively launched “Industry 4.0” or digital transformation strategies, hoping to promote re-industrialization through the development of the digital economy.

Against this backdrop, Chinese enterprises have actively participated in regional cooperation in fields such as digital infrastructure, e-commerce platforms, and digital technology R&D, covering the construction of data centers, 5G networks, and cross-border communication networks, as well as the development of cross-border e-commerce and mobile payment systems.

At the same time, enterprises and research institutions from both sides are carrying out cooperation in digital technology fields such as artificial intelligence, gradually forming a regional digital innovation network, providing new momentum for industrial upgrading and digital transformation.

Overall, the complementary division of labor and investment collaboration between China and ASEAN provide a solid foundation for easing structural frictions and bolstering regional industrial resilience.

Moving forward, Chinese enterprises can deepen their integration into ASEAN supply chains, fostering higher-level cooperation in resource development, manufacturing support, technological innovation, and green transition.

Expanding high-quality investment to localize more mid-to-high-end manufacturing in ASEAN will not only drive local industrial upgrading and job creation but also help China optimize its industrial structure and achieve a more rational global allocation of capacity.

Meanwhile, leveraging platforms like the Belt and Road Initiative and RCEP, both sides should strengthen policy dialogue and standards alignment, while improving mechanisms for investment protection, talent development, and supply chain security to foster a more open, stable, and predictable business environment.

By pursuing this complementary, collaboration-driven path, China and ASEAN can achieve a new cooperative equilibrium within this historical “re-convergence,” gradually defusing industrial competition and building a mutually beneficial, sustainable regional industrial community.

Conclusion

History since World War II demonstrates that expanding economic cooperation, facilitating industrial transfer, and supporting latecomers’ industrialization are vital to a nation’s global or regional leadership.

Beyond military or ideological strength, sustainable leadership stems from building an inclusive, mutually beneficial order that shares the fruits of development.

As a priority in China’s neighborhood diplomacy and a key region for the Belt and Road Initiative, ASEAN is at a critical juncture of economic recovery and industrial restructuring. Member states are eager to enhance manufacturing competitiveness and rebuild supply chain resilience to achieve sustainable development.

In recent years, China’s substantial direct investment, intermediate goods exports, and infrastructure cooperation have significantly strengthened the region’s industrial base and supply chains, laying a solid foundation for Southeast Asia’s re-industrialization—a contribution widely recognized by ASEAN governments.

Many have explicitly identified deeper economic and industrial ties with China as a crucial pathway to modernizing their own industries and infrastructure. Regional frameworks, including the upgraded China-ASEAN Free Trade Area and Belt and Road projects, have further provided a stable institutional environment for industrial synergy and factor mobility.

As ASEAN advances its re-industrialization, China should closely attend to evolving industrial policies and the region’s expectations for deeper cooperation. Backed by its 1.4-billion-person super-sized market, China has accumulated substantial capital and technological resources through unleashing consumption, guiding resource flows, and stimulating regional demand.

This capacity not only sustains China’s own development but also creates real opportunities for other developing countries in the region to industrialize and achieve technological leaps.

By further unlocking domestic demand and maintaining open cooperation in outward investment, industrial collaboration, and technological exchange, China is well-positioned to share its hard-earned experience in industrialization, infrastructure, and industrial upgrading with neighboring countries.

Through deeper supply chain integration, expanded mutually beneficial investment, and enhanced exchanges in technology and talent, China can inject new momentum into regional growth while creating opportunities for more developing countries to participate in the global industrial system and upgrade their economic structures.

In this process, China’s development dividends are poised to interact more closely with the development needs of regional partners, helping to shape a more balanced, inclusive, and sustainable pattern of industrialization across Asia.

Share

Leave a comment

China’s Commercial Rocket Odyssey

When I headed to Washington in 2017 as a tech journalist, Falcon 9 had already nailed its booster landings. The following year, the Falcon Heavy’s maiden flight sent a cherry-red Tesla Roadster into a Mars-crossing orbit, recovering both side boosters and the core stage.

I still vividly remember the sheer awe this technological marvel inspired in China. Looking back, that moment marked the absolute peak of China’s fondness for America, right before the trade war began.

In the years that followed, as tariffs and the pandemic sent bilateral ties into a downward spiral, Musk’s SpaceX remained one of the very few bright spots in the Chinese public’s eyes — in particular, the Starship recovery.

Chinese commercial launch companies have made no secret of their strategy to “cross the river by feeling for Musk,” actively adopting his technological frameworks and commercial methodologies. No wonder, their rockets in development are routinely dubbed “China’s Falcon 9,” much like “China’s Starlink” in the satellite internet sector—and more recently, the orbital data centers Musk has been championing.

Beyond SpaceX, only two other players worldwide have achieved rocket reusability, both within a recent year. Blue Origin’s New Glenn made its first booster landing in November 2025, becoming the second company globally to recover an orbital-class rocket, despite a setback in May.

On July 10, 2026, the Long March 10B successfully completed its maiden flight from Hainan’s Commercial Spaceport, achieving first-stage recovery via a sea-based net capture system. This marked China’s first controlled rocket recovery and a world first for net-based recovery.

The Long March 10B is now classified as a commercial rocket, a shift from the past when launches were strictly viewed as national missions funded by space agencies like NASA, ESA or CNSA. However, the rocket still carries a distinct state-backed DNA, as it was spun off from the China Academy of Launch Vehicle Technology (CALT), a major state-owned aerospace enterprise that handled national missions.

Another state-backed contender, the Long March 12A developed by SAST, made its maiden flight on December 23, 2025. The first-stage recovery (also designed to be in Falcon-9-style) attempt failed. The mission was deemed a “basic success.” However, industry insiders felt its recovery performance fell short of Zhuque-3’s first try. Reportedly, the Long March 12A could make its second test flight as early as the end of this month.

The Zhuque-3, developed by LandSpace and launched last December, is considered a true-blooded commercial rocket.

Like LandSpace, a growing queue of Chinese rocket firms is gearing up for similar tests, bolstered by the construction of new land- and sea-based launch facilities across the country and the backing of local governments in Shanghai, Yantai, Guangzhou, Chengdu and Haikou, all of which are cultivating commercial spaceflight as a new engine for future economic growth. The following section will dive into these players.

Notably, after their reusability is achieved, they still face the challenges of true re-flight capability and securing a cost advantage. Yet, this commercial path has become an industry consensus. The real question is how to get there. Companies have very different answers on where to start, what propellant to use, and the pace at which to push forward.

In a blog-style post on WeChat, a writer known as "Lao Tan" analyzed the strategic choices facing China's commercial rocket industry. He pointed out that every divergence reflects the distinct corporate identity, capital strategies and engineering philosophies of these companies.

The Trajectory of Ascent

The genesis of the global commercial launch industry has been anything but uniform. In the West, the consensus was clear: liquid propulsion was the only viable starting line.

SpaceX cut its teeth on the Falcon 1, a diminutive liquid-fuelled rocket that endured three humiliating failures before finally reaching orbit, a crucible that forged its founding engineering cadre.

Rocket Lab took a similar liquid path with its kerosene-burning Electron, spending a decade and dozens of launches building the case for its forthcoming Neutron, a medium-to-heavy reusable workhorse.

Blue Origin opted for a dual-track strategy, mastering suborbital reusability with its New Shepard demonstrator while quietly laying the groundwork for the heavy-lift New Glenn.

Even Relativity Space, despite the maiden flight of its 3D-printed Terran 1 ending in a failure to reach orbit, remains committed to the liquid path, now pushing toward a 2026 debut for its larger Terran R under new ownership.

China’s commercial space sector, however, has charted a decidedly different course. Early Chinese private launch firms largely bypassed liquid engines in favor of solid-fuel rockets. For companies such as LandSpace (蓝箭航天), OneSpace (零壹空间), i-Space (星际荣耀), Galactic Energy (星河动力) and CAS Space (中科宇航), solid motors offered a mature technology base, a shorter supply chain, and, crucially, a faster route to a maiden flight.

Yet this shortcut has a terminal point. Solid propulsion is economically unviable for medium- and heavy-lift vehicles and is fundamentally incompatible with reusability.

The industry is unanimous: the transition to liquid fuel is not a choice but an imperative for any firm hoping to compete in the reusable-launch era. The true divergence now lies not in the destination, but in the mechanics of the pivot.

The First Schism: Solid vs. Dual

As the industry matures, it has bifurcated into two distinct strategic camps, driven by differing philosophies on capital allocation and risk management.

The first camp comprises the liquid purists: firms that have abandoned solid propulsion entirely to concentrate all available resources on liquid-fuelled vehicles. LandSpace exemplifies this approach. After its Zhuque-1 solid rocket suffered a third-stage anomaly during its 2018 maiden flight, the company ruthlessly pivoted, abandoning solid motors altogether to focus on liquids.

The gamble paid off: its Zhuque-2 became the world’s first liquid oxygen-methane rocket to reach orbit, and its larger Zhuque-3 achieved orbital insertion on its maiden flight in December 2025, albeit with a landing anomaly.

The second camp favors a dual-track strategy, balancing the commercialisation of mature solid rockets with the parallel development of next-generation reusable liquid vehicles. Galactic Energy, CAS Space, and i-Space belong to this pragmatic cohort.

Galactic Energy’s Ceres solid rocket has completed 21 successful commercial launches, making it the most prolific private launcher in China, while its liquid reusable Pallas-1 is slated for a maiden flight between late 2026 and early 2027.

CAS Space has logged 14 sucessful launches with its Lijian-1 solid rocket (one failure), and its liquid Lijian-2 completed its maiden flight this March.

Meanwhile, i-Space’s Hyperbola-1 solid rocket has achieved four successes from eight attempts, even as it gears up for the maiden flight of its liquid Hyperbola-3 in late 2026, which will also attempt an offshore recovery.

The advantages of this dual-track model extend far beyond the obvious: the cash flow generated by solid launches insulates these firms from the boom-and-bust cycles of venture financing.

More importantly, operating solid rockets allows companies to stress-test the entire commercial launch ecosystem—from customer acquisition and launch execution to ground telemetry and in-orbit delivery.

Galactic Energy, for instance, has served 27 commercial satellite operators, building a comprehensive service infrastructure. This proven track record acts as a bridge of trust, securing early customers for the far more complex liquid vehicles to come.

The divergence between these two strategies is not merely a technical dispute over propellants. It is a fundamental disagreement over how to manage scarce enterprise resources and mitigate existential risk in a notoriously unforgiving industry.

The Second Schism: Methane vs. Kerosene

Having committed to liquid propulsion, the industry’s next great debate has centerd on the choice of propellant.

The liquid oxygen and methane (LOX-methane) camp argues that methane’s clean-burning properties, which result in significantly less carbon buildup, make it the ideal fuel for long-term reusability. This is the path chosen by SpaceX’s Starship, Blue Origin’s New Glenn, and Rocket Lab’s Neutron.

Conversely, the LOX-kerosene camp points to the ultimate proof of concept: SpaceX’s Merlin engines on the Falcon 9. With over 600 booster reuse flights and individual boosters flying more than 35 times, kerosene has undeniably proven its viability for commercial reuse.

Globally, these two chemistries are evolving in parallel, and even the same companies are hedging their bets. SpaceX relies on kerosene for its workhorse Falcon 9 but has pivoted to methane for the next-generation Starship. Blue Origin similarly uses liquid oxygen and liquid hydrogen for its suborbital New Shepard, but opts for methane for the heavy-lift New Glenn.

The lesson is clear: fuel selection is dictated by a firm’s specific technological base, supply-chain realities, and mission profile, rather than any inherent, universal superiority of one propellant over another.

China’s commercial launch sector is currently navigating this exact same bifurcation.

In the methane corner, LandSpace’s Zhuque-2 made history as the world’s first LOX-methane rocket to reach orbit. i-Space has also committed to methane for its Hyperbola-3, with both vehicles targeting maiden flights before the end of 2026.

Meanwhile, the kerosene camp is mounting a formidable challenge through engineering ingenuity. Galactic Energy’s Pallas-1 features the Cangqiong-50 engine, which utilizes a pintle injector to allow continuous thrust modulation between 32% and 105%. Through relentless iteration, the rocket maker has mitigated the coking (carbon buildup) issues that traditionally plague kerosene engines. By July 2026, the Cangqiong-50 had accumulated over 20,000 seconds of ground hot-fire testing, with 40 units already assembled and delivered.

Space Pioneer (天兵科技) is tackling the coking problem from a different angle: its medium-to-heavy reusable Tianlong-3 will use nine Tianhuo-12 engines on its first stage, burning coal-derived aerospace kerosene to reduce carbon deposits. CAS Space’s Lijian-2 and OrienSpace (东方空间)’s Gravity-2 have also adopted the kerosene route.

The strategic appeal of LOX-kerosene lies in its sheer thrust density. The density-specific impulse of LOX-kerosene is roughly 20% higher than that of LOX-methane, allowing rockets of the same physical dimensions to carry more propellant and deliver greater payload capacity. Furthermore, having been stress-tested by over 600 reuse flights on the Falcon 9, LOX-kerosene currently boasts the highest level of engineering maturity in the reusable launch market.

The Third Schism: Run, or Walk Before You Run

Once the propellant is chosen and the hardware is built, engineers face their most pragmatic dilemma: attempt a booster recovery on the maiden flight, or secure orbital insertion first and defer reusability for later?

History offers a cautionary tale for patience. When SpaceX launched the Falcon 9 in 2010, it merely demonstrated a reusable architecture; it lacked grid fins, landing legs, and any intention of a landing. It took three more years and five subsequent launches before the company tested re-entry burns and achieved a soft ocean splashdown. Landing legs were not added until 2014, and the first successful land-based recovery did not occur until the 16th flight in December 2015.

Blue Origin’s New Glenn similarly failed to stick its maiden landing, succeeding only on its second attempt. Rocket Lab has opted for an even more conservative trajectory with its Neutron rocket, planning an initial ocean splashdown to gather telemetry before attempting a precision land-based recovery.

Yet, a bold exception has emerged from China’s state sector. On July 10, 2026, the Long March 10B achieved a historic first: a successful sea-based net recovery of its first stage during its maiden flight. However, this was no reckless gamble. The Long March 10B is a direct derivative of the Long March 10 family, and its recovery mechanisms had already been rigorously validated during low-altitude demonstration flights of its sibling, the Long March 10A. The 10B’s success underscores a vital aerospace axiom: a maiden-flight recovery is only viable when the underlying technology has already survived exhaustive ground and subsystem testing.

China’s private launch sector is currently split over this exact calculus.

The aggressive camp is betting on catching the booster immediately. LandSpace’s Zhuque-3 successfully reached orbit on its maiden flight in December 2025, though its first-stage recovery encountered anomalies. i-Space has similarly announced plans to attempt a recovery test during the maiden flight of its Hyperbola-3 liquid rocket.

Conversely, a more pragmatic cohort is choosing to walk before they run.

CAS Space’s Lijian-2 successfully reached orbit on its maiden flight this March without any recovery hardware. The company has since announced a phased approach: it will eventually integrate reusable Liqing-2 engines and adopt a cluster-recovery architecture to gradually build reusability.

Space Pioneer’s Tianlong-3 also skipped recovery equipment on its first stage, viewing reusability as a future milestone. Tragically, this cautious approach did not save the vehicle, as the rocket failed to reach orbit during its April 2026 maiden flight.

Galactic Energy has perhaps articulated the most methodical version of this conservative strategy. Its Pallas-1 is designed from the ground up for reusability, featuring deep-throttling engines, pre-installed grid fins, and landing mechanisms. Yet, its maiden flight will not attempt a landing. Instead, the company is relying on exhaustive ground validation: its Cangqiong-50 engines have accumulated over 20,000 seconds of hot-fire testing, simulating the equivalent of 25 full flights. This prolonged testing has already exposed component degradation issues that short-duration tests would have missed—defects that, if discovered mid-flight, would have been far costlier to rectify. For Galactic Energy, a phased rollout backed by relentless ground testing is the only logical path to reusability.

This schism reflects a fundamental divergence in risk-reward calculus. Attempting a maiden-flight recovery accelerates the acquisition of flight data, but it carries a high probability of failure that can delay commercial operations. Prioritising orbital insertion first boosts the chances of a successful maiden flight and faster commercialization, but it inevitably stretches out the timeline for validating reusability.

The Fourth Schism: 3D Printing

Beyond propellant choices and recovery timelines, a quieter but equally consequential dimension of competition is emerging: the extent to which additive manufacturing, or 3D printing, is integrated into engine production.

The fundamental value proposition of 3D printing lies in consolidation. By fusing hundreds of discrete parts into a single monolithic structure, the technology drastically reduces welding and assembly requirements. Theoretically, this can slash manufacturing lead times by over 70%. Furthermore, topology optimisation allows engineers to strip away excess mass and design intricate internal flow channels, ultimately unlocking superior engine performance.

Globally, the penetration of this technology varies. Rocket Lab has been a pioneer, 3D printing the core components of its engines from day one, with hundreds of such engines now successfully powering orbital-class launches. SpaceX has also embraced the technology, incorporating 3D printing into the turbopumps and combustion chamber components of its Merlin engines, and making it a foundational manufacturing pillar for its next-generation Raptor engine.

China’s leading commercial launch firms have similarly adopted additive manufacturing across the board.

Galactic Energy extensively uses 3D printing for the thrust chamber and expansion nozzle of its Cangqiong-50 engine, as well as for its pintle injector, turbopump casing, and gas generator. This hybrid manufacturing approach has boosted engine performance by over 3%, reduced production costs by 15%, and shortened manufacturing cycles by 30%, enabling the batch delivery of more than 40 engines.

At LandSpace, over 70% of the components in its Tianque-series engines are 3D printed, with several key parts supplied by the domestic additive manufacturing specialist Bright Laser Technologies (铂力特). Similarly, i-Space has applied 3D printing to critical components of its JD-2 engine, including the turbopump and complex piping networks.

Broadly speaking, the industry is transitioning from merely proving that 3D printing works to demonstrating that it can scale. The depth of additive manufacturing integration directly dictates whether an engine can be mass-produced at lower costs and in shorter timeframes. While this metric rarely makes the front page of space news, it is quietly becoming a decisive variable in the long-term economic viability of commercial launch providers.

Conclusion: No Right Way, Only the Finish Line

From initial propulsion to fuel chemistry, from recovery pacing to manufacturing techniques, each divergence reflects the unique commercial style, capital strategies, and engineering philosophies of the firms involved.

There is no single, universally correct answer. Looking back from the vantage point of 2026, it is clear that LOX-methane pioneers have yet to achieve rapid mass production, maiden-flight recovery remains a high-stakes gamble with limited odds of success, and firms that started with liquid propulsion have had to absorb punishingly high upfront costs. The true merits and flaws of each trajectory can only be validated over a sufficiently long time horizon.

Yet, one conclusion is inescapable: the 2026–2027 period marks a critical inflection point for China’s commercial launch industry, transitioning from the fundamental question of “can we reach orbit?” to the economic imperative of “can we achieve reusability?”

A wave of liquid-fuelled rockets is making their maiden flights, recovery tests are intensifying, and the window of opportunity in capital markets has swung open. The firm that first closes the commercial loop of “recovery and reuse” will seize the commanding high ground in the next phase of the industry’s evolution.

Galactic Energy’s phased strategy, LandSpace’s audacious maiden-flight recovery challenge, Space Pioneer’s heavy-lift thrust trajectory, the offshore recovery experiments by i-Space and Space Epoch, and CAS Space’s modular architecture—each path possesses its own internal logic and its own inherent risks.

Unlike the United States, where SpaceX stands alone at the apex (with second-placed Blue Origin still a distant second), China’s commercial launch sector has yet to consolidate. It remains unclear who the ultimate victor will be—and indeed, there may be more than one. Data shows that in the first half of this year, the number of newly established enterprises in China's spacecraft and launch vehicle manufacturing sector surged by 185.7% year on year.

Yet an inescapable reality is that Chinese firms are still more than a decade behind SpaceX. Elon Musk recently noted on X that the V3 satellites slated for Starship deployment will boast communications capabilities roughly an order of magnitude greater than the current V2 satellites launched by the Falcon 9. Looking ahead, SpaceX plans to deploy V3 and next-generation Starlink satellites at scale, alongside V2 and subsequent generations of direct-to-cell satellites, expanding its overall constellation by another order of magnitude.

China’s satellite-internet ambitions—led by the state-backed GW network and Shanghai-backed SpaceSail—remain far from matching Starlink’s sheer scale. The primary challenge for these latecomers is stark: launch cadence and payload capacity per rocket are still woefully inadequate.

Historically, China has excelled at industrial scaling. But pitted against a rival as formidable as Elon Musk, the question of what scorecard China’s commercial space sector will post over the next five years remains open.

Share

Leave a comment

On August 19, Unitree Robotics became the first pure-play humanoid robot stock on China’s A-share market, soaring as much as 630% on its first day and touching a market capitalization of 444 billion yuan (62 billion USD) – nearly 800 times trailing earnings.

For a company that booked 278 million yuan in net profit last year, that valuation implies a level of future conquest that would make even the most ardent techno-optimist blush.

The celebratory headlines write themselves: “Humanoid Robot First Share” lands with a bang, locking in massive paper gains for a star-studded investor roster that includes Meituan, Tencent, Alibaba, Ant Group, Sequoia China, MPCi and Liang Wenfeng’s DeepSeek.

Yet beneath the ticker-tape euphoria, a cooler reading of the fundamentals suggests the market may be pricing not just Unitree’s present, but a sci‑fi future that remains stubbornly out of reach.

For investors accustomed to a skeptical lens, three uncomfortable questions demand answers: Is the valuation anchored in reality? Is the business model sustainable? And is “humanoid” even the right bet to begin with?

The scarcity premium – and its perils

Unitree’s spectacular debut owes much to a simple supply-demand quirk: only 7.44% of its shares are freely tradable, one of the tightest floats in STAR Market history. For context, most hard-tech STAR listings debut with 8–12% in public hands; Unitree’s float ranks among the lowest ever recorded.

That scarcity, combined with the hype around “the first true humanoid play”, drove opening-day fireworks that saw the stock hit 1,100 yuan before settling at 845 yuan – still a 460% gain. The retail lottery for IPO allocations was equally extreme: the winning rate was just 0.018%, meaning one successful bid for every 5,500 applications – roughly twenty times harder to win than the recent CXMT offering.

But scarcity cuts both ways. When the lock-ups expire – and when the inevitable profit-taking begins – the same thin liquidity that amplified the rally will magnify the fall.

History offers a cautionary tale: UBTech Robotics, listed in Hong Kong, surged 264% post-IPO only to give back most of those gains within months. In 2025, the Shenzhen-listed special-purpose acquisition vehicle Swancor Advanced Materials – a rumored backdoor listing vehicle for rival Agibot (智元) – briefly soared more than tenfold before crashing back to earth. Unitree’s own after-market performance is also like to follow a similar arc once the initial euphoria fades.

The manufacturing mirage

Strip away the acrobatic backflips, the choreographed dance routines, and the rhetoric of “embodied artificial intelligence”, and Unitree remains, for now, a hardware assembler.

In the first nine months of 2025, 73.6% of its humanoid revenue came from research and education – universities and laboratories buying robots as experimental platforms, not factories deploying them for productivity gains. Commercial consumer applications – showroom guides, stage performances, and tourism displays – accounted for another 17.4%.

Genuine industrial use, the holy grail that would justify a trillion-yuan addressable market, contributed less than 10% of humanoid sales. In fact, the company’s own prospectus reveals that the combined share of revenue from scientific research, education, and consumer entertainment actually rose from 66.6% in 2022 to 78.1% in the first three quarters of 2025. The robots are selling, but they are largely ending up on researchers’ benches and hobbyists’ desks, not on factory floors.

That is not a criticism; it is a reality check. Humanoid robots today are closer to costly science kits than to assembly-line workers. Their dexterity and reliability remain limited; their economic return for most customers is negative.

Unitree itself has slashed average unit prices from 593,000 yuan in 2023 to 166,000 yuan in 2025 – shedding more than 400,000 yuan in just two years – and competitors like Noetix Robotics (松延动力) now offer bipedal units below 10,000 yuan. Rental rates for humanoids have crashed from as high as 10,000 yuan per day to under 1,000 yuan. Price wars erode margins – Unitree’s gross margin dropped from 60.7% to 56.0% in just one year – and they signal commoditisation before scale has even been achieved.

Worse, the growth trajectory is already showing cracks. In the first half of 2026, Unitree reported revenue of 1.15 billion yuan, up 48.5% year-on-year – respectable but a sharp deceleration from previous triple-digit growth. More troubling, non-recurring net profit fell 19.3% year-on-year, with the company blaming surging R&D and marketing expenses.

The company has now posted two consecutive quarters of year-on-year profit decline. That is not the profile of a business galloping toward a trillion-yuan valuation; it is the profile of a company fighting to stay ahead in a brutally competitive, commoditising market.

The data story – more promise than proof

The bull case for Unitree rests on a seductive narrative: higher shipments generate more real-world interaction data, which in turn trains superior embodied-AI models, creating a virtuous circle.

Unitree shipped roughly 5,900 humanoids in the first half of 2026 – a solid global No. 2, but behind Agibot’s 8,400 units and 44% global market share, according to Smart Analytics Global (SAG). Just a year earlier, Unitree was still the global leader; the crown has now slipped.

Yet the quality of that data matters more than quantity. Most Unitree robots operate in controlled academic environments, not messy, high-stakes factory floors. Contrast that with Tesla’s Optimus, which learns from its own automotive production lines, or with Figure AI (Helix 02), which partners with BMW and other industrial giants. Without deep, messy, industry-specific operational data, Unitree’s “brain” may remain chronically underfed.

The company’s own prospectus concedes that revenue from industrial applications – manufacturing, logistics, inspection – remains negligible, meaning its data feedback loop lacks the most valuable signal: real-world productivity metrics.

Moreover, the embodied-AI bottleneck is not hardware – it is simulation-to-reality transfer, generalisation, and real-time adaptation. As of early 2026, the total globally available corpus of real-world physical interaction data stands at roughly 500,000 hours – less than one-twentieth of what is needed to train a truly general-purpose embodied foundation model.

The industry is still years away from solving this, and meanwhile, synthetic data and video training remain imperfect substitutes. The data moat that investors hope Unitree is building may turn out to be more of a puddle than a lake.

The form-factor fallacy – and the critics who matter

Another challenge to Unitree’s entire thesis comes not from a rival manufacturer, but from the intellectual godmother of modern AI, Fei-Fei Li (李飞飞). In a recent interview, Li argued bluntly that “the humanoid direction is misguided” (Of course, she could be wrong). Her reasoning is systematic and devastating: the human form is an evolutionary compromise, not an engineering optimum. If we needed only to climb trees, she observed, our fingers would have evolved very differently. Non-structured environments are technically the hardest challenge and should not be the starting point. The real bottleneck in physical AI is not morphology but data and simulation – and humanoid bipedal locomotion is among the least efficient ways to navigate most real-world spaces.

Li’s BEHAVIOR Robot Suite project explicitly uses a two-arm, wheeled platform with a four-degree-of-freedom torso – no bipedal legs, no human mimicry. Her core logic is simple: robot form should serve function, not narcissism. Capital efficiency will eventually break the spell of “human self-worship” and replace it with purpose-built designs.

She is not alone. Turing laureate Yann LeCun has publicly dismissed most humanoid robots as “pre-programmed gimmicks” with no practical utility, noting that “the intelligence of all current humanoids is less than that of a house cat”. K1Wave’s founding partner Cao Bin (曹斌) made an even more pointed economic argument in a recent dialogue: “There are 8 billion humans on this planet. Many tasks are already done cheaply by people. Building a humanoid to replace them is driven by capital, not productivity. As long as the marginal cost of human labor is below that of the robot, the rational choice is to use people, not build humanoids.”

If these critics are even half-right, Unitree’s entire product strategy – doubling down on bipedal, anthropomorphic design – may be building toward an architectural dead end. The industry could pivot toward specialized, task-optimized machines, leaving humanoids as a niche curiosity, much like the short-lived vogue for general-purpose home robots that never quite materialized.

The competitive landscape – and the shadow of China’s “body” trap

Unitree’s IPO arrives just as Agibot, founded by former Huawei staffs, dethroned it as the global shipment leader – a symbolic blow that underscores a broader structural risk. Chinese manufacturers excel at hardware integration and cost reduction – the “body” of the robot. The country’s supply chain for motors, reducers, sensors, and structural parts is world-class and getting cheaper by the quarter. Unitree’s ability to drive down unit prices from 593,000 to 166,000 in two years is a testament to this manufacturing muscle.

But the “brain” – foundation models, simulation software, high-end AI training chips – remains largely American-dominated. Nvidia’s GPUs, OpenAI’s and Google’s foundational research, and the simulation ecosystems developed in Silicon Valley still set the pace.

The historical verdict – early leaders might not win

The most sobering lesson for Unitree’s shareholders comes from business history. Nokia dominated the feature-phone era with a 60% global market share, only to be obliterated by the smartphone revolution. Yahoo was the undisputed king of the early internet, yet it sold its core assets for a pittance. General Motors, once synonymous with the automobile industry, filed for bankruptcy in 2009. Even iRobot, the pioneer of consumer robotics, held over 80% of the US vacuum-robot market before losing its way and delisting from Nasdaq after three consecutive years of revenue decline.

The pattern is consistent across technology upheavals: early leaders are rarely the final victors when the underlying technology, customer needs, and business models are all in flux.

As China Merchants Securities (招商证券) noted in a recent research note, the humanoid industry today is roughly analogous to the autonomous-vehicle market in 2015–2016 – a moment when the technology was just about “usable” in a handful of pilot projects, but still far from delivering clear economic returns to customers.

The eventual winners in electric vehicles – Tesla, BYD, and a handful of others – were not the early darlings. Many of the first-movers are now forgotten or bankrupt. The same destiny might await most of today’s humanoid pioneers, Unitree included. (Note: To be honest, Unitree wasn't the first. Dataa Robotics actually started it, but they've lost their momentum.)

Conclusion: a beacon, not a bible

Unitree’s listing does provide a vital service: it establishes the first public valuation benchmark for a nascent industry. That “anchor” will shape financing rounds, mergers and acquisitions, and investor expectations for years to come. But anchors can drag as well as steady. At 440 billion yuan, the market is pricing Unitree as if it will not only maintain its lead but also capture a disproportionate share of a multi-trillion-yuan market that remains entirely hypothetical.

The more sober-minded forecast from China Construction Bank International (建银国际) – a 109 billion yuan fair value based on 2026 sales – looks almost quaint by comparison.

For now, Unitree is a remarkable Chinese success story in engineering, brand-building, and go-to-market speed. It has built a thriving business selling sophisticated robots to the world’s research community.

But the gap between a splendid prototype and a profitable, scalable industrial franchise remains vast. Investors who treat its IPO as a coronation rather than a milestone may find that the champagne, like the stock, goes flat sooner than they expect.

The real race – for commercially viable, economically justified, general-purpose robotics – has barely begun. And in that race, the first runner out of the blocks is rarely the one who crosses the finish line first.

Share

Leave a comment

China’s private rocket company’s Zhuque-3 (ZQ-3), often dubbed the “China’s Falcon 9,” successfully recovered its first stage booster on its second attempt. It might not be as explosive as a first-try success, and it missed out on the glory of being China’s very first booster recovery, but it’s still a seriously impressive achievement. It is land-based, and in a Falcon-9 style, standing on the pad with legs.

Just like the Falcon 9, it uses landing legs for recovery (which visually looks like the launch sequence played in reverse). That said, the Falcon 9 actually had way more early failures before getting it right.

The video clips shows that two minutes and some seconds after the launch, the first-stage booster successfully separated from the second-stage rocket carrying a satellite. Less than ten minutes after launch, the booster, having undergone two retro-propulsion decelerations, slowly landed at a recovery site in Minqin County, Gansu Province, about 390 kilometers away from the launch site.

The size of this rocket are comparable to those of SpaceX’s current workhorse, the Falcon 9, and its overall technological approach is, in a sense, similar to that of Starship, utilizing a stainless steel rocket body and liquid oxygen-methane engines, which could enable it to achieve even lower launch costs than the Falcon 9.

The Zhuque-3 employs a landing-leg recovery method that slightly sacrifices payload capacity, allowing the rockets to be refurbished and relaunched in a shorter time frame, as engineers only need to maintain the rocket itself without having to repair ground recovery equipment.

Aug. 19’s success came after a recovery failure last year.

Now, Commercial spaceflight is currently one of the hottest technology sectors. At present, China has more than 20 rocket companies and over 80 rocket models.

When LandSpace (蓝箭航天) was founded in June 2015, SpaceX’s Falcon 9 had not yet successfully achieved reusable recovery, and China’s aerospace industry had just opened up to private capital. As one investor put it, the entire industry had “nothing but opportunities.”

Today, LandSpace has over 1,200 employees. The Shanghai Stock Exchange accepted the IPO application of LandSpace Technology on December 31, 2025.

LandSpace’s rise can be attributed to its early and steadfast investment, as well as its highly self-reliant development model. LandSpace founder Zhang Changwu recognized early on that key subsystems of a rocket “cannot be bought off the shelf.”

LandSpace began developing its own engines in 2017 and was also one of the first companies in the industry to build its own test stands, launch sites, and rocket factories. This has allowed LandSpace to master most of its critical technologies and iterate rapidly in-house, without relying on external supply chains.

The well-known Chinese blog LatePost (晚点) conducted an exclusive interview with Dong Kai (董锴), deputy chief designer of the Zhuque-3.

Dong graduated from Harbin Institute of Technology with a major in aircraft design. He originally thought he would enter the military-industrial system and become one of those who “work in obscurity.”

In 2009, he joined the General Design Department of the Rocket Research Institute at China Aerospace Science and Technology Corporation (CASC, 中国航天科技集团) as an engineer. During that period, he received comprehensive engineering training and also participated in the development of the Long March 5, then China’s largest-thrust rocket.

In 2021, he joined LandSpace and was involved in the entire design and development process of the Zhuque-3.

In Dong’s account, there were no dramatic twists or sci-fi tales of overtaking competitors on the Zhuque-3 project—they simply followed the plan and got the job done.

Subscribe now

Below is the transcript of the exclusive interview.

Recovery Is Only the First Step

**LatePost:** After the previous recovery failure, what improvements did you make before this launch?

**Dong Kai:**

First, we sacrificed some payload weight to reinforce the thermal protection structure. Second, we added gas cylinders inside the tail compartment to purge the air and lower methane concentration before ignition.

Last time, the failure occurred during the landing ignition phase. According to the plan, the center engine would ignite first, followed by up to four others depending on conditions to slow the descent, and finally the engines would be throttled down from five to one as the landing legs deployed.

Last time, the anomaly happened in the tail compartment right after the center engine ignited but before the other four fired.

**LatePost:** How much time elapsed between the first and second launches, and what did you do in between?

**Dong Kai:**

The launch was on December 3. By the next day, we had almost all the telemetry and optical data. We didn’t have much time to celebrate or dwell on the disappointment. By December 5, we had already begun the technical fault review for the maiden flight.

The aerospace industry follows a rigorous methodology – what we call “presumption of guilt” combined with “no conclusion from a single piece of evidence.” For example, if an explosion occurs in the tail compartment, there must be an open flame, combustion, and propellant involved. Okay, where did the propellant come from? Where did it leak? Following this chain of reasoning, you break it down into a set of root-level events.

Every risk that is ruled out must be supported by objective evidence. You can’t dismiss a possibility based on subjective judgment. The process is like criminal investigation – eliminate all impossibilities, and whatever remains, no matter how implausible, must be the truth. There’s only one truth – just like in *Detective Conan*.

**LatePost:** How long did that take?

**Dong Kai:**

About two to three months. For us, a rocket is a significant investment. If we let the real culprit go, that tuition fee would have been wasted.

Once we identified the fault, we moved to improvements. The engineering side is relatively straightforward – the rest is about the cost in time and materials to fix it. In the end, these are all solvable engineering problems. We’re not the type to invent hardships for ourselves – we don’t show off. We choose the most achievable path.

**LatePost:** A key logic in commercial space is to collect data through successive tests and improve. Did the data collection from the maiden flight meet expectations?

**Dong Kai:**

Yes, it did. The key is to see whether the actual flight data matches the expectations from our ground simulations and analyses. Sometimes a successful flight could just mean the design margin was large enough to absorb errors, not that our predictions were accurate.

We did see some data that deviated from expectations. When we spot deviations, we adjust the ground models – subtracting where necessary, adding where needed. For instance, we changed the recovery ignition from five engines to three based on the data we gathered.

There’s a saying in aerospace: “If not necessary, do not add entities.” (Note: Occam’s razor theory) We initially went with five engines because another vehicle had experienced insufficient aerodynamic braking, requiring more engine thrust to compensate. After the maiden flight, I realized that risk wasn’t as significant – three engines could bring it back. Adding two extra engines would only introduce additional failure points, which could end up costing us more than we gain. Every plan is a trade-off between benefit and cost.

**LatePost:** Now that recovery has succeeded, what’s next?

**Dong Kai:**

We’re aiming to reuse the recovered rocket as soon as possible.

After this rocket returns, we’ll disassemble it very carefully to see where damage occurred after this flight. If each component has a “health bar,” our future work is to make sure every health bar is measurable, and repairable – and after repairs, we put it back on the launch pad.

The main improvement areas for Zhuque-3 are increasing payload capacity, enhancing reliability, and shortening the reuse cycle. My personal aspiration is to get Zhuque-3 past the threshold of 10 launches per year.

**LatePost:** How soon do you plan to reuse it?

**Dong Kai:**

Given LandSpace’s stainless steel production approach and engine manufacturing capacity, on average we can build a Zhuque-3 in dozens of days. I want to shorten the recovery-reuse cycle to less than the build time – the process of bringing the rocket back, doing adjustments, and sending it up again should take less time than building a new one. If it takes four or five months to refurbish, we might as well build a new one.

In gaming terms, any spell that accelerates or rewinds time is a rare item. If I build 10 rockets and shorten the reuse cycle to two months, that’s 10 times 6 – 60 launches a year. If I cut the reuse time to one month, our launch capacity doubles again.

**LatePost:** Compared to Falcon 9, where does Zhuque-3 stand?

**Dong Kai:**

On a scale of 1 to 9, LandSpace is only at Level 6, which is laboratory validation. We’ve just entered Level 7 – actual flight validation.

It gets harder the higher you go. Level 8 would be marked by the ability to fly again. Level 9 – the highest – is scaled, routine launches with extremely high success rates.

If you’re familiar with *A Record of a Mortal’s Journey to Immortality*, if Falcon 9 is a master, then Zhuque-3 is merely at its foundational stage. Once we can reuse it, we’ll reach a higher level.

**LatePost:** Were there any regrets from the design to the launch of Zhuque-3?

**Dong Kai:**

Not many in terms of product design. The maiden flight result was less than perfect, but I wouldn’t call it a regret. The only regret is that the maiden flight came a bit later than hoped.

**LatePost:** What’s the bottleneck for LandSpace’s next phase of launch capability growth?

**Dong Kai:**

With LandSpace’s current launch pads and production capacity, we haven’t hit the ceiling yet. LandSpace can build 20 rockets a year.

Of course, this is also related to launch requirements and the current strictness of launch permit approvals. In the early stages of exploration, everyone tends to be more cautious. Even Musk has complained about the FAA – though now the FAA treats him fairly well.

**LatePost:** Under what conditions would policy become more relaxed?

**Dong Kai:**

Some things work on the principle of “seeing is believing.”

So if you ask me what the bottleneck is, I’d say: we haven’t launched Zhuque-3 enough times yet. Our reliability data isn’t comprehensive enough to put everyone completely at ease. We just need more practice – more launches, higher success rates, and things will naturally accelerate.

Zhuque-3’s Core Tech Is All In-House

**LatePost:** What was the decision logic behind building Zhuque-3’s body with stainless steel?

**Dong Kai:**

To solve manufacturing challenges.

I worked on the Long March 5. When the rocket body went from 3.35 meters to 5 meters in diameter, we encountered many difficulties. We wondered if there was a simpler solution – and stainless steel was the answer.

First, the welding process is easier than with thin-walled aluminum alloys. Thin-walled aluminum pressure vessels are almost exclusively used in aerospace – few other industries use this technology, so the supply chain is limited. With stainless steel, the supply chain is abundant.

Second, it’s cheap – around tens of thousands of yuan per ton.

Third, it has good heat resistance, eliminating the need for an additional thermal insulation layer. That’s not just about cost and weight – with insulation, we’d have to maintain it after every recovery, adding significant workload and demanding high repair standards.

Fourth, it offers scalability – stainless steel can be made very thin.

**LatePost:** SpaceX announced in 2018 that it would use stainless steel for its rockets. If stainless steel is so good, why didn’t anyone use it before?

**Dong Kai:**

LandSpace had considered it too. Back then, the aluminum alloy supply chain was too limited. Because of its narrow application range, aluminum alloy supply and production were almost entirely concentrated within the state aerospace system. We could have developed it ourselves, but that would have meant retreading the same path others had already walked. From a timeline perspective, that wasn’t the optimal solution.

Also, in those early days, the environment wasn’t as favorable – you had to first prove “I’m capable, I can do this.” Some of our choices were forced by circumstance.

Plus, people thought stainless steel was too heavy – it has higher strength than aluminum alloys but also higher density, so it ends up heavier.

In fact, based on material strength, you only need millimeter-thick rocket walls. But at that thickness, production becomes problematic – welding causes distortion. A 20-meter-long stainless steel cylinder might be round at the front during welding but end up oval at the back, because distortion accumulates. After too many welds, it no longer meets the tolerance requirements for docking with other segments.

Welding itself is simple. Plenty of manufacturers in the Yangtze River Delta and Beijing-Tianjin-Hebei region that build storage tanks could do it. But they can’t weld materials this thin, nor can they maintain the geometric tolerances at this scale.

**LatePost:** So how did you solve the welding problem?

**Dong Kai:**

We focused on process development, using laser welding – the process with the lowest heat input – for the job.

Thin-wall welding is only an issue at a 4.5-meter diameter. If you scale up to 9 or 10 meters – like Starship – the strength requirements increase, wall thickness goes above 2 millimeters, and welding becomes less challenging.

LandSpace has a department called Tianma Laboratory (天马实验室) – positioned similarly to Lockheed Martin’s Skunk Works (one of the world’s most famous secret advanced R&D facilities, which developed the F-22 Raptor). It’s our elite team dedicated to solving stainless steel process challenges. We now use laser welding, essentially automated.

But we didn’t start with automation – we accumulated knowledge through repeated trials, observing how much 2-millimeter material would deform and where, then compensating in the robot’s path planning. We also studied the material itself – not just deformation but original properties like temperature during welding. Through extensive time and data accumulation, we gradually worked down from 2 millimeters to thinner gauges, eventually achieving minimal error and good yield rates at very thin specifications, then froze the data and parameters.

Every bit of thickness reduction directly translates into improved payload performance.

**LatePost:** So the welding improvements happened almost entirely in LandSpace’s own labs, without suppliers?

**Dong Kai:**

At present, the entire automated welding production for the rocket is developed by LandSpace itself. We don’t reinvent the wheel where suppliers can do the job, but rocket welding isn’t something the outsourced supply chain excels at.

In the early stage, when we wanted to validate stainless steel strength, we found welding workers in the Yangtze River Delta and gave them 2-millimeter sheets to try, regardless of technique – manual argon arc welding, whatever. That only solved the “whether we can do it” question. To truly compete on performance, we had to rely on ourselves.

Also, with outsourced supply chains, quality control is harder to manage. If the pressure vessel fails, the consequences are disastrous. Our rocket body has strict requirements for shape retention and error control.

**LatePost:** When you approached steel mill workers in the Yangtze River Delta to weld a rocket, how did they react?

**Dong Kai:**

We told them the diameter and volume – they said, okay, no problem.

You have to trust the ingenuity of the private supply chain – and their cost control is remarkable. But when we said the thickness must not exceed a certain limit, with specific upper and lower bounds, plus geometric tolerance requirements for the tank surfaces, alignment of segments, hole parallelism, and roundness – they said, “We’ve never seen requirements like this.”

Normally, their customers worry about them cutting corners and want thicker welds. Welding thin is not their specialty. Few other steel-using industries have such demands, and they have no reason to invest in such difficult processes just for rockets.

**LatePost:** Is the stainless steel you use special, or ordinary?

**Dong Kai:**

It’s off-the-shelf material – tens of thousands of yuan per ton.

Though we only produce small quantities, we treat the rocket as an industrial product, not a luxury item. We don’t need to select a highly customized material.

**LatePost:** After switching to stainless steel, which other parts of the rocket needed modification or redesign?

**Dong Kai:**

Few major hardware changes. We reconsidered some things at the system level – since we were using stainless steel, we pushed it to its limits. For example, in the return strategy, with greater structural strength, we could choose more aggressive trajectories with higher dynamic pressure.

If switching to stainless steel meant adding burdens elsewhere, that would indicate a fundamental flaw in the approach. We don’t have “unlimited ammunition” – technical choices are made to solve problems, not to invent challenges to prove ourselves.

**LatePost:** Not having “unlimited ammunition” – how does that change things for a commercial space company?

**Dong Kai:**

Commercial space companies are very cautious when choosing rocket diameters, because different diameters require different infrastructure investments.

The state aerospace system doesn’t have to worry about this – they can pick the optimal data, whether 4.37 meters or 7.5 meters, and the launch site and ground support will accommodate it.

Commercial space companies think about reducing unnecessary investment. We only choose diameters that the state programs have already used – 4.2 meters, 3.35 meters – so we can use existing ground equipment and tooling interfaces. If not necessary, do not add entities.

**LatePost:** How did you complete the entire rocket in 28 months?

**Dong Kai:**

There was preparatory work. The Tianque 12A engine used on Zhuque-3 was developed during the Zhuque-2 era. And the process validation for Zhuque-3 started with test articles as early as late 2022. These accumulations and trial-and-error helped us avoid repeating mistakes.

In engineering, there are no shortcuts – only fewer detours.

We never emphasized mandatory overtime. Instead, we clearly communicated project milestones across all departments. In a highly motivated organization, when everyone else is pushing toward deadlines, how can you tolerate falling behind? That’s basically how those 28 months went.

How to Manage an Aerospace Project

**LatePost:** How do you break down the challenge of launching something the size and weight of a skyscraper into space and bringing it back safely?

**Dong Kai:**

Project decomposition is organized around major ground tests – modal tests, power system tests, and so on. Each test has corresponding requirements for each subsystem, specifying what state must be achieved by when. It’s like a military operation – each unit must arrive early, never late.

**LatePost:** So each subsystem has autonomy, as long as it meets its targets on time?

**Dong Kai:**

The top commander doesn’t need to micromanage down to the platoon level. As the army group, I tell each division where and when to be – the rest is up to the division commander. This is a key feature of LandSpace’s management – full delegation of authority.

**LatePost:** Is that how it works inside China’s state aerospace system?

**Dong Kai:**

I don’t think it’s a matter of inside or outside the state system – it depends on the rocket program. If a program faces strong external competition and must be completed by a certain deadline, of course you delegate.

The traditional approach does tend to be more granular – even specifying when certain documents should be completed. This places very high demands on the planning department. We know we’re not yet capable of that level of control, so we delegate some autonomy.

**LatePost:** You said the ideal state is to let people be self-driven. That’s a high bar, and rockets aren’t exactly an area where success comes easily.

**Dong Kai:**

This kind of highly motivated organizational culture isn’t achieved through lectures or indoctrination.

First, through systems and institutions. LandSpace advocates “pragmatic innovation and decisive execution,” encourages trial and error, and doesn’t make people worry that “doing more means making more mistakes.” Rewards and consequences are clearly defined.

Second, by providing real-world opportunities. In any organization, no matter the size, the truly effective force is the top 20–30%. These people are often driven not just by compensation but by a sense of achievement and self-actualization. LandSpace happens to provide a platform where they can keep getting real combat experience and building that sense of accomplishment.

Think about it – what frustrates a highly skilled engineer the most? Spending all day on PowerPoint presentations and never having time to do what truly matters. We give them a stage.

That sense of honor is built through battle after battle. I believe this logic applies to most knowledge-based work.

**LatePost:** Rocket engineers need backgrounds in physics and materials – but now many AI companies are also recruiting these brightest minds. How do you attract top talent?

**Dong Kai:**

Our definition of talent is somewhat different. LandSpace wants “moldable talent.” In an industry as innovative as commercial space, I don’t think what people have learned before counts for that much – it just means they can help us avoid a few detours. What we value more is growth potential and learning ability.

The prerequisite is shared values. People who choose aerospace more or less have a sense of national mission. We offer compensation that removes the need for them to weigh personal sacrifice – not “I chose aerospace, so I have to endure poverty.”

**LatePost:** You mentioned that many in aerospace have deep passion and aspiration for this field. How did your own feelings for aerospace develop?

**Dong Kai:**

I first attended a rocket maiden flight in 2016 – the Long March 5. At that moment, nothing else mattered – that was enough. The moment a rocket lifts off is an immense emotional shock for everyone who built it.

Many people love this from childhood. For me, it was in college – when Shenzhou V launched in 2003 and Yang Liwei went to space – that I suddenly felt this industry was truly noble.

I graduated and started working in 2009, just in time for the 60th anniversary of the founding of the PRC. Whether it was learning patriotic songs or the early orientation training, there was a strong patriotic atmosphere.

When you enter this industry, you have to give up a lot because of secrecy requirements – maybe not as extreme as working on weapons, but there are restrictions. Back then, one of the songs we learned was *The Motherland Will Not Forget*. It felt different from what we learned in middle school – it gave you a different feeling. When you’re surrounded by that atmosphere, you can’t help but be influenced and genuinely believe what you’re doing is noble.

Even now, I still feel that way. I can’t explain it, but I feel like I carry a historical mission. From entering the大院 (compound) to leaving it and moving into commercial space – it’s all driven by that original aspiration.

**LatePost:** What’s the most important thing your experience at CASC gave you?

**Dong Kai:**

Engineering discipline training. I was lucky – right after graduation, I joined the Long March 5 and the new-generation rocket program. New projects are great for professional growth.

The commanders I worked with back then – I still admire them. It’s like that scene in *The Decisive Engagement* where the commander says, “Liu Yalou, take note – make the following deployments” – you feel you’re watching a top-level commander. I’ve worked with people like that, and now I unconsciously emulate their approach. That’s the 50-year legacy of China’s aerospace program.

Sometimes I see press portrayals of the state teams as competitors to be disparaged. I think that’s a huge misunderstanding – it greatly underestimates China’s decades of aerospace accumulation and its organizational capabilities.

**LatePost:** What aspects of the national teams do you think are underestimated?

**Dong Kai:**

First, they always have to control the floor – the minimum acceptable outcome. That’s a different logic from commercial space, which takes risks in pursuit of higher ceilings. It’s like why you go to a top-tier hospital instead of a street clinic – the clinic might advertise “we once cured a terminal case,” pursuing the best possible outcome, while you go to a major hospital because its baseline is reliable. The same logic applies to state aerospace decision-making.

Second, all martial arts come from Shaolin Temple (天下武功出少林). China’s commercial space industry’s true “Whampoa Military Academy” is the state system – most of our early employees came from there. If you consider the state system as part of China’s commercial space landscape, it’s the flagship, bearing the burden of trial and error. We’re taking fewer detours today because they’ve already blazed the trail.

**LatePost:** Looking at SpaceX’s history, many of their early efforts were things NASA had already done – just done more cheaply. But eventually there’s a gear shift – they also have to take on the role of leader.

**Dong Kai:**

Yes, exactly. On a broader scale, SpaceX is the big brother of global commercial space now.

Musk Has Proved It Can Be Done, And We Can Too

**LatePost:** Over the past two decades, every aspect of aerospace technology has changed – stainless steel materials, engine technology. If you take a systematic view, what are the major shifts you see in the field?

**Dong Kai:**

In 2009, China’s aerospace was always talking about catching up with the world. Today, if not for SpaceX, China would be number one – the state program has surpassed NASA. The United Launch Alliance (the Lockheed Martin-Boeing joint venture serving the U.S. military and NASA), once considered unassailable, has transformed dramatically. Who would have thought Boeing’s “space black shop” would also face setbacks?

Second, the development of heavy-lift rockets has exceeded expectations. Back then, people felt building a heavy-lift rocket was a lifelong endeavor. For Long March 5, if you count from feasibility studies, it started in 1986, engine development in 1996, project approval in 2006, and launch in 2016 – forty years in total. When I first joined the Long March 5 program, senior colleagues told me, “You’re lucky – you get to work on a new rocket,” meaning I’d have a chance to go into battle, like a soldier finally getting to fight a major war.

In retrospect, the changes from 2016 to 2018 were dramatic. SpaceX went from Falcon 9 to Falcon Heavy in just over two years. For me, that was an even bigger shock than watching Falcon 9 land. And Starship even more so.

Third, there are now over 80 rockets awaiting launch in China – named and unnamed. In the past, the Long March series had just over a dozen. For those of us in the industry, it feels like the spring of commercial space has arrived.

When I first started working on rockets, I thought maybe after Long March 5, if I was lucky, I’d get to work on heavy-lift rockets. Now the landscape has changed entirely – the entire global aerospace industry has a new vitality.

When Musk first talked about launching 42,000 satellites, we all thought that number was inflated. That became a lesson for me – never make judgments too quickly.

We’ve always had respect for SpaceX – not for the grand narratives, but for the fact that after three Falcon 1 failures, they dared to organize a fourth launch – “They can do that? They dare to do that?” After that, it seemed like everyone had a psychological anchor: I can fail three times.

**LatePost:** Beyond Zhuque-3, what are LandSpace’s next technology priorities?

**Dong Kai:**

The core priority for Zhuque-3 going forward is engine upgrades to improve payload. The Tianque 12B engine should be delivered around the end of this year or early next year. Meanwhile, LandSpace is also developing the Lanyan full-flow engine, which has undergone many test firings, but it will take considerably longer to integrate into the full rocket.

Larger-scale rockets are still in the technology demonstration phase. Project timing will depend on market demand and also on how Zhuque-3 operations perform. At this stage, our priority is to increase Zhuque-3’s launch frequency to 30–50 times a year.

**LatePost:** That number doesn’t sound very aggressive.

**Dong Kai:**

In 2023 and 2024, the total launch volume for all of China was about 160 tons per year. What does that mean for Zhuque-3? In reusable mode, that’s just over a dozen launches.

**LatePost:** Historically, even the U.S. state teams couldn’t afford failures. Now people think commercial space has more room for trial and error. But rockets are still expensive – how does LandSpace balance experimentation with safety and success?

**Dong Kai:**

It’s not that contradictory. First, we absolutely do not tolerate management or process risks. Second, we have systems in place to ensure personnel safety – unmanned operations, automation, remote control.

**LatePost:** But as a commercial company, you still have to control costs.

**Dong Kai:**

Our core reason for cost control is survival – not improving gross margin by a few percentage points on a financial statement. We don’t calculate how much margin increases if we cut costs from 50 million to 30 million. Rather, if we save that 20 million, we can run more tests and have more opportunities to iterate. There are red lines and bottom lines we cannot cross in design decisions – beyond that, we cut unnecessary or non-essential expenses.

**LatePost:** Still sounds difficult. What’s the fundamental advantage of the commercial approach?

**Dong Kai:**

Greater supply chain flexibility.

The state system has strict quality control requirements for its suppliers – they use supplier certainty to counter the uncertainty of new products.

The state team’s supply chain is constrained not only by economic indicators but also by various regulatory requirements. It’s not just a straightforward transaction – if something goes wrong, accountability must be traced back to individuals. This model raises costs and extends timelines, but it keeps the floor under control.

Using market-based supply chains gives us access to cheaper components – and “cheaper” doesn’t mean squeezing supplier margins. It’s because China is a major industrial power with many highly competitive sectors, good efficiency, and high yield rates. In the trade-off between cost and reliability, we can be bolder. But higher flexibility is both an advantage and a disadvantage – controlling the floor can also be more difficult.

This places greater demands on our quality control. We can’t just issue a statement of work to a supplier and accept their word that “it’s done.” We need our own ways to verify and test whether suppliers actually meet the standards, rather than building reliability on their commercial credibility. In the past, rocket designers signed off on documents and their job was done. Now we’re also responsible for outcomes. Comparing the two models, commercial space places higher demands on technical personnel.

**LatePost:** In terms of organization, personnel, and collaboration, how different is it from your time in the state system?

**Dong Kai:**

It looks similar on the surface, but fundamentally quite different.

Ultimately, it’s about quality control. In the state system, the organizational philosophy is to assign dedicated personnel to every task, break systems down into finer pieces, control the floor, and ensure stability even with staff turnover. The cost is increased system complexity.

From an engineer’s perspective, almost all quality issues arise at handoff points – just like in warfare, breakthroughs happen at the seams between two units. If a unit fights alone, it knows its responsibility and won’t let the enemy through. The traditional system relies on more robust procedures to ensure handoffs go smoothly.

The result: the state aerospace system has dozens of times more people than us. LandSpace, with 1,000–2,000 employees, is considered large in commercial space – but they have tens of thousands, yet their efficiency isn’t 10 times greater, because a significant portion of their workforce is dedicated to managing the risks of organizational complexity. They’ve found their balance.

Commercial space inherits this system but reduces complexity. We keep the critical architecture and the principle of each unit guarding its own territory, but we don’t divide personnel too finely – we maintain higher integration. A task that might require five or six handoff stages could be handled by a single team that bridges them all, potentially reducing handoff risks. Increasing individual responsibility may increase risk, but it can also reduce it. It’s all about balance at different stages of organizational development.

Our COO said late last year that LandSpace will remain a startup for a long time – solving quality issues and reducing management complexity remain key tasks.

**LatePost:** As commercial space becomes more visible to the public in recent years, have the types of talent you attract changed significantly?

**Dong Kai:**

Commercial space companies tend to do more experienced hiring. The density of fresh graduates with top talent isn’t as high as in the state system.

New graduates benefit from the structured training in the state aerospace system. We do have development programs, but not as comprehensive as “Shaolin Temple” – we tend to throw people into the battlefield.

For experienced hires, we’ve also brought in people from other industries, like automotive. Traditional engineers aren’t as cost-sensitive – they assume things should cost what they cost. But people from automotive tell us, “In our industry, this doesn’t cost nearly that much.”

**LatePost:** Can you give an example?

**Dong Kai:**

Take ambient temperature sensors for rockets – they cost tens to hundreds of yuan each. People from automotive say, “We use these for a few yuan – how can you be charging so much?”

**LatePost:** Doesn’t aerospace require specific certification standards for all components, which drives up cost?

**Dong Kai:**

We certify based on results. For components where we can control the outcome, the validation process is faster. I even think this is likely the trend for commercial space. Some things may be called “automotive-grade” on the ground – but if they’ve flown to space and proven themselves, I’d say calling them “space-grade” isn’t a stretch.

**LatePost:** When did you start believing that a private company building a reusable rocket was definitely achievable?

**Dong Kai:**

We believed China’s reusable rockets would definitely succeed.

From a technical logic standpoint, LandSpace had accumulated power system experience with Zhuque-2. The Zhuque-3 maiden flight was also successful. The remaining critical technology was the 3.3-kilometer recovery phase. Given the technical conditions at the time, we couldn’t guarantee 100% success, but we couldn’t see any rigid technical obstacles in any part of the process.

But really, you don’t need to overthink the final outcome. If we’re talking about operational costs and results, we trust our founder Zhang Changwu’s judgment. If he’s bold enough to take the bet, we don’t need to overcomplicate things – we just focus on the engineering, make the most of limited resources, and maximize our chances.

However hard it gets, it can’t be harder than Musk’s early days – even Armstrong said it wouldn’t work. That’s the significance of pioneers – they prove something can be done. Since he can, so can we.

Share

Leave a comment

Washington is set to force a stark choice on 35 Pax Silica and AI Opportunity Statement signatories: align with the U.S. on AI or sign up with China's WAICO initiative, believed to be China’s competing framework. Those opting for Beijing are threatened to be shut out of the American-led coalition, according to a U.S. official and an internal document seen by Reuters.

Last year, Washington launched the Pax Silica initiative aimed at securing supply chains for AI models, semiconductors and critical minerals, amid a fierce technology rivalry with Beijing.

“Pax” comes from Latin, meaning peace, stability, and long-term prosperity. “Silica” refers to silicon’s precursor (silicon dioxide) — the base material for semiconductors and many electronic components essential to AI.

The initial partners under Pax Silica were Australia, Finland, Japan, Qatar, India, Singapore, Korea and Israel. In June this year, a total of 35 nations joined Pax Silica, the U.S.-led initiative focused on securing supply chain for advanced technologies such as semiconductors, AI and rare earth elements.

List of the 35 Nations That Signed the Joint Statement on Partnership for AI Opportunities:

Over on the other side of the Pacific, all eyes were on Shanghai from July 17 to 20, as the city hosted the 2026 World AI Conference and High-Level Meeting on Global AI Governance. Just before the event kicked off, 29 countries inked the World AI Cooperation Organization (WAICO) charter, officially launching the world’s first AI-focused intergovernmental organization — with Shanghai as its permanent headquarters.

The 29 countries that signed the agreement to establish the World Artificial Intelligence Cooperation Organization (WAICO) are founding members:

Reactions have been starkly divided. On one side, delegates from Algeria, Venezuela, Cameroon, and even Italy have offered effusive praise for China's open-door approach and its ambitions to lead the global AI agenda.

Yet critics have been equally vocal, with some Western outlets pointing to a conspicuous absence—"no one from Silicon Valley" attended World Artificial Intelligence Conference (WAIC)—and noting that WAICO counts no Western nations among its founding members, raising doubts about its clout on the world stage.

Apparently, for now, the initiative's backers remain largely within Beijing's orbit: Italy's words of support stopped short of a signature, and a closer look at the membership lists reveals that Kazakhstan is the only country straddling both camps—a founder of WAICO and a signatory to Washington's rival Pax Silica coalition.

(Note: The disparity in commitment is telling: Kazakhstan's accession to Pax Silica was formalized by Deputy Prime Minister Zhaslan Madiyev, while President Kassym-Jomart Tokayev himself signed the WAICO agreement during a July 16 state visit to China.)

What China put on the table at this year's WAIC was not a breakthrough model or a next-generation semiconductor. Its real play was two messages: open-source and inclusive. That messaging, in turn, signaled the quiet launch of something larger—an "open-source Belt and Road" designed to rally the Global South around Beijing's AI vision.

China has, for the most part, chosen to maintain a low-key stance (韬光养晦) in its WAICO agenda, insisting that placing a bet on Beijing does not mean placing one against Washington. Yet the establishment of WAICO has still triggered an immediate U.S. backlash—a dynamic that forms the essential backdrop to the internal draft obtained by Reuters.

That said, if we do view through the lens of an AI-era rivalry, one question looms large and unresolved: five years down the road, which bloc will come out on top—the Pax Silica bloc marshaled by Washington, or Beijing’s WAICO?

Subscribe now

In a recent GBA Review op-ed, CUHK-Shenzhen public policy professor Huang Ping argues that while the U.S. may hold its Western allies in line, the Global South remains out of reach. The two camps, he says, are essentially competing for the same prize: middle powers that have yet to choose sides. Here is a summary of the key takeaways from this article with some adaptations:

The “AIIB for AI”: AI Is Not Just Technology—It’s the Next Infrastructure

To grasp the significance of WAICO, one must first recognize a fundamental premise: artificial intelligence is not merely a technology or an industry—it is the infrastructure of the future, much like railways, highways, electricity, and telecommunications in their time.

This is a reality unfolding before us. Railways reshaped the movement of people and goods; electricity reorganized industrial production; telecommunications transformed the dissemination of information. AI, in turn, is reshaping the very processes of knowledge production, decision-making, and value creation. Once AI becomes embedded in a country’s economic and social systems as infrastructure, it ceases to be a mere “application” that can be swapped out at will.

Notably, the infrastructure lock-in doesn’t occur at the product level. Electricity locks you in through grids and standards, not through any specific generator. The same applies to AI—product-level lock-in can be undone by open-source alternatives, but embedding at the ecosystem and standards level creates a truly lasting relationship. Understanding this distinction is key to grasping the novelty of China’s latest move.

WAICO’s True Precedent Is Not the Belt and Road Forum—It’s the AIIB.

This year’s WAIC was fundamentally different from its predecessors. Over the past eight years, WAIC had been an industry gathering centered on exhibitions, forums, and business matchmaking. But in 2026, the conference saw China’s top leader attending the opening ceremony and delivering a keynote address. More significantly, the WAICO agreement was formally signed on the sidelines, establishing the world’s first intergovernmental organization dedicated to AI. This was no longer an industry event—it was a diplomatic act of institutional creation.

If we look for a precise parallel in China’s recent multilateral practice, WAICO is not the Belt and Road Forum (which is a summit mechanism, not an international organization)—it is the Asian Infrastructure Investment Bank (AIIB) of 2015.

Like the AIIB, WAICO is a Chinese initiative with a formal charter, a permanent secretariat, membership status, and a permanent headquarters. Both represent the pivotal leap from fragmented bilateral cooperation to a standing multilateral institution.

But the most important lesson the AIIB offers WAICO lies not in institutional design, but in membership composition. In March 2015, the United Kingdom defied explicit U.S. opposition to become the first Western nation to apply for AIIB membership, with Germany, France, and Italy quickly following suit.

Of the 57 founding members, a significant number were core U.S. allies. The AIIB never asked any country to choose sides between China and the U.S. WAICO will almost surely chart the same course.

The Open-Source “Belt and Road”

If the Belt and Road Initiative (BRI) was the globalization of infrastructure, then WAICO marks the launch of an “AI Belt and Road.” And the fundamental difference from its traditional counterpart is this: the “AI Belt and Road” China champions is open-source.

The original BRI was essentially a “hard connectivity” model: railways, highways, bridges, and ports were built through Chinese investment, Chinese construction, and delivered as turnkey projects. Once a bridge was completed or a road paved, the Chinese teams would phase out, with operational rights gradually handed over or maintained through concession agreements.

The logic of open-source AI is entirely different. An open-source model delivers not a finished product, but the method itself. What the recipient gains is not a completed road, but an entire toolkit of road-building knowledge and capability—one that will deeply integrate with and continuously evolve alongside local industries, governance, and government operations.

In this sense, what participants will find “indispensable” is not any particular Chinese product, but the entire technological ecosystem that grows around China’s open-source models. Trading “non-lock-in” for “indispensability”—that is the essence of the “open-source Belt and Road.”

Now, a telling trend is visible on Hugging Face. If 2025 marked the moment Chinese open-source models surpassed American ones, 2026 has brought a new development: Chinese open-source models are now holding their own against closed-source competitors at the frontier.

At the same time, more U.S. applications are turning to open-source options, and a rising number of American models are being developed atop Chinese foundation models. Perhaps most strikingly, it is U.S. chipmakers and infrastructure players that have been the most aggressive in embracing open source — chiefly because it gives them the cheapest and fastest way to run China's frontier open-source models on American-made silicon.

Sure, the trend remains a work in progress, and no conclusion is yet in sight — U.S. closed-source models continue to lead in raw performance. Yet history suggests that technical superiority at a given point in time does not guarantee eventual market dominance. After all, the outcome, for now, is still very much an open question.

Open-Source — Trading Transparency for Trust

The two defining features of WAICO are open-source and inclusive access.

Open-source is, at its core, the opening of methodology — “teaching a man to fish” rather than “giving him a fish.” This enables technology-late countries to truly localize and independently develop their technological capabilities — precisely the “data sovereignty” and “tech sovereignty” that the European Union has long championed but struggled to achieve in practice.

At WAIC 2026, China called on nations to seize the “historic opportunity” of open-source AI, pledged to help developing countries build AI capacity, and warned that unequal access to technology could lead to “new historical injustices.”

One question inevitably arises: Will China’s open-source strategy change? My assessment is that it will not — and will, in fact, be further reinforced. There are two reasons. First, China’s AI sector does not enjoy a first-mover advantage — since it is not the leader, open-source is actually the most effective path to accelerate catch-up and expand ecosystem influence. Second, China’s AI credibility in international markets cannot overtake that of the U.S. in the short term — open-source is precisely the most direct way to trade transparency for trust. From a purely market logic perspective, therefore, open-source is not a value choice but a rational strategic one.

**Trump Card No. 2: Inclusive Access — Making AI a Global Public Good**

Inclusive access is another defining feature of China’s AI approach — making AI as affordable and accessible as renewable energy, cheap and available enough for ordinary people around the world to truly use. At its core, China aims to position AI as a global public good, rather than an exclusive tool for a privileged few.

By contrast: to this day, Anthropic’s most powerful Mythos model was once fully suspended from foreign access by the U.S. government on national security grounds. Following partial relaxation, access has been limited to approximately 100 vetted U.S.-based institutions. Its pricing — $25 per million tokens for input and $125 per million tokens for output — has earned it the title of “the most powerful, and the most expensive” model ever created. At such price points and access thresholds, AI has effectively become a resource that only a small number of privileged institutions can truly possess.

Of course, a fair caveat is in order: as intellectual property requiring massive R&D investment, the pricing of AI is inherently complex. High pricing is not inherently “bad,” just as inclusive access is not inherently “good.” This is not a value judgment — rather, it is an observation that China and the U.S. are now charting two distinct development paths in AI.

China has long labored under a cloud of global skepticism — though the more accurate diagnosis is that Washington has framed Beijing as a rival, and much of the world, whether by choice or by pressure, has fallen in line. Had this open-source, inclusive agenda been advanced by the United States instead, it would almost certainly have been met with open arms.

Pax Silica presents a very sharp profile: having expanded from 7 countries in December 2025 to 35 today, it explicitly covers critical minerals, energy inputs, advanced manufacturing, semiconductors, AI infrastructure, and logistics — this is a supply chain coalition built around the full AI and semiconductor industrial chain, with a clear exclusionary orientation toward China. Pax Silica knows exactly what it is.

Yet, WAICO’s profile, by contrast, remains rather blurry. Its 29 founding members include no core Western countries, no Pax Silica members — they are largely partners showing up out of friendship and trust. But what exactly do they expect to gain from joining WAICO? Models? Capacity building? A voice in governance? The structure of their demands remains unclear. WAICO knows what it opposes — technological monopoly and the digital divide — but it must quickly answer another question: what does it deliver to its members?

Yet one fact bears repeating: for all the lack of overlap between the two camps, the majority of nations fall into neither column. That unclaimed space in between — that is the true battlefield.

Can Pax Silica Lock In the Middle Ground?

The conclusion goes: It can lock in its Western allies — but it cannot lock out the Global South. What Pax Silica holds, it already possessed.

For the U.S. alliance network, Pax Silica is tightening its grip — and will only tighten further. The reasons are threefold: deep economic entanglement with U.S. markets and the dollar system, structural military dependence on Washington, and the political calculus that “this version of America is temporary — we can wait until 2028.”

Together, these factors ensure that U.S. allies have not and will not abandon Washington. For these nations, Pax Silica’s compliance standards and procurement rules will only become more stringent. Cracking open the U.S. alliance from within is not impossible — but the cost would be high and the payoff uncertain.

But, what it cannot lock in are the middle powers aligned with neither side.

At its core, Pax Silica is a high-bar “trusted supplier” club. It offers members standards, market access, and security assurances — but it cannot deliver what developing countries need most: affordable, autonomous AI capabilities. That is precisely where open-source and inclusive access meet the demand — and it is precisely why WAICO’s founding membership draws first from the Global South.

Even more consequential are the middle powers that appear on neither list: Gulf states pursuing economic transformation and hedging their AI bets; Southeast Asian nations embedded in China’s industrial supply chain while maintaining security ties with Washington; Latin American countries chasing digitalization — newly appearing on Pax Silica’s invitation list while remaining long-standing partners on Beijing’s cooperation roster; and some European peripheral states nursing grievances against both Brussels and Washington. What these countries share is a reluctance to choose sides — they want both. Hedging, not taking sides, is their default strategy.

A Race for the Same Swing States

So to answer the question head-on: Pax Silica can hold its Western allies, but it cannot seal off the Global South. In that sense, it cannot contain the globalization of Chinese AI.

But “cannot contain” does not mean “cannot win.” Pax Silica has already begun expanding into Latin America (Argentina, Chile) and Central Asia (Kazakhstan) — territories traditionally seen as China’s cooperation heartland. This shows that Washington has read the map correctly: the decisive battle is not in the alliance circle, but in the middle ground.

The middle powers are live pieces — and both rosters are still growing. The rule of this race is simple: whoever allows these middle powers to engage with both sides at lower political cost will keep them in their orbit. And that, in turn, hinges on whether China can steer clear of several major pitfalls.

Pitfalls China Should Avoid

The first three pitfalls correspond directly to the controversies that have dogged the Belt and Road Initiative; the fourth, paradoxically, draws on a positive lesson from the AIIB — and it is the one most worth heeding.

Who builds: Avoid State-Led, Embrace Market-Led

As an intergovernmental organization, WAICO’s very establishment carries the unmistakable stamp of state diplomacy — and that is a necessary and legitimate step in building international legitimacy.

But if the subsequent rollout of the “open-source Belt and Road” continues to be government-led, it will fall into a trap that demands close vigilance. The criticism that the BRI was overly state-driven cannot be allowed to repeat itself in the AI era.

The reasoning is straightforward: when AI cooperation is driven by states rather than enterprises, it is naturally viewed through the lens of geopolitical rivalry — difficult to interpret as pure market behavior or technological collaboration.

This only reinforces the narrative that “Chinese AI going global is a state strategic tool.” Conversely, when the lead actors are businesses operating on commercial logic, responding to customer demand and product competitiveness, such cooperation is far more palatable to the international community and more likely to sidestep ideological scrutiny.

The answer to “who builds” should be: governments set the stage, businesses take the spotlight. What governments should do is provide the top-level institutional design and political endorsement — as WAICO has done. But when it comes to ground-level cooperation in each country and each sector, the drivers must be genuine market players: Alibaba, ByteDance, DeepSeek, and the like.

How to Build: Avoid Full-Chain Exports, Embrace Localized Integration

This is the direct response to the “insufficient localization” controversy. If China replicates the “packaged full-chain export” model — with Chinese firms providing everything from chips to models to applications, transplanted wholesale into overseas markets — it will struggle to take root commercially (ignoring the differences in local industrial bases and real demand), and politically, it will be read as “technological colonialism,” playing directly into the negative narratives about Chinese AI.

The right path is deep localized integration: China provides the open-source model as the foundational “base” and the methodology. But the development of specific application scenarios, industry adaptation, and even fine-tuning of models should be left largely to local developers, local enterprises, and local industrial ecosystems. Only then can China truly “go inside” rather than merely “export its products” — the former is a long-term relationship embedded in local development logic; the latter is just a one-off commodity transaction.

How to Balance the Books: Avoid Showpiece Projects, Embrace Real Demand and Commercial Sustainability

This merges the “debt trap” and “white elephant” controversies into one — and it is the pitfall most easily obscured by early enthusiasm.

What has damaged the BRI’s reputation most is not projects that failed to be built, but projects that were built and then left unused. The AI version will be no different: if an overseas AI cooperation project lacks genuine local demand and a sustainable operating model, it is not an achievement — it is ammunition.

Narratives of “compute colonialism” and “data colonialism” are already waiting in the wings. The first batch of failed projects will define the entire initiative’s reputation — and the margin for error on an “open-source Belt and Road” is even narrower than it was for infrastructure.

Early projects, therefore, should prioritize quality over quantity: ten half-dead showcase projects are worth less than three that genuinely integrate with local industries and sustain themselves. On financing, China should also be wary of falling back on large-scale policy loans — once AI cooperation becomes entangled with debt, the “debt trap” narrative will be seamlessly transplanted onto WAICO. Commercial sustainability does not lower the bar for projects — on the contrary, it raises it.

How to Position: Avoid Bloc Confrontation, Embrace “Intermingling”

This is the most important pitfall, because it directly determines whether WAICO can win over the middle powers.

If WAICO is positioned as an “anti-Pax Silica bloc” — whether by Beijing’s own messaging or by Washington’s narrative framing — the political cost for middle powers to join will skyrocket. That would be tantamount to forcing them to choose sides — precisely what they resist most. Turning WAICO into a bloc would only help Pax Silica complete its own containment.

The AIIB’s positive precedent is right there: the United Kingdom joined despite U.S. opposition, and core U.S. allies were simultaneously AIIB members — no one saw a contradiction, because the AIIB never asked anyone to break with Washington. That very “intermingling” is the source of its legitimacy.

WAICO should embed this logic directly into its institutional design: include no exclusivity clauses in its charter; proactively invite Pax Silica members to participate as members or observers; and leave interfaces open for U.S. companies and institutions to engage. Joining WAICO must become something that “requires no explanation to Washington.”

This gives rise to a concrete yardstick: WAICO’s true success will not be measured by a longer list of friendly nations — it will be measured by the day it signs up its first western nation from Pax Silica framework (not only Kazakhstan). Before that happens, WAICO is just a circle of friends; after that, it becomes an international organization.

The window of opportunity will not stay open forever.

Pax Silica is still expanding its ranks, and the U.S. has made clear its intent to impede WAICO's further growth — all while the middle powers listen to pitches from both camps.

The coming 12 months will determine whether WAICO can make the leap from a like-minded club to a truly global institution.

Share

Leave a comment

On August 14, the U.S. Court of Appeals for the D.C. Circuit issued a mixed ruling in the case of SZ DJI Technology Co. (深圳大疆) v. Department of Defense, a high-stakes challenge to the company’s placement on the Pentagon’s “Section 1260H List” of Chinese military companies.

The court affirmed the Defense Department’s finding that DJI receives technological support from the Chinese government, and rejected DJI’s due process and disparate-treatment claims.

However, on the pivotal question of whether DJI contributes to the Chinese defense industrial base, the appeals court found that the district court had applied the wrong standard of review—relying on post-hoc justifications from government attorneys or constructing its own rationale from the administrative record, rather than reviewing the actual basis the agency used at the time of its decision.

Accordingly, the appellate panel affirmed in part, reversed in part, and remanded the case for further proceedings.

A critical caveat: the appellate court did not rule that DJI is not a Chinese military company, nor did it order the Pentagon to remove the firm from the 1260H list. Therefore, the ruling represents a narrowly procedural win for the Chinese drone maker.

The “Assistance” Finding: Substantial Evidence, Not Hard Proof

The Defense Department had argued that DJI’s designation as a “National Enterprise Technology Center” (NETC, 国家认定企业技术中心) by China’s National Development and Reform Commission (发展改革委)— China’s economic planning body and an agency the Pentagon characterizes as part of the military-industrial planning apparatus—proves the Shenzhen company receives government tech support, including tax breaks, cash subsidies, and state-backed capital funds.

DJI countered that the record at best shows eligibility for such benefits, not actual receipt. In its pleadings, the company flatly asserted it had “never received any assistance from the government through its NETC recognition.”

The appeals court found the Pentagon’s determination met the “substantial evidence” standard—a low bar that does not require the U.S. government to itemize subsidies or trace funds to a bank account.

As the court explained, “The DoD states that the ‘NETC enjoys many policy benefits,’ not that the NETC enjoys eligibility for such benefits. That language naturally suggests that DJI is currently receiving policy benefits.” The court further noted that the nature of those perks—such as preferential tax rates on imported equipment—are the kind companies would be expected to use if available. DJI’s own history of seeking and receiving other government subsidies only reinforced the inference.

DJI’s assertion to the contrary, the court held, was just a self-serving statement by its lawyers, unsupported by evidence, and insufficient to rebut the administrative record.

Procedural Claims Rejected

The court also summarily dismissed DJI’s due process and equal-treatment arguments. On due process, the panel found that DJI had failed to show that its blacklisting had “broadly preclude[d]” it from pursuing its chosen business—the high threshold required under the “stigma-plus” doctrine.

Despite losing some contracts and facing state-level bans, DJI continues to sell drones globally and remains the market leader, with an estimated 90% of the consumer market and nearly 70% of the broader drone sector.

As for disparate treatment, DJI pointed to Nokia Bell, Volkswagen, and Nissan as similarly situated companies that also hold NETC status but were not listed (the whataboutism). The court was unpersuaded, noting there was no evidence the Pentagon had ever reviewed those firms and decided against listing them. An agency, the court held, is not obliged to survey the globe for every potential comparator before designating one company.

The Crux: How to Define “Contribution” to China’s Defense Industrial Base

The real turning point—and the source of DJI’s appellate victory—lies in the court’s analysis of whether DJI contributes to the Chinese defense industrial base. DJI had argued that even if it received industrial or technological support, that does not automatically mean it enhances China’s military capabilities.

Notably, the Pentagon’s unclassified report included a section accusing DJI of contributing to the “Chinese Defense Industrial Base,” but the substantive explanation was entirely redacted. Neither DJI nor the public could see the factual basis or reasoning behind that conclusion.

However, the district court, without reviewing the classified justification, instead pieced together support from other parts of the report and accepted new explanations offered by government lawyers in litigation.

The appeals court found this approach violated the Chenery principle—the bedrock administrative-law rule that a court may uphold agency action only on the grounds the agency actually invoked at the time of its decision, not on post-hoc rationalizations supplied by its attorneys.

As the panel put it: “Courts ‘must judge the propriety of [agency] action solely by the grounds invoked by the agency.’” The court emphasized that it is “powerless” to supply rationales the agency itself did not provide. Otherwise, the court would be impermissibly second-guessing the Department of Defense's factual determinations.

Crucially, the appeals court did not hold that the Pentagon lacked evidence, nor did it rule that DJI’s drones lack military utility. It simply found that, without reviewing the classified record, the district court could not rely on public materials to sustain the listing.

The appellate court also declined to conduct its own de novo review of the classified materials, noting that it is “a court of review, not of first view.” The case now returns to the district court, which may examine the classified administrative record to determine whether the Pentagon’s actual, contemporaneous reasoning supports the “contribution” finding. The district court may also decide whether DJI or its counsel may access the classified materials under some form of restricted review.

The Mootness Question

On remand, the district court will first confront a procedural question: the Pentagon published a new 1260H list in June 2026, again naming DJI, this time under the amended statutory criteria.

The Pentagon may argue that the new listing renders DJI’s challenge to the January 2025 designation moot. The appeals court, however, found that on the current record, mootness is not established—it remains plausible that each annual designation carries distinct reputational harm, and it is unclear whether the new rationales are factually supported and independently sufficient.

What the 2026 Redesignation Means for DJI

Even if DJI ultimately prevails on the 2025 designation, that victory would not automatically remove it from the current 1260H list. The June 2026 listing invokes new grounds—including the “Single Champion” award, affiliations with the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission (SASAC), and presence in a military-civil fusion enterprise zone. To challenge its current status, DJI would need to separately attack those new justifications or seek to fold them into the existing litigation.

The Broader Legal Landscape: What Does “Contribution” Really Mean?

The 1260H statute provides multiple paths to designation. An entity may be listed if it is found to have ownership, control, or affiliation with the People’s Liberation Army, the Armed Police, the Ministry of State Security, MIIT, SASAC, or the State Administration for Science, Technology and Industry for National Defense. Alternatively, a company may be designated if it qualifies as a “military-civil fusion contributor to the Chinese defense industrial base.”

That second path first requires the company to meet one of several statutory categories—for example, receiving government support tied to military-industrial planning, maintaining certain government or military affiliations, being located in a military-civil fusion park, holding defense-related qualifications, or offering products on military procurement platforms.

The real dispute is whether, after meeting one of these categories, the Pentagon must separately prove that the company actually “contributes to the Chinese defense industrial base.”

In the Hesai Technology (禾赛科技) case, the government initially argued that no separate showing was required—that satisfying any category automatically qualifies as contribution.

The district court rejected that interpretation, holding that it would render the phrase “to the Chinese defense industrial base” virtually meaningless—a violation of the canon that every word in a statute should be given effect.

The 2024 amendments, which added broad categories like “Single Champion,” “little giant,” and military-civil fusion parks, only sharpened the issue. If those industrial-policy designations alone suffice for listing, then a vast array of ordinary industrial beneficiaries, park tenants, and even foreign multinationals’ Chinese subsidiaries could be swept in.

A separate “contribution” threshold, the court suggested, is needed to distinguish general industrial ties from genuinely defense-significant connections.

The Pentagon has a plausible counterargument: Congress may have deliberately chosen these categories as proxies or indicators of contribution, precisely to avoid requiring the Pentagon to trace every product to a military end-user.

The DJI case does not definitively resolve this question, because both parties agreed that a separate contribution finding is required. The appellate court “assume[d] without deciding” that such a requirement exists.

What Standard of Proof for “Contribution”?

Even if a separate contribution requirement is recognized, the next question is: what must the Pentagon prove, and to what degree? In Hesai, the plaintiff argued that only direct supply to the Chinese government or military for defense purposes should count. The district court deemed that standard too narrow, noting that the defense industrial base includes private firms with the capability to research, design, manufacture, or maintain weapons systems, subsystems, or components—not just direct contractors.

The court also rejected the government’s expansive view that any theoretically dual-use product qualifies. Instead, it adopted an intermediate standard: the company’s products or technologies must have a “substantial military application or benefit.” Merely possible, hypothetical, or incidental military uses are insufficient.

This test breaks down into three sequential questions:

1. Specificity: The Pentagon cannot merely point to a militarily significant industry—AI, biotech, cloud computing, or lidar, for example. It must identify which of the company’s specific products, technologies, or capabilities have military value.

2. Materiality: The military use must be real and significant. Direct military procurement, defense cooperation, military certification, or actual use by armed forces are the strongest evidence, but not the only permissible proof. A technology with an unmistakably important battlefield function may suffice.

3. Attribution: The Pentagon must tie the military application to the company itself. Evidence that other firms in the same sector supply the military, or that a generic technology could conceivably be weaponized, is not enough to show that *this* company enhances China’s defense industrial base.

In Hesai, the government’s evidence included reports that lidar is critical for autonomous military vehicles, academic articles on China’s lidar industry and national security, and even a competitor’s product (Ouster) allegedly found on Russian or Chinese military vehicles.

The district court acknowledged the evidence was not overwhelming but deferred to the Pentagon given the low “substantial evidence” threshold and the deference courts typically afford agencies on national security and military technology matters.

The Hesai appeal is now pending before the D.C. Circuit, and its outcome is likely to clarify the “contribution” standard. A likely eventual framework: the company’s own products, services, or industrial capabilities must have a specific, significant, and non-accidental military application, supported by company-specific evidence that plausibly shows how that capability directly or indirectly strengthens China’s defense industrial base. Moreover, the Pentagon must articulate that reasoning at the time of designation—not after litigation begins.

Implications for Alibaba and WuXi AppTec

DJI’s procedural victory has different implications for other Chinese companies on the 1260H list, notably Alibaba and WuXi AppTec (药明康德).

For Alibaba, the “contribution” question could be a key vulnerability. Cloud computing, big-data analytics, and large-language models may indeed have significant military applications. But the Pentagon cannot simply assert general military value—it must identify which of Alibaba’s specific products or capabilities have material military use, and provide individualized evidence linking Alibaba to the PLA, defense contractors, defense research projects, or military systems.

That said, Alibaba faces a higher hurdle: the 2026 listing cites not only its status as a “military-civil fusion contributor” through its MIIT affiliation, but also an indirect relationship with SASAC, potentially placing it on another statutory path to designation. Even if Alibaba prevails on the “contribution” issue, it may still need to challenge the Pentagon’s legal and factual basis for the “affiliation” finding.

WuXi AppTec’s situation is quite different. The Pentagon’s stated rationale is not that WuXi “contributes to the Chinese defense industrial base,” but rather that it is indirectly held by SASAC and connected to the State Administration for Science, Technology and Industry for National Defense and the PLA.

WuXi AppTec has already obtained a preliminary injunction, as the district court found significant factual gaps in the government’s evidence—including a misreading of a fund-ownership data point and an overstatement of the significance of third-party lab services.

Thus, the “contribution” standard from DJI and Hesai is only tangentially relevant to WuXi AppTec’s current litigation. WuXi AppTec’s strongest argument for now is that the Pentagon failed to accurately prove the ownership or affiliation relationships required by statute. Only if the government later redesignates WuXi AppTech under the “military-civil fusion contributor” path would the general national-security significance of biotech need to be translated into a company-specific showing of contribution to China’s defense industrial base.

Bottom Line

DJI’s appellate ruling is a notable but circumscribed triumph. The company succeeded in forcing the court to look at the Pentagon’s actual, classified reasoning—but it did not secure an order for delisting. The case now goes back to the district court for a behind-the-scenes review of classified evidence. Meanwhile, the broader legal battle over what counts as “contributing” to China’s defense industrial base continues, with the Hesai appeal poised to offer more definitive guidance. For DJI and the other listed companies, the road ahead remains long, and the outcome far from certain.

Adapted from an article by the Dongbuyaqiao Research Institute (东不压桥研究所)。

My Final remark: In a sense, U.S. federal courts do provide a viable legal platform for Chinese companies to challenge the Department of Defense's blacklist decisions, as evidenced by cases like WuXi AppTec and Xiaomi.

However, even if a company ultimately prevails or is removed from the list, the damage to its business reputation and lost cooperation opportunities during the listing period is real and often irreparable.

In fast-moving technology fields, the Pentagon achieves its aim of obstructing Chinese competitors simply by tying them up in litigation—win or lose.

Share

Leave a comment

During an August 7 conversation titled "China Shocks USA: The Fight Is Already Over" on the YouTube channel Neutrality Studies with Kyoto University's Pascal Lottaz, Professor Wang Wen (王文), Dean of Renmin University's Chongyang Institute for Financial Studies, offered a compelling analysis of the current geopolitical landscape.

From the Chinese scholarly perspective, he explained, Trump 2.0 is not merely a disruptive presidency but a profound systemic crisis in U.S. governance — one that is paradoxically quickening the unraveling of American hegemony. Yet Wang emphasized that China has responded not with fear, but with strategic maturity: shifting from a defensive posture to one of active influence, and channeling external pressure into a powerful engine for homegrown innovation and rising national confidence.

He said that betting for China is not betting against the U.S., but to improve the lives of its 1.4 billion people through long-term domestic planning, and he believes the two countries are now escaping the Thucydides Trap toward a more stable and constructive relationship.

Wang also rejects Western realpolitik theories of inevitable security competition, arguing that China competes for development, not domination, and envisions a multipolar world based on mutual respect, inclusive cooperation, and a shared future for mankind. He also shares his views on Iran, Taiwan risks, BRICS, the Belt and Road Initiative, global governance, green energy and AI.

Subscribe now

Following is the transcript of the dialogue:

Pascal Lottaz: Welcome back everybody to Neutrality Studies. Today for the first time with Professor Wang Wen, the Dean of the Chongyang Institute for Financial Studies at the Renmin University of China. Professor Wang, welcome.

Wang Wen: Thank you, thank you for your opportunity to give me.

Pascal Lottaz: It’s fantastic to have a Chinese scholar here and you’ve been, actually you are one of the very attentive people who try to analyze also international relations and that’s why I would like to ask you, you know, a little bit mainly about the China’s relationship today with the United States and maybe let’s start with Donald Trump because it’s hard to get past him. How do you think, how do Chinese people actually view the second Trump administration?

Wang Wen: Yeah, well, this is a very important question, you know, Trump is so famous, also is very famous in China, every Chinese people know him. But the problem is that gradually, you know, compared to eight years ago or two years ago when he got the office again, I think the Chinese people now have found many consensus about Trump 2.0. Let me show you a few consensus of Chinese viewpoint on Trump 2.0.

First is that Trump 2.0 is a concentrated outbreak of inherent contradictions within the American system. The transactional American first policy is rapidly eroding American global leadership. As I said, compared to Trump 1.0, who took office in 2017, Trump 2.0 is no longer merely about impacting the Washington establishment, but a systematic transformation of U.S. governance rules, purging the federal public office system, unlimited expansion of executive power, indiscriminate imposition of non-reciprocal tariff on allies, and significant reduction in global public goods investment.

So now we can see from the Chinese people, I mean, especially Chinese scholars, most of the Chinese academics agree that all of Trump’s policies are driven by short-term electoral and commercial interests, lacking any medium or long-term global governance planning. This utilitarian hegemony is actively undermining the global order framework built by the U.S. after World War II, objectively accelerating the multipolarization process. So I think for Chinese people, especially for Chinese scholars nowadays President Trump destroyed the American hegemony. That’s, I think, a very important consensus on Trump 2.0.

Pascal Lottaz: You know, there are people in my world, of this YouTube bubble that try to analyze U.S. hegemony, that argue that overall Trump is not that exceptional, because overall the Trump administration continues with this hegemonic approach, the war on Iran, the proxy war on Russia, the undermining of the global order that doesn’t benefit the United States anymore. So it’s still kind of the continuation of U.S. expansionism, and also it’s ratcheting up of war rhetoric with China, and the U.S. bases are not going away, and the energy shock is rather also causing problems for the Chinese economy. What do you and Chinese scholars make out of such arguments? Do you think this is, you know, just the latest version of a long-term strategy, or is this, you know, Donald Trump as a kind of a very bizarre administration that’s actually harming U.S. interests more than it helps?

Wang Wen: Yeah, I agree with you. You know, for most of the Chinese scholars, they think that competing with China is a long-standing strategic consensus between the two U.S. parties. But the problem is that Trump 2.0, businessman-like transaction thinking, allows more controllable negotiation space in the China and U.S. games. Whether it’s Biden or Trump, the underlying logic of the U.S. position in China as it’s a primarily strategic competitor remains unchanged as you mentioned right now. However, the biggest difference from the Biden era is that Trump is not fixated on ideological confrontation. All negotiations are measured by economic gains and reciprocal exchanges.

In 2025, I mean, last year, the U.S. raised the tariffs on Chinese goods to 145 percent. However, many U.S. domestic businesses protested and infliction of pressure, forced the U.S. to actively seek multiple rounds of trade negotiations with China. So I think there were a lot of cases confirming our shared judgment. Trump’s tariff stands on China is a means to an end, the real goal is to gain economic benefits. So this leaves a window for China-U.S. relations to manage the differences and engaging in pragmatic cooperation in certain areas.

And another point I want to share with you is that, because you mentioned the containment or Trump’s competitive strategy, but nowadays for China side, we think the comprehensive external pressure has not hindered China’s development. Instead, it has forced us to build a safer, more diversified and independent development system, transforming pressure into long-term growth resilience. So that’s what will make us more and more confident.

Look at the conditions happened in the past eight years. I give you an example. When Trump launched his first trade war in 2018, there was considerable anxiety in Chinese public opinion. But what was the result by 2026, I mean this year, the blockade spurred independent innovation. China is rising in semiconductors, ships, new energy vehicle, photovoltaics, robotic, larger language model and more. So I don’t know whether you read my very interesting article written last year, which was cited by media outlet in over 20 countries, including The Economist. My articles last year is that, thanks to Trump for making China great again. This article stated the extreme external pressure cannot hurt Chinese style modernization. Instead, it has allowed us to fill gaps in our industrial chain, expand our global cooperation and achieve a qualitative leap in our strategic resilience.

Pascal Lottaz: So yeah, that is a very good title for a very good article. It actually made me wonder whether you and other Chinese scholars actually kind of think like, thank you, Trump, for being a teacher without wanting to be one. Is that kind of a sentiment that is currently around in China that, okay, we can learn a lot from what’s going on over there, let’s make the best out of it?

Wang Wen: Well, yes, I agree with you. Nowadays, more and more Chinese people, when they face Trump 2.0, the Chinese are becoming more and more confident. I must tell you a fact, China may be better at dealing with the U.S. than the U.S. dealing with China. China is not afraid of Trump 2.0. On the contrary, China has successfully responded to Trump’s trade war, high-tech war, and other containment and suppression efforts. So I think this self-confidence is the proactive self-confidence in dealing with external games.

As I said, eight years ago, when Trump 1.0 came to power, when faced with the sudden trade war, China responded more passively and took advantage of the situation. However, after the Trump 2.0 cycle, China has learned to proactively shape U.S.-China relationship. Let me give a few examples.

At the economic and trade level, we have accurately implemented rare earth export controls, energy product tariff counter measures, and unreliable entity list reciprocity constraints. And we no longer passively endure unilateral barriers. At the global governance level, we rely on BRICS, Shanghai Cooperation Organization, Belt and Road Initiative Expansion, and cooperation in Global South to pigeon against the U.S. unilateralism. And more importantly, very interesting is that China’s experience is that at the communication level with the U.S., we take the initiative to conduct hierarchy dialogues, distinct different topics such as economic, trade, climate change, humanities, and security, to set cooperation agendas, proactively manage risks, and avoid escalation of conflicts with the U.S.

So now, it’s very obvious that, as you mentioned, in the past eight years, the mentality of Chinese people has changed very clearly. In the past, we passively adopted to the rules set by the United States. But now, we have the ability to balance the differences, proactively create opportunity for bilateral cooperation, and no longer dictated by the reins of the U.S. policy. That’s one kind of Chinese self-confidence.

Pascal Lottaz: That is very interesting, because what I hear from your words is that you are actually rather positive, rather optimistic about China’s ability to deal with the United States. So, does that influence now the future planning of the U.S.-China relationship, that China doesn’t feel that they just need to deal with anything being thrown at them, but that they can also now start long-term planning of how to proactively approach the U.S. and balance things? Is that the thinking in Beijing these days?

Wang Wen: Oh, well, I think you mentioned why China has more and more self-confidence deal with the United States. I think the fundamental confidence is that China is in China’s own development path. So, because, let me explain to you, because I think many Western public opinions have fallen into a cognitive misunderstanding. They think that China-U.S. relations are a zero-sum competition, and they believe that China’s development goal is to surpass and defeat the United States. I think this is precisely the biggest misunderstanding of Chinese development logic.

Let me tell you that the development goal of China has never been to defeat any country, but to improve the people’s livelihood, achieve common prosperity for all people, and complete national rejuvenation. So now, I think, even though it has been in a competitive environment among the major powers for a long time, from Europe, from Japan, or from the U.S., China’s policy focus has always been based on the domestic demand.

I don’t know whether you know, this year is the first year of our 15th Five-Year Plan. I don’t know whether you know of 15th Five-Year Plan. 15th Five-Year Plan focuses on boosting domestic demand, technological innovation, urban-rural balance, and ecological protection, ensuring tens of millions of jobs every year, improving the pension and medical system, and continue to revitalize the countryside. All strategic starting points are the needs of a better life for the 1.4 billion Chinese people.

But look at Trump’s rule, look at American society. There are so many polarizations between rich and poor in the United States. There are so many that the welfare of the people at the bottom has shut. So I think our confidence is that our track has never been to catch up with the United States, but to continue to do our own things well. So we become more and more confident because we do very well, much better than before. So China become much more better than before. And we think that we can have more confidence to deal with the U.S..

Pascal Lottaz: This is very interesting, you know, because what you’re saying is the goal of China is self-improvement and not to hurt the other one. And that’s the motivating factor. But you know, when we listen to a lot of the rhetoric that comes out of Washington or even among the think tanks and so on, they now have a goal of disturbing or preventing China from building the Belt and Road Initiative. And they’re very proud, you know, that they picked off Venezuela. And now, you know, Colombia is going to get a right wing government and they’re going to leave the Chinese orbit. The West keeps thinking about who do we control and who does China control. But you’re saying that’s not the point. That’s not what we are after. So in this sense, China is probably not very worried about Venezuela and Colombia because it’s like they do what they do and then China will do something different, right?

Wang Wen: Yeah, I think in the future, the ultimate area of U.S.-China competition is not tariffs or technological barriers or occupation to the other countries, but a contest of domestic governance capability. That’s why we believe that time is on China’s side in the long term, in the long run, favoring stable and continuous development policy.

So that’s why in media, we often report the American society is divided, American policy is volatile, and each administration overturns many of its predecessors’ policy. So in contrast, over the past four decades, China has consistently implemented its medium and long term plan with strategic such as targeted poverty alleviation, common prosperity, new style productivity, and green transformation continues taking efforts. So I think gradually and gradually, in the long term, China’s policy continuity, massive market, complete industrial system, and institutional advantages that coordinate development and security will continue and to narrow the gap in comprehensive national strength between China and the U.S..

So that’s our fundamental basis for our long term confidence. But look at the United States, the U.S. has so many conflicts with the outside world, that will stop and prevent the U.S.’s sustainable development in the long term. So we compare it, you can find why Chinese people have more and more confidence.

Pascal Lottaz: But does it not worry you when you see how the United States is trying to actively hurt its other powers on the world stage, like the war with Russia, NATO’s war with Russia, and especially now the war with Iran, doesn’t it worry Chinese strategic thinkers like you that you think like, okay, we might be next, they might try to transform Taiwan or the Philippines into Ukraine for China in order to drain our resources? Is this not being discussed?

Wang Wen: I think maybe last year or in the past, we often worry about it. We worry about that, maybe whether China will be the next one who will be conquered or the war launched by the U.S.. But nowadays, I mean, especially this year, we become more confident because I don’t know whether this year, this May, the two leaders of U.S. and China, we cognize a new consensus. The two leaders reached a new consensus, the constructive, strategic, and stable relationship. The core of this relationship is cooperation as a primary focus, orderly competition, manageable differences, and prospect of peace.

So this is extremely important. So I think this year is very, very important, which was misunderstood or ignored by a lot of public opinion in the West. This year, in my opinion, the U.S. and China are escaping the Thucydides’ trap. Both countries are very rational. Maybe as you may know, although some countries such as Japan or the Philippines and even Europe hope for a war between the U.S. and China, the leaders of China and the U.S. are very rational, at least at the strategic level.

In my opinion, even President Trump is a very controversial person, but Trump is the most rational person in Washington political circles regarding China policy. This year, the two leaders between China and the U.S., will have a full face-to-face meeting. This is very, very rare. And because of the, as I said, the constructive, strategic, stability relationship, this is a very important new consensus.

And more importantly is that we are experiencing a new era of interaction between major powers in President Trump’s 2.0 administration. For example, the U.S.-China-Russia relationship. Perhaps, I guess, at the APEC summit this November in Shenzhen, China we will see a photo of President Xi Jinping, President Trump, and President Putin together. A new type of a big power relationship. Of course, I don’t see this as a G3 level action. I just want to say that I see it as a beginning of a post-post-Cold War era, led by three major powers, just as the United States, Soviet Union, and the United Kingdom led the establishment of the Yalta system after World War II.

So that also demonstrates that China’s response and China’s countermeasure to Trump’s 2.0 over the past few years, and even to the U.S. containment policies over the past 20 years or so, has been very, very successful. Because the U.S. agrees with China as a new consensus, the constructive strategic stability. So the era, in my opinion, the era of vicious competition between China and the U.S. is passing. So that’s why, as you mentioned, I’m very, very optimistic about China and U.S. relations.

Pascal Lottaz: It’s so interesting that you’re saying that, because to me, it seems it’s going the other way. It seems to me that the United States is gearing up for more, not just competition, but actual fight. Maybe not direct kinetic fight with China, but at least the trade war to try to hurt Chinese interests. But you’re much more optimistic. And I just want to ask, what you just portrayed, is this your expectation of what will happen? Or is it your assessment of what China’s strategy is probably going to be? Where China wants to push the relationship?

Wang Wen: Yes, as I said, China’s strategy of development is not to conquer other countries. China just focus on ourselves and focus on the broader global trends. So if we observe the world, as well as China and the U.S. future from a more crucial perspective of broader global trends, China has achieved a systematic and forward-looking strategic layout in the forefront of global future.

If we observe U.S.-China competition decades ago, that condition was characterized by the U.S. leading and China catch up. Now, in the cutting-edge emerging fields that will determine the global landscape for the next 30 years, China has already completed a forward-looking and systematic layout, achieving a struggle for leadership and a pace of advancement in multiple core future track.

Let me explain to you. For example, in the track of low-carbon and green development, China is the only country in the world to have built a full-chain large-scale and industrialized green energy system with a pace of development far existing that of the U.S.. You look at four cores of low-carbon industrial globally, Wind power, photovoltaics, lithium battery, and the new energy vehicle, China production capacity, output, and export account for more than half of a global total. So look at the future. It can be said that the standards, industrial chain and implementation model for global green transformation in the future are basically defined, supplied, and lead by China. That’s one future track.

The second track is about the modern social governance. In the field of modern social governance, China has formed a global unique and highly efficient governance system domestically, transforming its institutional advantage into long-term development advantage. So if you come to China, you can see that rely on a systematic approach to digital governance, grassroots governance, and emergent response, China has built the world’s largest digital governance system, enabling a precise and efficient implementation of public service, social control, and emergent response. If you look at China’s routine pandemic prevention and control, if you look at the disaster emergent response, if you look at China’s rural revitalization, if you look at China’s common prosperity, if you look at China’s full-scale construction of smart city, smart community, China has achieved a stable, precise, inclusive, sustainable, modern governance model. So I think this is what we call its domestic governance capability.

So these domestic governance capability balance efficiency and affairs, balance development and security, balance short-term livelihood and long-term development. So I think this is a very powerful domestic governance capacity. That is a core advantage that American-style system cannot replicate. So it’s also the underlying support for China’s long-term stable development and continuous accumulation of momentum.

And another very important, and you might be very interested in another track, that is now competitive very seriously. That is AI and future computing power. We can compare China and the U.S.. China now has achieved multi-dimensional advisement in application implementation, space computing power, and industrial ecosystem. Yes, I agree. So U.S. advantage, I cannot deny, U.S. advantage still has a lot of advantages in fundamental areas, such as underlying algorithms and high-end GPU chips. But the problem is that China’s overall layout leads in the future direction, such as total electricity, AI industrialization, scenario implementation, and space-based intelligence computing power.

So what I want to say is that now China focuses on our domestic construction. Our top leader, focus 99% energy on the domestic construction, you know, push and promote our body healthy and healthy, stronger and stronger. As long as our body, our national body is strengthened, become much stronger than before, no any external strengthens can conquer us, can occupy us, can beat us. So I think that’s very, very important. That’s a different strategic method. For United States, their leaders focus on the external expansion, external conflict with other countries. But for Chinese leaders, we focus on the domestic development. That’s very, very different. So in the future, I think time belongs to China.

Pascal Lottaz: This is really very interesting because you’re not conceptualizing it as China becoming number one because that’s not the point. China wants to continue to develop whatever the other ones do, I don’t care. Like “we just continue developing to the moon and beyond”. But there is this danger that the United States or others and the Western countries including Europe might start using violent means. I mean, it doesn’t, what happened to Iran and the war that was started by the United States by an aggressive attack again, like a breach of the UN charter and whatnot. Isn’t that a threat or how does this war and also, you know, the continuous from my view lies by the United States when trying to use diplomacy as a way to pretend that everything is fine but then strike militarily when it’s convenient for them. Doesn’t that create kind of, isn’t that a threat to China or how does this war impact China’s assessment of U.S. capabilities?

Wang Wen: I think, first of all, I admit that the U.S.-Iran conflict is a huge tragedy. The war now has crossed the red line of the earlier tacit understanding and the risks are escalating now especially since July, the resumption of the hostility. Both sides have shifted from simply striking military targets to attacking each other’s energy infrastructure even breaking the previous consensus of not touching civilian facility. So now I think the Middle East, the area is sliding into a deeper vortex of confrontation. So that’s a very, very dangerous conditions in Middle East. It’s very obvious U.S. now is hijacked by the Israel.

But in reality, you ask me, Chinese people, whether Chinese feel the threat from U.S.? No. I think from the Iran conflict with U.S., it shows that U.S. is still a paper tiger lacking the capacity for a protracted war. Because we all know that many American scholars and analysts believe that this war, United States has already lost. The U.S. has changed nothing. It failed to push for regime change in Iran. It failed to gain control of Strait of Hormuz. It failed to control global oil prices. And it failed to protect the interest of Gulf states. The U.S. is a complete loser in this war.

So I predict, because of the domestic political polarization in the U.S., coupled with approaching the midterm elections making the White House unable to bear the cost of the prolonged, high-intensity war. And now U.S. Gulf allies are also generally unwilling to be directly involved in the long-term conflicts. So that’s why I think as Chairman Mao Zedong said many, many years ago that the U.S. is a paper tiger. And this war between U.S. and Iran, U.S. would lose this war far worse than the Vietnam War and Afghanistan War, in my opinion.

And historically, we can observe the U.S. war history with the other country. The U.S. has never been able to conquer or defeat a country with a population exceeding 30 million. And how about Iran? Iran has 80 million. It’s a middle and a big power regionally. And also Vietnam. Vietnam that time in the 1960s, 50 million or over exceeding 30 million.

So I think the U.S. especially need to respect those regional power. The U.S. need to respect those countries with civilized tradition, including Iran, to be honest. For China, of course, U.S. need to respect us with a long-term civilized tradition. Of course, China’s population is much bigger. So that’s why I think for Chinese people, we observe the war between U.S. and Iran. It could promote us more confidence than before because the U.S. cannot conquer Iran, how dare he conquer China?

So this is our observation. Sometimes we just feel it is a very, very big tragedy, big tragedy for those people who died in the war, especially for those Iranian people, especially for those primary school students in Iran. So that’s my observation and opinion.

Pascal Lottaz: Maybe the very last kind of complex that I just would like to ask your opinion about, it’s like China’s cooperation with the BRICS states and these regional corporations with Russia, Iran, but extending on South Africa, Brazil, and so on and The Belt and Road Initiative. Should we understand China’s interaction with other powers the way that you described earlier, just as part and parcel of trying to improve China? And in that context try to improve relations and try to improve trade with others? Or are there more geopolitical aspects behind it like trying to build a network to resist, for instance, Western pressure when it comes to the banking network or to trade networks? How should we understand China’s larger foreign policy approach?

Wang Wen: Well, I think you launched a very important question about China’s future. If I can understand it like this, now, as a think tank scholar, we think for our country, because nowadays China’s national domestic strength has become much more powerful than before. We grew up very quickly. Our economy value has become much bigger than before. So what should we do? What do we want to the world? This is a new question to us. Every day we think about it. And also then we’re consulting for our government. And also our government, you know, gradually China has our new strategic ambition, this ambition.

What about China’s foreign policy ambition? I think at least five points, if I can share with you:

First, we want to pursue a peaceful development environment shared by all countries. China hopes to maintain universal peace, oppose hegemonism, power politics, and various camp confrontations and create a stable development environment for all humanity. So therefore, as you can see in the past many years, China has launched a Global Security Initiative to prevent the world from falling into new turmoil and division, allowing all countries to pursue development with peace of mind. That’s one ambition.

Second ambition is that we want to promote common prosperity globally. China wants to break the unfair pattern of a few countries monopolizing the dividends of development so that the fruits of development can benefit people all over the world more widely and equitably. So therefore, through the platform, as you mentioned, Belt and Road Initiative and Global Development Initiative, China is helping countries in Global South especially in Africa to address infrastructure shortcomings, cultivate independent development capability and promote a more diversified source of global economic growth.

And third strategic ambition for China’s foreign policy in the future is that we want to build a more just and reasonable global governance system. So that’s why I think China hopes to change the old governance rule and the old governance system domestic by the West and promote the reform of the international order towards consultation, joint construction, and shared benefits. So therefore, China advocates the genius of multilateralism and support the UN in planning a central role and enhance the voice of a developing country in the international affairs through the mechanisms such as AIIB, BRICS Cooperation you mentioned and the Shanghai Cooperation Organization, you mentioned, ensuring that global affairs are handled through joint consultation among all countries.

And I think the fourth foreign policy ambition of China is that we want to safeguard diverse and inclusive ecosystem of human civilization. Because China hopes, as you know, China has a long civilization history, China hopes that different civilizations and different development paths in the world can be respected. So we disagree so-called crash of civilization launched by Samuel Huntington many, many years ago.

And the final strategic foreign policy ambition of course is China working together to address a common challenge to human survival and development. Because we all know nowadays we have so many common threats such as climate change, digital governance, and public health. So China just hopes to encourage countries around the world to transcend differences and jointly solve common problems facing all humanity. So that’s why China is promoting the world’s largest low-carbon revolution. China hopes to work with all countries to safeguard our shared home earth.

So my conclusion is that we should, when we observe China, when we think about China’s rise, we need to jump from the traditional international relations theory, the Realistic International Relations Theory (Note: Realists believe that states and nations will always act to increase their power, regardless of the moral consequences of such actions). Because if you think about China in terms of the realism theory of international study, you can find that, maybe you can got a conclusion that, oh yes, China rise up, then China will become the next hegemony. No, if you jump from the Western international relation, realistic theory, you can see that what China wants has never been to dominate the world, but to work with all countries to build a community with a shared future for mankind, so that the world is free from the hegemonic oppression and development gaps and achieve common progress for all mankind.

So from this point, you can explain everything China did in the past many decades. For example, we have even a neighbor conflict, neighbor differences between Philippines. We have full military strength to maybe can push the military reunification with Taiwan, but we didn’t. We didn’t do that. We just keep patience and solve it step by step. This is resilience and patience of new big power strategy. So I think this is very interesting.

So how to judge the new style of big power’s rising, China? I think we need to push our thinking innovation. So as a think tank scholar from China, we just try our best to explain China’s rising up logic to the foreigners. That’s my opinion.

Pascal Lottaz: This is fascinating. I actually wanted to ask you exactly about that, about Western realism. And just the very last point, because I think John Mearsheimer is also a famous scholar in China, and one of his main points that I struggle with is that he builds the entire thinking of realism around the inevitability of security competition. And if I understand you correctly, then you would probably disagree and say like, security competition is not what we do. We compete, but we compete for development, not for security. And hence, I would understand you as saying like, no, in this point, Mr. Mearsheimer doesn’t correctly assess what China actually wants. Is that the case?

Wang Wen: Yeah, actually Professor John Mearsheimer, he’s my long-term friend. I invited him to deliver a lecture in our university twice. I think that, yes, we have a lot of debate, but we respect each other very well. We often contact by email. That’s the new style, like the new style relationship between the U.S. and China in the future.

I think we do not reject the fair competition with the United States and we continuously seek a model of mutual respect and a peaceful coexistence. I think in the long run, China and the U.S. are destined to coexist and competition will be the norm. But conflict and confrontation are by no meaning inevitable. Because as I said, China’s original aspiration for development remains unchanged. All its forward-looking plans, industrial upgrading, governance reforms, the technological breakthrough have never had the ultimate goal of defeating or surpassing the United States, but rather of addressing its own development shortcomings, improving the quality of life for its people, and achieving sustainable national development.

So I think, compared to passively responding to the United States in the past, China’s capacity rises up very quickly. In the past, China followed the rules and adopted to the existing landscape. But now, we proactively plan our own path, shake the new rules, and lead the new trends. Because we are increasingly convinced that time is on the side of long-termism, time is on the side of stable development, and time is on the side of inclusive and win-win cooperation. That’s our thinking, our trust, and our idea.

Pascal Lottaz: Professor Wang, that makes me very happy to hear that, because quite often, my discussions in the end, they become very gloomy about the future of humanity and whatnot, but you’re giving me a very positive outlook. I’m very grateful for that. People who want to find your work and who would like to follow you or your institute, where should they go?

Wang Wen: Yes, they can click our website. So welcome, any friends, contact me. We can debate, we can discuss, we can make friends. And I think that now, we need more and more thinking innovation. In the past, we trusted the Western theory of international relations, but now I think only those Western international relations theories are not enough to understand the new world.

For China, we wake up, we rise up. But the problem is that after we become much more powerful and stronger than before, every day, we learn from the U.S. successful experience. But simultaneously, we learn from the U.S. the negative lessons. We always caution, we always suggest our policy makers that, No, don’t learn from the U.S. those negative issues. Our ambition is to become a much better big power, much more civilized big power, much more beautiful world. That’s our long-term strategic ambition.

Pascal Lottaz: Well, and I would encourage everybody to engage and actually to join in building a cooperative framework for global development. I find that a very, very good idea. We will certainly have you back here on this channel, Professor Wang. I hope in the future also with other scholars, maybe from Russia and other places in order to chart that out. Professor Wang Wen, thank you so much for your time today.

Wang Wen: Thank you.

Finally, a quick observation:

Seemingly, after the interview, Pascal was not persuaded by Wang's arguments. His repeated use of "That is very interesting" throughout the questions strikes me less as a sincere commendation. Still, he gave Wang the space to articulate his views completely — and in that sense, the conversation fulfilled its purpose. The same logic applies to the broader U.S.-China dynamic.

Share

Leave a comment

On the evening of August 15, 2026, a long, low blast from a freighter’s horn echoes across Ningbo-Zhoushan Port (宁波舟山港) as a massive vessel called Dubai Tower (IMO 9433066) operated by Sea Legend Shipping (海杰航运) eases from its berth and sets a course for Britain’s Port of Felixstowe—not by the customary southern routes, but north, along the Russian Arctic coast.

After departing, it will successively call at key Northern European and Baltic ports, including Felixstowe (UK), Rotterdam (Netherlands), Hamburg (Germany) and Gdynia (Poland).

This is no routine sailing. It marks the inaugural voyage of the world’s first weekly liner service linking China and Europe via the Arctic, now entering regular commercial operation.

This voyage carries high-value, temperature-sensitive cargo such as energy storage cabinets, power batteries, photovoltaic modules, and new-energy components. The cargo sources are primarily concentrated in Yangtze River Delta cities like Ningbo and Yiwu, which are the peak-season export delivery demands of the European market.

During the precious golden window of open water between August and October each year, the service will run on a fixed weekly schedule, with a rotation of seven high-ice-class vessels shuttling between six major Chinese ports and key European hubs.

Sea Legend Shipping is set to deploy seven ice-strengthened vessels, with its entire fleet covering multiple capacity tiers ranging from approximately 962 TEU to 4,890 TEU, including ships of various ice classes such as ICE 1A, ICE 1C, and ICE 1D, some of which are LNG-powered. This year, it plans to operate eight consecutive voyages on a weekly schedule, achieving normalized operations on the Arctic route during the summer window period.

The materialization of this commercial route signals that Arctic shipping has moved decisively from pilot projects and theoretical feasibility into scaled, normalized business.

Yet beneath the compelling narrative of shortened voyages and streamlined logistics lies something far more consequential. The normalization of Arctic shipping is not merely a byproduct of better icebreakers or a warming planet. It is a concentrated expression of a shifting global geopolitical order, intensifying great-power rivalry, and a deepening crisis in the governance of the polar regions.

Along this frozen corridor, commercial calculation and national ambition intertwine; the fragility of the polar environment collides with humanity’s relentless push outward.

Subscribe now

A Ledger That Finally Adds Up

The fundamental driver behind the Arctic route’s transition from a theoretical shortcut to daily operations is simple: the economics of polar shipping have, against all odds, produced a viable business case. This is not merely a geographical convenience; it is a meticulously calculated cost-benefit equation.

First, there is the alluring arithmetic of time. From Ningbo-Zhoushan to Felixstowe via the Northeast Passage takes roughly 20 days—15 to 20 days faster than the traditional Suez Canal route, about five days quicker than the much-touted China-Europe rail freight, and a staggering 6,800 nautical miles shorter than the alternate route around Africa’s Cape of Good Hope. In an era when supply chains are judged by the hour, such compression means a significant boost in capital turnover.

Then comes the hard-nosed cost calculus. While Arctic transits require expensive icebreaker pilotage fees and pricier low-sulfur fuel, the drastically shorter distance substantially cuts overall fuel consumption. By comprehensive estimates, the per-container cost on the Arctic route is roughly 40 percent lower than that of rail freight; compared to the Suez Canal route, the per-container economics are at least competitive.

Moreover, the Arctic’s natural deep-freeze offers a distinct advantage for cold-chain logistics, reducing temperature-control energy use and cargo spoilage for high-value goods such as photovoltaic components, energy-storage equipment, and precision machinery.

On this maiden voyage, the containers included cross-border e-commerce products, solar panels, and power batteries—strong evidence of the route’s niche appeal in specific high-value supply chains.

“Last year, we observed the trial voyage of the Arctic route and were quite interested in it,” said Fan Zhiguo, logistics director at Huizhou EVE Energy Co., Ltd., on the maiden voyage. This time, the Chinese battery manufacturer is conducting a small-scale trial shipment of power battery products, hoping to fully test the entire process of transportation, temperature control and customs clearance during the inaugural voyage.

Third is the carbon ledger, which has become a matter of financial survival. With the European Union’s Carbon Border Adjustment Mechanism now in force, carbon emissions are no longer an abstract environmental slogan but a tangible line item on corporate balance sheets. The Arctic route cuts single-journey emissions by roughly 50 percent—a massive reduction that translates directly into green competitiveness for Chinese exporters, offering crucial breathing room against stiff carbon tariffs.

Fourth, and perhaps most urgently, is the security premium. In recent years, the Red Sea crisis, chronic congestion in the Strait of Malacca, and frequent geopolitical flare-ups have exposed the profound fragility of global supply chains. On July 23, 2026, a Houthi attack on two Saudi oil tankers sent global oil prices soaring back above $100 a barrel, underscoring the deteriorating stability of traditional maritime corridors.

The Arctic route neatly bypasses these high-risk chokepoints. This geopolitical risk premium is a powerful magnet pulling capital toward the pole.

Still, perspective is essential. As France’s Le Figaro has calculated, during the August–October 2026 shipping window, Chinese cargo volume on the route is estimated at roughly 20,000 TEUs, or 240,000 metric tons of goods. That pales against the Suez Canal’s annual throughput of 15 million TEUs. Thus, as many foreign analysts note, the Arctic route’s absolute commercial value in the short term remains modest; its core significance lies more in strategic hedging and symbolism.

Bypassing the Chokepoints

To the global shipping industry, the real impetus behind the Arctic route’s normalization is not simply saving money or days at sea. It is about securing a viable geopolitical alternative in a dangerously volatile world.

Today, the Red Sea, the Bab el-Mandeb Strait and other traditional maritime lifelines are under persistent threat. Iran-backed Houthi forces have repeatedly issued warnings and carried out attacks on commercial shipping. The security and predictability of the traditional Asia–Europe sea lanes have sharply deteriorated, insurance premiums have skyrocketed, and shipowners face enormous asset risks.

Geographically, the Arctic route sidesteps these powder kegs entirely. Sea Legend Shipping, a Chinese-backed shipping company established in Singapore in 2022 and headquartered in Shanghai, has explicitly stated that the route is insulated from regional political turmoil, ensuring safe and smooth cargo transport. Industry experts put it bluntly: geopolitical tensions have become the single strongest accelerant for Arctic shipping’s development.

The route’s viability also owes much to changing physical conditions. Global warming—an ecological catastrophe in its own right—has ironically opened a window for polar navigation. Research from the University of Southampton shows that Arctic winter sea-ice extent fell by 5.8 percent in 2024–2025, the largest decline on record. Although ice cover rebounded slightly in 2026, it remained at the second-lowest level ever recorded.

The U.S. National Center for Atmospheric Research predicts that the Arctic could see its first ice-free summer by 2030, with September sea ice possibly vanishing entirely by 2040. In fact, the Northeast Passage has already experienced ice-free conditions in summer since 2007, with ice-free periods lasting 30 to 40 days.

From late July to mid-August 2025, temperatures across much of the Arctic Circle remained unusually high for 14 consecutive days—the longest such stretch since 1961—and the Finnish city of Rovaniemi logged temperatures above 30°C for three straight weeks. Such extreme climatic shifts are fundamentally reshaping humanity’s physical relationship with the Arctic.

From Trial Runs to Routine

The journey from occasional “experimental voyages” to today’s regularized service has never been purely a market-driven affair. Behind it lies years of sustained infrastructure investment and systematic state engagement.

Consider the arc of progress. On August 15, 2013, the Yong Sheng, a vessel operated by China’s COSCO Shipping, departed from Taicang Port and arrived in Rotterdam roughly 27 days later—the first Chinese commercial transit of the Northern Sea Route.

By September 23, 2025, Sea Legend Shipping’s Istanbul Bridge had cut that transit to 20 days from Ningbo-Zhoushan to Felixstowe, a vivid testament to the route’s growing technical and commercial maturity.

Underpinning this leap is a deep, institutional synergy between China and Russia. In June 2024, China’s NewNew Shipping (新新航运公司) signed a cooperation agreement with Russia’s state nuclear energy corporation Rosatom to establish a joint shipbuilding venture dedicated to designing and building high-ice-class containerships.

The first five ARC7-class polar container vessels are already under construction, with the inaugural ship expected to launch in 2027. Separately, a China–Russia joint venture, Northern Sea Route Shipping Line (NSRSL), was formally established in late 2024, aiming to transition from summer–autumn seasonal operations to year-round service by 2027.

China has simultaneously upgraded its domestic support infrastructure. In July 2025, China Maritime Safety Administration under China’s Ministry of Transport added an “Arctic Route” module to its official systems; the following month, the Beihai Navigation Safety Center began broadcasting Arctic sea-ice forecasts to global shipping.

China now not only has ships capable of sailing the route but also the “eyes” and “brain”—meteorological, ice, and navigation services—to support them. In addition, China’s 15th Arctic scientific expedition in 2025 saw the Xue Long 2 (雪龙二号) and the manned submersible Jiaolong conduct their first deep dives under Arctic ice, gathering critical baseline data for safe navigation.

Russia, too, is mustering national resources for Arctic development. As the only country operating nuclear-powered icebreakers, Russia currently fields 45 icebreakers, five of which are nuclear-powered, and plans to expand that fleet to nine by 2035. The much-anticipated Leader-class heavy nuclear icebreaker is under construction at the Zvezda shipyard, with the first vessel expected to enter service by the end of 2027.

Russia is also advancing its “Floating Nuclear Power Plant” program, aiming to deploy five such units by 2027, two of which are being built by China’s Huisman Heavy Industry in Nantong. The results are tangible: in 2025, international transit cargo without port calls along the Northern Sea Route reached 3.2 million tons, a record high; Russian officials project that by 2026, total freight volume on the route could reach 37 to 40 million tons.

Despite the compelling business case, the global shipping industry remains deeply divided, with major European operators notably reluctant—even resistant.

Mediterranean Shipping Company, the world’s largest carrier, has explicitly ruled out using the Arctic route, citing fundamental safety concerns and unacceptable risks to crews. Denmark’s Maersk, which conducted a trial Arctic voyage in 2018, has since stayed away, citing limited economic returns and high ecological risks.

European hesitation springs from two main sources. First, there is genuine unease about the ecological fragility of the polar environment and the potential for irreversible damage. Second, and more decisive, is the shadow of the Russia–Ukraine war and the sanctions regime that followed.

European companies are loath to place strategic reliance on a route governed by Russian regulation, fearing blowback from geopolitical entanglements. Industry analysts note that traditional shipping lines value long-term reliability over short-term speed; many therefore prefer the well-trodden Cape route.

Nevertheless, the Arctic shipping race is drawing more players. South Korea has designated Arctic shipping as a key policy priority, planning trial voyages from Busan to Felixstowe. Supporting infrastructure is also booming: global orders for ice-class vessels reached 167 in 2025, a ten-year high, with another 164 slated for 2026—a clear indication that the supply chain is betting on the Arctic’s long-term potential.

China’s Century-Old Arctic Desire

China’s connection to the Arctic runs far deeper than most people realize. The normalization of this route is not merely a triumph of modern logistics; it is the echo of a century-long Chinese journey toward the high seas and the wider world. To grasp why China can so confidently assert its place in Arctic affairs, one must trace a remarkable historical thread.

In 1908, Kang Youwei (康有为), the late-Qing constitutional monarchy reformer then in exile, sailed to the Svalbard archipelago, reaching latitudes north of 84 degrees. In an age when China was weak and foreign gunboats ruled, the aged intellectual stood atop the planet, witnessed the midnight sun and composed poems in awe. He could hardly have imagined that, more than a century later, his homeland would steer steel leviathans through those same frozen waters with such commanding self-assurance.

The frozen realm that so moved Kang had once been a hunting ground for empires. In 1596, the Dutch navigator Willem Barents first sighted Svalbard; in 1607, the English explorer Henry Hudson discovered rich whale stocks there. Soon, whaling fleets from many nations converged, turning the archipelago into a bloody abattoir.

To impose order and prevent outright war over these resources, multiple countries signed the Spitsbergen Treaty in 1920 (now commonly known as the Svalbard Treaty), which recognized Norwegian sovereignty over the islands while granting all signatory states equal rights to commercial, industrial, scientific, and access activities in the archipelago and its territorial waters.

Here history takes a dramatic turn. As a victor power of World War I, China’s Northern Warlord government (北洋政府) received an invitation from France in 1925 and formally signed the Svalbard Treaty. But amid the chaos of warlord infighting and political instability, this vital instrument of national polar rights was shelved. It was never publicly reported; no one ever exercised the rights it conferred. And so China’s lawful Arctic stake lay dormant in the archives for nearly seven decades.

Not until 1991, when the Chinese scientist Gao Dengyi (高登义)—while browsing a copy of the Arctic Pilot gifted by a Norwegian friend—stumbled upon the astonishing fact that China was a signatory to the Svalbard Treaty. As Gao later recalled, the moment he read those words, he “shot up so abruptly that my head slammed hard against the bed frame.” That sharp thud awoke China’s dormant Arctic rights and threw open the door to the nation’s polar ambitions.

Armed with this discovery, Gao campaigned tirelessly, urging the establishment of a Chinese Arctic research station. He conducted multiple field expeditions to the Arctic, wrote detailed reports, and pressed the case with authorities.

His efforts gradually awakened China’s polar science community to the region’s strategic and scientific value. In 1999, the icebreaker Xue Long (雪龙) made China’s first Arctic voyage, lasting 71 days and covering 13,210 nautical miles. In 2004, China established the Yellow River Station in Svalbard, becoming the eighth country to maintain a research presence there.

At that moment, the yellowed treaty of 1925 at last became a red flag flying over the ice—China’s Arctic rights transformed from legal text into living reality.

By 2025, with the coordinated deployment of the Xue Long 2, Jidi (极地), Shenhai-1 (深海一号), and Tansuo-3 (探索三号) and the Jiaolong (蛟龙)’s first manned dives under Arctic ice, China’s polar capabilities had reached world-class levels. From Kang’s astonished gaze at the midnight sun, to Gao‘s sharp collision with a bed frame; from a forgotten treaty to the world’s first weekly China–Europe Arctic container service—this is not merely a technological leap. It is a nation’s century-long Arctic dream finally breaking through the ice.

Intensifying Great-Power Chess Match

The normalization of the Arctic route is also a major move on the geopolitical board. Today’s Arctic finds itself caught between collapsing governance frameworks and escalating great-power rivalry.

Geographically, the Arctic sea routes consist of three main passages: the commercially most advanced Northern Sea Route (NSR) along Russia’s Arctic coast; the Northwest Passage through Canada’s Arctic archipelago; and the Transpolar Sea Route crossing the central Arctic Ocean. The NSR, regulated by Russian law, is currently the most promising ice corridor. According to Rosatom chief Alexei Likhachev, transit cargo on the NSR reached 3.2 million tons in 2025, with 107 transits in total, including 15 container-ship transits—roughly five times the volume of a decade earlier.

Yet this commercial fervor stands in stark contrast to the severe fragmentation of Arctic governance. At the global level, the UN Convention on the Law of the Sea and the International Maritime Organization provide basic legal frameworks; at the regional level, the Arctic Council once served as the central platform for coordination.

But after the Ukraine crisis, geopolitical competition in the Arctic intensified sharply. With Finland and Sweden joining NATO, the buffer zone between Russia and the Alliance in the region has vanished, and the specter of military confrontation has returned.

The Arctic Council has been effectively paralyzed by boycotts, and the Ottawa Declaration’s original exclusion of traditional security issues now appears hopelessly naive against the harsh realities of power politics.

Resource competition remains a powerful lure. The U.S. Geological Survey estimates that the Arctic holds roughly 22 percent of the world’s undiscovered, recoverable oil and gas resources, including about 45 percent of global natural gas reserves and 25 percent of yet-to-be-discovered oil. Substantial deposits of diamonds, gold, silver, iron, nickel and rare-earth elements are also present.

These sanctions imposed after Ukraine crisis have hit Russian Arctic energy projects hard. Yamal LNG exports to Europe have plummeted, costing Russia an estimated $30 billion in lost revenue over three years; the Arctic LNG-2 project has been stalled by sanctions. Yet sanctions have not entirely halted great-power interest in the Arctic; behind the scenes, U.S.–Russia bargaining is taking new forms.

Since Donald Trump returned to the White House in January 2025, reports of prospective U.S.–Russia cooperation in the Arctic have surfaced repeatedly. During the February 18 Riyadh meeting between U.S. and Russian officials, Arctic issues unexpectedly became a topic of discussion; at the February 27 Istanbul meeting, the two sides held dedicated talks on Arctic energy and shipping development.

In March, President Vladimir Putin told Russia-1 television that Russia is ready to cooperate with foreign partners, including the United States, on rare earths and other critical minerals. Trump responded at a Cabinet meeting: “I'd like to buy minerals on Russian land too if we can.”

Even more striking are potential U.S.–Russia deals in natural gas. Russia has proposed resuming gas supplies to Europe via the U.S.; Washington has suggested American involvement in Russian energy projects and companies.

Options reportedly on the table include U.S. investment in both the Nord Stream pipeline projects and the Ukrainian transit pipeline; direct U.S. equity stakes in Gazprom; and U.S. purchases of Russian gas for re-export to Europe. Between March and April, media reports indicated that a U.S. energy consortium had drafted a framework agreement with Gazprom for post-sanctions cooperation, including a $3 billion acquisition of 49 percent of Nord Stream AG and a swap of sanctions relief for a 15 percent stake in Gazprom itself. Such transactional realpolitik is now reshaping the Arctic energy landscape.

At the same time, Arctic governance faces a triple dilemma: tension between global governance and regional jurisdiction; conflict between geopolitical competition and peaceful stability; and the impossible choice between resource development and ecological preservation.

The U.S. National Security Strategy released in 2025 openly dismissed climate governance, and in March of that year, the administration opened 1.56 million acres (0.6-million-hectare) of the Arctic National Wildlife Refuge to oil and gas leasing. This approach—treating the Arctic as a “resource hunting ground” and a “strategic fortress”—only deepens the crisis of polar governance.

The Risks: Technology, Emergency Response and the Fragile Ecosystem

Despite its clear advantages in speed and geopolitical hedging, the Arctic route remains fraught with serious operational and safety risks, as multiple foreign media reports and industry experts have warned.

Technologically, high-latitude navigation faces unstable GPS signals, weak communications, and rapidly changing ice and sea conditions—conditions that can swiftly trap or disable a vessel.

In terms of emergency response, the Arctic lies far from global search-and-rescue systems; salvage, firefighting, and wreck-removal capabilities are scarce. Any accident would be immensely difficult and costly to handle. Allianz data show that more than 500 maritime accidents occurred in Arctic waters over the past decade, nearly half caused by mechanical failure or equipment breakdown.

Moreover, the risk of oil spills and ecological pollution is exceptionally high, with potentially irreversible consequences for the fragile polar environment.

Operationally, the Northeast Passage’s shipping window remains concentrated in the summer–autumn melt season (July to October), with the minimum ice cover occurring around mid-September. Outside that period, icebreaker escort is required, limiting the window significantly. In terms of volume, the Arctic route’s current cargo throughput is still modest compared to Suez’s 15 million TEUs; it is far from capable of serving as a global supply-chain backbone.

Geopolitically, the route is essentially controlled by Russia, exposing operators to secondary sanctions risks—the very reason European carriers, wary of entanglement with Moscow, have held back.

Industry experts conclude that the Arctic route can serve as a useful supplement to traditional shipping lanes, diversifying Asia–Europe trade corridors, but given its short navigational window, high operational risks, limited scale, and inadequate support infrastructure, it is unlikely to displace conventional mainline routes for the foreseeable future.

China’s Arctic Position and the Path Ahead

Amid the growing complexity of Arctic affairs, China has maintained a steady, principled stance. In January 2018, Beijing issued its first Arctic policy white paper, defining itself as a “close stakeholder” and a “contributor, participant, and builder” in Arctic affairs.

This identity rests both on its legal status as a signatory to the Svalbard Treaty and on its substantive contributions to Arctic science, environmental protection, and shipping.

Beijing maintains that while the sovereignty of Arctic states should be respected, the legitimate rights of non-Arctic states should also be guaranteed; and that development must be pursued within the bounds of ecological sustainability.

Yet China’s Arctic overtures have drawn a wary, and often hostile, reception in the West. The successful 2025 Arctic test voyage and the Jiaolong’s deep dives under polar ice were misinterpreted by some as “military reconnaissance” and “resource grabbing.

For the foreseeable future, this is unlikely to change. Over the nearer horizon, three scenarios present themselves.

The first—and most probable—is that the Arctic becomes neither China’s backyard nor NATO’s inland sea, but settles into a brittle equilibrium of routine confrontation and grudging division of labor.

The second, considerably less likely, is that the polar region turns into the frozen front of a new Cold War, with global governance collapsing altogether.

The third and least probable—but most transformative—is that the Arctic, under a reinvigorated global framework, transforms from a frozen commons into an open high seas.

Three variables will determine which path prevails.

The first is the pace of warming. If an ice-free Arctic summer truly arrives by 2030, the physical barriers that have long shielded the region from mass commercial traffic will crumble. That would be a decisive advantage for China, which commands the world’s most formidable shipbuilding industry. If, on the other hand, warming stalls, Russia’s nuclear-icebreaker monopoly will remain unassailable.

The second is the relationship of U.S.-Russia relations. Should a Trumpian brand of transactional diplomacy persist, Washington and Moscow may cut private deals over Arctic resources and shipping rights—more likely leaving Beijing sidelined. But if tensions between the two powers remain high, China’s leverage as Russia’s indispensable polar partner swells considerably.

The third and perhaps most critical variable is China’s own strategic discipline. Beijing’s greatest asset in the Arctic is not its fleet of icebreakers, but a piece of paper: the Svalbard Treaty of 1920, which grants signatories equal rights to commercial and scientific activities in the archipelago.

If China overreaches—if it tries to build dual-use facilities in Greenland or elsewhere that could be construed as military in nature—it will hand Western critics the evidence they need to cement the “militarization” narrative and push for China’s expulsion from the Arctic Council’s observer ranks.

But if Beijing stays the course—science first, commerce second, politics kept at arm’s length—it may gradually earn the status of a de facto stakeholder, not by fiat, but by sheer persistence.

In this sense, the container ship now sailing from Ningbo-Zhoushan is precisely a waypoint in this strategy of trade primacy—a calculated test of how far ice and diplomacy will bend.

Share

Leave a comment

Initially, distillation was discussed as a technique, but later, the word “attacks” was appended to it. Before you settle on a stance, it's necessary to look into exactly what it is.

This week, China’s financial podcast LatePost published an in-depth analysis on AI distillation by Cheng Manqi. It is by far the most comprehensive piece I’ve seen on this hot-button topic.

Amidst the backlash from the U.S. government and select AI companies against Chinese LLM distillation, this piece doesn't try to mount a direct defense. Rather, it shines a light on the technical nuances that have been buried under political rhetoric and commercial rivalry.

Subscribe now

Some view it as an undignified form of theft. Others argue it has been unfairly stigmatized by a few industry leaders acting out of self-interest, and that it is, fundamentally, just an optimization technique.

Over the past few months, many loose threads in the AI industry have converged on a single focal point: distillation.

Recent events and shifts tied to this technology include: open-source models inching closer to the most powerful closed-source models; a public letter signed by 77 U.S. companies opposing hasty restrictions on open-source models; Anthropic accusing Chinese companies of using massive networks of fraudulent accounts to siphon data (on two separate occasions); a $1.5 billion copyright infringement settlement; and Zhang Yiming’s explicit pledge against distillation at a company-wide meeting for ByteDance’s Seed team.

This old technology has been repeatedly discussed, debated, and often misunderstood in 2026. But what exactly is distillation? How is it executed at a massive scale? Can it serve as a true competitive moat for a model development team? And what are the costs?

To answer these questions, we interviewed nearly a dozen model researchers and industry practitioners from various companies, and synthesized their insights with public research and technical reports to reconstruct the past, present, and broader implications of distillation.

The Origins of Distillation: Built for Compression, Not Superiority

Distillation is not plagiarism. It does not involve stealing software code, nor can it directly extract another model’s weights or complete training datasets.

The type of distillation currently at the center of controversy—the kind used to make models stronger—is technically a method for acquiring high-quality data. It involves repeatedly querying a more powerful “teacher model” to generate responses, and then using these “prompt-response” pairs to train a “student model” until it achieves comparable performance.

The concept is far from new. In 2015, Geoffrey Hinton, who had recently joined Google Brain, alongside Jeff Dean (then head of Google Brain) and young researcher Oriol Vinyals, published Distilling the Knowledge in a Neural Network. This paper was the first to formalize earlier concepts of model compression (dating back to 2006) under the term “distillation.”

This was nearly two and a half years before Google introduced the Transformer architecture that now underpins large language models. At the time, Hinton and his colleagues applied distillation to image recognition models. The method involved having the student model learn the probability distribution of the teacher model’s outputs. For instance, in identifying a cat, deep learning is a statistical process: 0.7 probability for cat, 0.2 for fox, 0.1 for dog.

The student model learns from this entire distribution of probabilities rather than just the final label. This is known as learning the “logits” (the raw scores that, after a Softmax conversion, become a probability distribution), a process referred to as “soft distillation.”

Soft distillation is typically “white-box,” as it requires the teacher model to fully expose its output probabilities to the student model.

Historically, this type of distillation occurred within the same organization, and its primary goal was not to enhance performance, but to compress it—using a model with fewer parameters to approximate the capabilities of a much larger one. While this inevitably results in some performance loss, it enables faster and cheaper inference.

To this day, compression remains one of the most typical use cases for distillation.

In autonomous driving, for example, companies first train highly capable, large-scale cloud models, and then use distillation and pruning to shrink them into smaller models that can run on edge chips in vehicles. Companies like Li Auto and XPeng have implemented this approach.

A more recent example is DeepSeek-R1 in early 2025. DeepSeek simultaneously released six smaller distilled models, all trained using the 671-billion-parameter R1 as the teacher. Four of these student models were built on Alibaba’s Qwen2.5, while two were based on Meta’s Llama 3, with parameter counts ranging from 1.5 billion to 70 billion.

DeepSeek first had R1 generate approximately 600,000 “prompt-reasoning process-answer” reasoning chains and 200,000 non-reasoning data points. During the post-training phase, this data was used for supervised fine-tuning (SFT) on the six smaller base models, endowing them with significantly enhanced reasoning capabilities.

Industrial-Scale Distillation Attacks

While distillation is a neutral technical method when used for compression, by February 2026, Google and Anthropic began publishing articles that paired the word “distillation” with “attack,” accusing certain companies of using it for unfair competition.

In February and June, Anthropic alleged that DeepSeek, Moonshot AI (Kimi), MiniMax and Alibaba’s Qwen collectively used roughly 50,000 fraudulent accounts to conduct over 44.8 million interactions with Claude in an attempt to extract its capabilities.

Why has large-scale distillation of leading closed-source models become so prevalent? The root of this shift lies in technological evolution, which can be traced through three key developments:

From Soft to Hard Distillation: From White-Box to Black-Box

In 2016, while pursuing his Ph.D. at Harvard, Yoon Kim and his advisor Sasha Rush introduced sequence-level knowledge distillation, adapting the technique originally used in image recognition to language translation tasks.

Sequence-level distillation no longer requires learning the probability distribution of the model’s intermediate steps. Instead, the teacher model first generates high-quality translations, and the student model learns from the complete “source-target” sequence pairs.

Though originally proposed to compress massive translation models and improve decoding speed, this innovation had a profound side effect: distillation no longer required access to the teacher model’s step-by-step probability distributions. It could be achieved simply by looking at the “final answer.”

This is “hard distillation,” and it can be performed via a black-box approach—meaning one only needs to call an API to receive the teacher model’s responses.

Sasha Rush later joined the AI coding company Cursor in March 2025. In June of this year, Cursor was acquired by SpaceXAI (the newly merged entity of SpaceX and xAI) for $60 billion, and its team has since been fully integrated into the company.

The Rise of Reasoning Models

In September 2024, just as the industry was awash in debates about scaling laws hitting a wall, OpenAI released its o1 reasoning model. The model introduced two major shifts:

First, it applied large-scale reinforcement learning (RL) during the post-training phase, teaching models to develop actual reasoning strategies. Second, during inference—the actual usage phase—it allocated more test-time compute, prompting the model to generate longer chains of thought when tackling complex problems, which in turn continued to boost its performance.

Both of these shifts supercharge the effects of distillation. For one, large-scale distillation is primarily a post-training technique; as post-training grows in importance, so does the return on investment for distillation. Furthermore, test-time compute doesn’t just yield a final answer. It also produces long-form reasoning processes, including chains of thought, tool calls, search trajectories, and self-correction steps. All of these outputs serve as premium raw data for distillation.

Four months later, DeepSeek’s technical report for R1 laid out the distillation process in even greater detail, revealing exactly how distillation makes models stronger. The R1 release highlighted several specific practices and findings:

First, the distillation process relied heavily on R1-generated “prompt-reasoning process-answer” data pairs. This is currently the gold standard for distillation, yielding better results than simple “prompt-answer” pairs.

DeepSeek, notably open in its approach, released R1 with its full chain of thought visible, stating it hoped this would “help the community distill better small models.”

Meanwhile, leading closed-source players like OpenAI, Anthropic and Google DeepMind have consistently hidden their full chains of thought and reasoning traces from users.

Because of this, industry insiders often talk about distillation in terms of reverse engineering.

Just this Monday (August 10th), researchers from the University of Tübingen in Germany and other institutions published a paper titled Stealing Reasoning Traces from Proprietary LLM APIs, detailing the methods they discovered to reconstruct these hidden reasoning traces.

It’s hardly a secret anymore—even if closed-source model providers actively try to hide them, chains of thought and reasoning traces can be reverse-engineered.

Fundamentally, these traces are just part of the model’s output, a byproduct of the inference stage. Whenever you interact with a model, its reasoning process is generated and leaves a digital footprint. This still falls under the umbrella of black-box and hard distillation.

Another key finding from R1 is that, during post-training, direct distillation yields greater performance gains than having a model learn through reinforcement learning (RL) on its own.

DeepSeek ran a comparative experiment using Qwen2.5-32B:

• Running over 10,000 steps of large-scale RL on Qwen-32B-Base resulted in a model that scored 47.0% on AIME 2024 (a benchmark for math problem-solving).

• Supervised fine-tuning (SFT) Qwen-32B-Base with 800,000 R1-generated data points resulted in a model that scored 72.6% on AIME 2024—a massive leap of over 25 percentage points compared to the RL approach.

According to the R1 technical report, they demonstrate that the reasoning patterns of large models can be distilled into small models, achieving superior performance compared to reasoning patterns acquired by small models through reinforcement learning,

While this is the result of a specific experiment and may not apply universally, it is a highly attractive finding. Long-horizon RL is inherently more difficult to train than SFT; it demands more robust infrastructure, is typically slower, and carries a significantly higher compute cost.

DeepSeek effectively published a playbook for smaller or less capable models to improve their reasoning abilities in a relatively economical, efficient, and predictable manner.

Multiple industry practitioners note that recent best practices lean heavily on SFT during post-training, with little to no RL, and still achieve excellent results.

Following the release of DeepSeek-R1, and throughout 2025 to the present, companies like Anthropic, OpenAI and Google have reported an increasing number of detected “distillation attacks.”

During this same period, deeper exploration into post-training popularized another, non-controversial distillation practice: on-policy distillation, which is currently primarily used for integrating capabilities during post-training.

The difference between on-policy and off-policy distillation lies in who generates the data. Most of the distillation mentioned earlier—where a model learns from another closed-source model—is off-policy, with the data generated entirely by the teacher model. In on-policy distillation, the student model generates its own reasoning and answers, while the teacher model provides feedback. This feedback can be in the form of token-by-token probability distributions (white-box) or evaluations of the reasoning traces and final answers (black-box).

Since the second half of 2025, Alibaba’s Qwen, Thinking Machines Lab, and Xiaomi’s MiMo have all contributed practices and improvements to on-policy distillation.

In its early-2026 MiMo V2-Flash technical report, Xiaomi introduced MOPD (Multi-Teacher On-Policy Distillation), later publishing a dedicated paper in June. The method involves training separate teacher models for math, coding and tool use, and then having the student model generate its own trajectories, receiving feedback from different teachers depending on the task. It addresses a new post-training challenge: directly mixing multiple capabilities into a single RL training loop often causes them to interfere with one another, leading to a zero-sum game.

Technical reports for both DeepSeek-V4 and Kimi K3 state that they used the MOPD approach during post-training to merge multiple expert models.

Using AI to Accelerate AI

The third thread driving the scale of distillation is that the process itself is becoming increasingly automated as AI capabilities improve.

Let’s return to the core of distillation: the “prompt-reasoning process-answer” data pairs.

First, the prompting stage can be shifted from humans to AI. Since 2022, research like Self-Instruct has been tackling the problem of high-quality prompts being scarce and expensive. In practice, companies can first filter high-quality real prompts from authorized user behavior, then use those as a seed to generate a larger volume of AI prompts. These are then fed to the teacher model at a much higher frequency to generate more responses. Think of it as using a small amount of yeast (real data) to leaven a much larger batch of dough.

The entire “prompt-reasoning-answer” data pairs can be rewritten and augmented in the same way. This reflects a foundational mindset in today’s AI industry: using AI and models to automate and accelerate AI itself.

Many specific steps in the distillation pipeline can leverage AI. This includes filtering high-quality prompts from massive volumes of real queries, evaluating “what constitutes high-quality data,” and increasing the diversity of synthetic data. Some of these steps can be handled directly by models, while others can be accelerated and optimized using increasingly powerful AI coding capabilities to build and refine various systems.

Since 2025, the methods and practices for large-scale distillation have matured. Its effectiveness and necessity have become increasingly apparent as the importance of post-training has grown.

Discussions surrounding “distillation” have also spilled out of the technical AI community. In public discourse, the concept has been simplified, misunderstood and even distorted. Distillation is no longer just a technical issue; it has moved to the center of the storm.

Misconceptions About Distillation: It’s Not a Silver Bullet, Nor a Secret

With a clearer understanding of how distillation works, we are better equipped to navigate the debates surrounding it.

Distillation is not a silver bullet, nor the most critical factor in model performance.
As mentioned repeatedly, the distillation that makes large language models stronger primarily occurs during post-training (and mid-training). However, a model’s overall performance is the result of the entire pipeline, from pre-training to post-training. The general consensus is that pre-training carries more weight. Therefore, distillation is not a silver bullet, nor is it the most decisive factor in a model’s performance.

In DeepSeek-R1’s technical report, using 800,000 R1 data points for distillation, the Qwen2.5-32B base model scored 72.6% on AIME 2024, surpassing the 70.0% achieved by Llama-3.3-70B-Instruct after distillation—even though Llama had more than double the parameters.

Following the release of K3, Ai2 researcher Nathan Lambert replied to a tweet about K3 topping the Frontend Code Arena leaderboard (which evaluates frontend coding skills), pushing back against the “it’s just distillation” talking point, and insisting that people finally need to acknowledge China's elite model-building capabilities.

Reaching a certain level of pre-training capability is the foundational prerequisite for Chinese open-source models like K3, GLM-5.2, and DeepSeek-V4 to achieve their current performance.

Distillation requires significant operations, experience and engineering know-how.
Distillation is often framed as a shortcut, and shortcuts imply ease. In reality, executing large-scale distillation today is a highly complex systems engineering challenge.

Based on descriptions from multiple practitioners, large-scale distillation has several difficult bottlenecks:

First, the ability to stably, frequently, and massively call leading models, combined with user operations. One commonly cited approach is setting up numerous proxy relays, offering discounts or other incentives to attract a specific cohort of users with genuine usage patterns—such as senior developers or STEM students and researchers handling heavy scientific workloads. In their daily use, they naturally generate high-quality, multi-turn prompts in real-world scenarios and receive model responses. These prompt-response pairs are then filtered and processed to serve as the seed for fermenting and augmenting more data.

This tests a team’s operational capabilities (knowing where high-quality users are and how to reach them) as much as its engineering capabilities (building a system stable enough to handle the load). It also requires ecosystem capabilities, as the entire process may involve partnerships with third-party companies or institutions.

Second, the team’s own ability to construct high-quality prompts and tasks. This requires a deep understanding of tasks, data, and the performance boundaries of current leading models. It overlaps significantly with the skills needed for large-model training itself.

Third, knowing how to effectively use the data. This requires building a robust data pipeline: ensuring a reasonable task distribution, and figuring out how to sample, filter, deduplicate, augment, and correct data, as well as determining the right formats and ratios.

This pipeline has measurable optimization metrics, such as what percentage of raw data ultimately makes it into post-training, and the efficiency and quality of the augmentation. The quality of the data pipeline directly impacts effectiveness, efficiency, and cost.

Continuous distillation is also incredibly expensive. Industry rumors suggest that companies’ spending and budgets for distillation this year range anywhere from $100 million to $1 billion.

Distillation isn’t a simple button you press to see a sudden leap in model performance. There are massive implementation challenges. You still have to calculate the return on investment.

Is Distillation a Moat for Model Teams?

So, given its relative complexity, can large-scale distillation serve as a competitive moat for a model development team?

Most practitioners we spoke with—whether from companies rumored to be distilling or those that aren’t—share a consistent view: for first-tier companies, distillation does not constitute a long-term moat.

Like many large-model technologies, the concepts and practices of distillation will gradually diffuse. Through talent mobility, open-source sharing, conference exchanges, and third-party service providers looking for clients, people and information in the AI space are constantly flowing. We repeatedly heard the same sentiment from researchers: “There are no real secrets in the large-model space.”

The competitive advantage provided by technical methods is mostly a first-mover advantage; those who do it first gain more experience. But it is not an insurmountable, winner-take-all moat like a network effect.

In the large-model space, the phenomenon truly considered to have a strong competitive moat is the data flywheel: if a company’s model is strong enough to reach a massive user base performing high-difficulty tasks, it will attract a continuous inflow of high-quality data. This data is unique, proprietary and unavailable to others. Once processed, it can be used to train an even stronger model, which in turn attracts more users to tackle even harder tasks.

In this flywheel logic, applications that interact directly with users hold immense value. For example, based on user agreements and permission settings, applications like Cursor, Devin, and Manus may be able to access more complete data and user behavior than the underlying models they call.

On August 12, following the release of Grok 4.6, Elon Musk replied to a tweet noting that Devin had integrated Grok 4.6, stating: “Grok 4.7 will exceed all current models.” He added, “SpaceX’s training corpus is so awesome & unique.” At $60 billion, the Cursor acquisition is starting to look like a bargain.

But the data flywheel has its own controversies. While model and application companies can collect data, can they actually use it for training? And as tasks become more difficult and scenarios more high-value, will customers and users be more inclined to keep that data to themselves? Furthermore, as the user base expands and becomes more diverse, is the process of panning for gold still worth the effort?

Some practitioners argue that in AI applications geared toward daily life assistance and entertainment, the vast majority of user-generated data is essentially garbage for training stronger models.

The Internal Dividing Line: Black Box vs. White Box

Perhaps the most fascinating phenomenon regarding distillation is that while no one wants to talk about it openly, most industry insiders we spoke to do not secretly believe it is a shameful practice that violates their technical principles.

The internal dividing line lies in the distinction between black-box and white-box distillation.

Currently, distillation targeting closed-source models is entirely black-box. The data obtained consists of outputs generated during the model’s public use—it is a byproduct of the model as a product.

So why shouldn’t other companies be able to act as users, asking the model questions and receiving answers? After all, real money changes hands for these queries and responses. (Of course, in practice, companies employ various tactics to exploit loopholes and drive down costs.)

The more contentious part involves reasoning traces. Since most models hide their complete chains of thought, the distilling party must use technical means to infer and reconstruct them. Yet, these reasoning traces remain a product of the inference stage.

Anthropic, Google and OpenAI will argue: “Our Terms of Service explicitly state that competitors cannot use our models to train or improve their own.”

But who was sued by The New York Times, accused of copying and using the newspaper’s 170-year archive of news built by generations of journalists and authors without permission? OpenAI.

Who downloaded massive volumes of books from piracy platforms, refused to pay for a single one, and was recently hit with a class-action lawsuit by American authors, resulting in a $1.5 billion settlement? Anthropic.

Even some American AI practitioners consider companies like Anthropic to be hypocritical. It echoes a scenario from Christopher Nolan’s new film Odyssey: you sent the Trojan Horse into the city, and now your own homeland is being invaded by forces from across the sea.

Large-scale distillation is a new issue born from recent technological developments. It is difficult to convince the majority that it is an unacceptable, immoral, or even evil act simply because a few companies claim it violates their Terms of Service.

Moreover, violating a Terms of Service agreement does not equate to legal infringement; it also involves other legal regulations and jurisdictional issues.

In its broadest sense—using data to improve models—distillation has become ubiquitous.

A widely circulated method for verifying distillation is actually ineffective: if you ask a model “Who are you?” and Model A claims to be Model B, that is not definitive proof that A distilled B. During the pre-training phase, everyone uses vast amounts of public internet data, much of which already consists of outputs generated by various models.

In July, when asked about distillation during an Axios interview, Nvidia founder Jensen Huang said that, broadly speaking, distillation, learning from AI, learning from other sources of knowledge, is a fundamental principle of intelligence.

For most model companies, including some American ones, the accelerator has been pressed. Few are likely to voluntarily pump the brakes on distillation in the near term.

Why ByteDance Chose Not to Distill?

While most companies prefer not to discuss distillation publicly, ByteDance is an exception.

At a company-wide meeting for the Seed team in late July, ByteDance founder Zhang Yiming explicitly stated his opposition to distillation.

ByteDance has adjusted its stance on distillation over time. In late 2023, it became the first major Chinese LLM vendor publicly flagged by OpenAI for allegedly using GPT outputs to improve its own models non-compliantly. At the time, large-scale distillation was not yet prevalent, and OpenAI itself noted that ByteDance’s API usage was “minimal.”

ByteDance later responded that GPT-generated data had been used for model annotation and evaluation, but the relevant data was removed from the training set by mid-2023.

According to LatePost, for over two years following this incident—including throughout 2025, when the scale of distillation expanded rapidly—ByteDance’s Seed team did not distill leading closed-source models. Instead, it acquired data through other means, such as hiring top prize winners from math and computer science competitions at high salaries to construct and annotate data.

During this same period, the business controversies surrounding TikTok in the U.S. remained unresolved. The restructuring deal for TikTok’s U.S. operations was only officially completed in late January 2026.

Around the Spring Festival in 2026, Seed experienced a period of wavering. At the time, OpenClaw was gaining massive traction, Anthropic’s revenue was surging, and the usage of Chinese open-source models like GLM-5 was skyrocketing. The quantitative improvements in coding capabilities had triggered a qualitative leap, while ByteDance’s models were relatively lagging in this area.

After a period of deliberation, based on Zhang’s remarks at the Seed all-hands meeting, ByteDance has made its choice: it will not distill external leading models.

According to an exclusive report by LatePost, Zhang’s core message at the all-hands meeting was:

Distillation can improve model performance in the short term, but it is fundamentally just replicating capabilities that Claude already possesses. Following this path means we can only continuously approach our competitors; it will be very difficult to truly surpass them.

He also wants Seed to build its own AGI moat from a more foundational level.

Even if refusing to distill means the company will temporarily fall behind domestic competitors technically, ByteDance will not take this shortcut to advance its model capabilities.

This leads to another debate regarding distillation: Is it truly impossible for a distilled model to surpass its teacher?

We asked multiple practitioners this question, and their answers were remarkably similar: Technically, it is not impossible, and there are potential organizational hazards.

Distillation has inherent technical flaws. For instance, a student model might learn the teacher model’s errors, biases, refusal habits and expression patterns.

However, distillation is only one part of model training. There are many other areas for improvement in the complete training process: pre-training data, architecture, algorithms, and infrastructure... If multiple optimizations accumulate, is it possible for a student model to outperform a specific teacher model?

Some research has shown that, in certain specific tasks, student models can indeed surpass their teachers. For example, in December 2024, Microsoft released Phi-4, a 14-billion-parameter model trained heavily on synthetic data from teacher models like GPT-4o. Phi-4 outperformed GPT-4o on two benchmarks:

• GPQA (graduate and PhD-level scientific knowledge and reasoning): Phi-4 scored 56.1%, compared to GPT-4o’s 50.6%.

• MATH (math problem-solving and reasoning): Phi-4 scored 80.4%, compared to GPT-4o’s 74.6%.

Notably, as a smaller-parameter model, Phi-4 directly incorporated teacher-generated data into its pre-training. However, when training ultra-large-scale models with trillions of parameters, it is difficult to use distilled data for the massive pre-training phase. Compared to processing web pages, code, and books directly, fetching teacher-generated data point-by-point is slow and expensive.

Nevertheless, this raises an intriguing hypothesis: as model inference speeds increase dramatically and prices plummet, could distilled data—or data constructed with the help of stronger models—be integrated more heavily into pre-training?

Could multi-teacher distillation be used to make models stronger? Theoretically, a student model could learn from multiple top-tier models like Claude and GPT simultaneously. This might introduce new technical challenges, such as training instability and interference between different capabilities when distilling from different base models.

Even more radical speculations include: Can leading companies achieve a “bootstrapping” effect by distilling their own models, essentially lifting themselves up by their own bootstraps?

From a purely research perspective, “whether a model using distillation can surpass its teacher” remains an open question to be verified and explored.

Zhang and ByteDance have provided their answer through action: they believe it cannot be done.

This likely ties back to the organizational hazards mentioned by multiple practitioners. Distillation is a relatively economical and fast-acting method. An athlete might theoretically take performance-enhancing drugs while maintaining a rigorous training regimen. But in reality, it is rarely sustainable, as it breeds complacency and laziness.

When a team allocates a disproportionate amount of its attention and resources to distillation during a given period, projects and individuals exploring more uncertain, long-term directions may be starved of resources and recognition.

ByteDance is placing a bet: mastering data capabilities independent of external rivals will build a more robust and long-term technological advantage in model development.

Starting in June, ByteDance began restructuring its data team, establishing a first-tier AI department, “AI Data and Security,” parallel to Seed and Flow.

As mentioned earlier, most people do not believe distillation constitutes a long-term moat. In the large-model space, the true strong moat is the data flywheel.

Data has also become a highly valuable, independent business segment. For example, the U.S. company Mercor primarily helps large-model companies find scientists, PhDs, and other professionals for data construction, annotation, and model evaluation. Reports indicate that in July, Mercor was seeking a new funding round at a $20 billion valuation.

In China, companies starting from data construction have already reached $3 billion valuations.

Under our exclusive report on “ByteDance’s No-Distillation Stance,” one comment stood out: “He [Zhang] might not be the most tech-savvy, but he is most likely the most attuned to human nature.”

No One Can Easily Hit the Brakes

When someone accuses you of wrongdoing, replying with “Didn’t you do the exact same thing?” does not resolve the conflict.

At this very moment, open-source models are increasingly matching the performance of leading closed-source models. And the objective reality is that the strongest open-source models come from China, while the strongest closed-source models come from the U.S.

Previously, Zhipu.ai was placed on the U.S. Commerce Department’s Entity List in January 2025. The Intelligence Authorization Act for Fiscal Year 2026, which took effect later that year, mandated the removal of DeepSeek from U.S. intelligence, national security systems, and related suppliers.

By April of this year, the U.S. House Committee on Homeland Security and the Select Committee on Strategic Competition between the United States and the Chinese Communist Party began investigating U.S. companies’ use of Chinese models, questioning why Cursor used Kimi K2.5 as the base for Composer, and why Airbnb used Alibaba’s Qwen for customer service.

Following the release of Kimi K3 on July 16, reports emerged that the U.S. government is considering restricting or even banning certain Chinese open-source models. On July 24, 77 companies and organizations, including Microsoft, Nvidia, Meta and Fireworks AI, successively co-signed an open letter titled Open Weights and American AI Leadership, opposing hasty restrictions on open-source models.

Jensen Huang’s very first tweet was sharing this open letter.

New restriction measures discussed in recent months are still under consideration and have not yet been implemented.

Companies like Anthropic, OpenAI, and Google are also adopting stricter technical measures to identify and ban accounts suspected of being used for distillation. Anthropic, for instance, stated it has built classifiers and behavioral fingerprinting systems to detect distillation traffic, identifying cross-account coordination, repetitive prompting, and reasoning-trace extraction. It is also strengthening identity verification for accounts belonging to educational, research, and startup entities.

The methods for reconstructing reasoning traces in the University of Tübingen’s “stealing reasoning tracks” study mentioned earlier have already been reported to the relevant closed-source model companies by the research team. By the time the team uploaded their paper, some of those methods had already been rendered ineffective.

All parties involved are making their own choices and preparations. The upcoming shifts will transform the entire industry chain, from compute, cloud services, and infrastructure to models, applications, and customer deployment.

The storm continues. The next eye of the storm may not be distillation; it is, after all, just one of many methods for optimizing models.

With so many companies investing massive resources and scrambling to win this race for intelligence, what is the ultimate goal? These are the questions that some are already beginning to ask.

Share

Leave a comment

A few days ago, I published an article covering Wang Xingxing’s background and Unitree’s product lineup. Despite the FCC’s ban on exporting its new models to the US, this Chinese robotics firm is still highly sought after by the capital markets—and there’s clearly more to the story.

Today’s piece, adapted from 36Kr by Qiu Xiaofen and Yuan Silai, dives into how Unitree evolved from a startup into a commercial powerhouse.

In the tourist city of Hangzhou, Unitree’s headquarters has become an attraction in its own right. Tour groups arrive year-round, posing for photos beneath banners at the entrance. The company’s exhibition hall occasionally opens to individual visitors, though during peak periods, the wait reportedly stretches to more than a month. So many third-party agencies have been charging fees to bring visitors—often under the guise of “official partnerships”—that Unitree has been forced to post conspicuous notices: UNITREE DOES NOT OFFER PAID VISIT SERVICES!

New residential developments nearby have begun marketing themselves as “Unitree’s Choice,” banking on rumors that early employees will be able to buy apartments in Binjiang District once the company goes public.

And Wang Xingxing, the 35-year-old founder, has become an unlikely celebrity. Internet forums dissect his wardrobe choices (black suit trousers, no belt), his interview mannerisms (hands pressed together, Tim Cook-style), and even his handwriting (forceful strokes revealing unique personal charisma). Netizens debate whether his Myers-Briggs type is INTP (introverted, intuitive, thinking, and prospecting) or INTJ (introverted, intuitive, thinking, and judging).

This summer, that public fervor collided with capital markets. In June 2026, Unitree Technology’s IPO application on Shanghai’s STAR Market sailed through the exchange’s listing committee in just 73 days—a “lightning-fast” approval.

The offering values the company at 61 billion yuan ($8.5 billion). Yet investors are already talking about a market capitalization exceeding 200 billion yuan, with some suggesting 400 billion is possible in the near term.

“Compared with overseas peers, that’s not an exaggeration,” said Yu Wenchao, a partner at Dunhong Capital and an early Unitree investor. More than one backer has compared Unitree to Apple and Wang to Elon Musk. “After the listing, anything is possible.”

Others find the euphoria baffling. “Have they gone mad? People in this circle have become this crazy?” said one investor who passed on Unitree. “Hundreds of billions? That’s absurd. Their 2025 revenue was only 1.7 billion yuan!”

Never has a hardware company commanded such frenzied attention with such modest sales. When Unitree’s shares began trading on August 10, the lottery rate for retail investors was less than 0.02%—equivalent to a static price-to-earnings ratio of roughly 219 times. The average P/E for its sector, “general equipment manufacturing,” is 38.56.

Unitree’s valuation rests on two foundations: the company’s own achievements, and the tectonic forces beneath it—national strategy, industry bets, and public emotion. Robotics has been written into China’s 15th Five-Year Plan. State-backed institutions have piled in. The listing pipeline has opened. Over the past three years, funding for embodied intelligence has surged exponentially, creating dozens of companies valued in the tens of billions.

For the public who watched Unitree robots dance on the Spring Festival Gala and dream of robot nannies; for investors and stock traders hoping for a windfall; and for industry peers watching their own valuations hinge on Unitree’s performance—the company has become a totem. If it sustains a hundred-billion-yuan valuation, backers of other embodied-AI startups can sleep soundly. If it crashes, the entire industry will shudder.

As the listing drumbeat intensifies, everyone waits for Unitree to deliver its verdict on what the public market will pay for China’s robot dream.

The Year Everyone Neglected Unitree

In March 2019, when Decent Capital’s team met Wang Xingxing, Unitree had less than a few hundred thousand yuan in the bank. Yet Wang seemed remarkably unhurried.

The office was two cramped rooms in a dilapidated building—perhaps 20 or 30 square meters—resembling a laboratory more than a company. Electronic components covered the walls; parts from robotic dogs littered the floor and desks. There was barely room to stand.

The meeting moved to the lobby downstairs, where automatic doors opened and closed in the January chill, letting in gusts of freezing air. “It was so cold!” recalled Zhou Hualin, a director at Decent.

Wang wasn’t particularly eager for money. He wanted 4 million yuan. Zeng Liqing, Decent’s founder, insisted on giving him more than 10 million. By then, Wang had been building Unitree for two years, and the company was valued at less than 100 million yuan. He had already been rejected by countless investors.

Investing in Unitree was not the conventional choice. In Hangzhou, another robotics firm, DEEPRobotics, was widely considered “China’s Boston Dynamics.” Its founder, Zhu Qiuguo, was a professor at Zhejiang University’s prestigious control science department, had led national-level projects, and had never struggled for funding. In 2019, DeepRobotics was valued at 200–300 million yuan—more than double Unitree’s price tag. “Zhejiang’s control science program was famous. Investors naturally chose the founder with the better pedigree,” one investor recalled.

Against this backdrop, Wang seemed unremarkable. Tian Jiangchuan, managing partner at Monad Ventures, later wrote in her internal notes that Wang’s background was “grassroots”—a graduate of a ordinary 211 university, with just three months at DJI on his résumé. Around strangers, he often wore a stern expression that seemed tense.

When investors asked how he planned to make money, Wang gave an honest answer: delivering packages up staircases in old residential buildings without elevators. It was an answer that made investors shake their heads.

The investment world faced a moment of reckoning, grappling with the question of how Unitree went unnoticed. But in 2019, most simply “couldn’t pull the trigger.”

Tian later diagnosed her own failure as “elitist arrogance.” He once believed the robotics industry required top-tier academic pedigrees. Wang’s degree from Shanghai University, Tian thought, didn’t match the industry’s demands or his fund’s preferences. “That was a colossal misjudgment.”

Unitree was also pursuing a technical path that diverged from the mainstream. Around 2016, when the company was founded, most domestic robotics firms were copying Boston Dynamics’ hydraulic approach. “That era was Copy From America,” said one early investor.

Hydraulics rely on high-pressure oil pushing pistons in joints—similar to human muscle contraction. Zhou distinctly remembers asking Wang why he wasn’t pursuing hydraulics. Wang became almost irritated.

“Hydraulics will never work,” he said bluntly. “First, hydraulic cylinders are high-precision machinery. Americans make them small and thin; Chinese manufacturers can’t do that. Second, the costs can’t come down. Third, hydraulic control precision isn’t sufficient for fine movements.”

Compared with hydraulics, Wang’s chosen electric-motor approach couldn’t deliver the same power output, but it was simpler, cheaper and more precise.

Time proved him right. In April 2024, Boston Dynamics officially retired its hydraulic Atlas and pivoted to the electric-drive approach that Unitree had championed since 2016.

Wang’s thinking was never purely theoretical; from the start, it pointed toward cost and scale. He told Vertex Ventures China founder Zheng Juncong that Unitree’s R&D started from a position of low cost, high reliability and practicality. The goal was “to truly bring quadruped robots into public life.”

To achieve this, Wang built a system from scratch.

Shortly after Unitree implemented direct-drive technology, domestic competitors began adopting a comparable strategic approach. However, the majority adhered to a standardized configuration: Swiss Maxon high-speed motors coupled with LeaderDrive reducers. The principle was simple—high-speed motors are small and fast but weak. Add a harmonic reducer, and it acts like an amplifier, trading speed for torque.

Unitree took the opposite path: quasi-direct drive (QDD), using low-speed motors with planetary reducers. Low-speed motors are larger and slower but inherently more powerful, so they require less amplification. More importantly, harmonic reducer gears are as thin as eggshells—fragile and costing thousands of yuan each. But planetary reducers are standard industrial components costing mere tens of yuan and are virtually indestructible.

Beyond the power system, Wang overhauled structural designs that had been industry standards for years.

Take motor placement. Most competitors followed the Swiss ANYbotics scheme: two motors in the thigh, one in the knee. The problem? Too much mass in the lower limbs, draining power and complicating maintenance.

Unitree concentrated all motors at the hip joint, leaving the lower leg as a lightweight linkage. This architectural innovation compensated for the electric drive’s relative shortcomings in explosive power and torque.

“After looking at that structure, I thought: This is a genius design. It must have gone through countless iterations to arrive at this optimal solution,” Zhou said.

History ultimately vindicated Unitree as the technological frontrunner. In its wake, most domestic robotics firms adopted Unitree’s underlying blueprint as the industry standard: electric actuation, low-speed high-torque motors paired with planetary reducers, and a fully proximal three-motor configuration.

But being too far ahead then came at a price: Wang simply couldn’t find suppliers for his core components.

As one industry veteran recalled, the supply chain for quadruped robots was practically nonexistent back then. The manufacturing was relegated to small workshops, and the concept of “integrated joints” was unheard of. Motors, reducers, and controllers were all sourced from separate vendors.

Piecing them together resulted in bulky, heavy modules that left the robotic dog barely able to take two steps on its own—rendering it practically useless.

The market was flooded with high-speed motors, but Wang needed the exact opposite: low-speed, high-torque motors to drive joint movement. Furthermore, while the market was saturated with outer-rotor motors designed for drones and high-speed hair dryers, Wang specifically required inner-rotor motors. To make matters worse, his custom motor designs were half the weight of industry benchmarks. There were simply no mature, off-the-shelf products that fit the bill.

The motors he could buy were entirely unusable, and custom manufacturing was out of the question; in its early days, Unitree was only producing a few dozen units a year. “If you knocked on the doors of listed companies like Wolong or Moons’, they wouldn’t even give you the time of day,” a former employee recalled.

After facing repeated setbacks, Wang had a crucial realization: “A complete-system manufacturer must master its core technologies, just like Apple. We have to reshape the supply chain according to our own vision, rather than being held hostage by suppliers.”

Unable to source externally, Wang bought silicon steel sheets and copper wire and began winding the motors himself.

“I was absolutely floored. This was so DJI,” Zhou recalled. Knowing that Wang had previously worked at DJI and was close to founder Frank Wang (Wang Tao), Zhou immediately recognized the origin of his scrappy, hands-on approach.

Becoming “a little DJI” was precisely why Dexun decided to bet on Wang. “As Shenzhen-based investors, we knew exactly how DJI grew,” Zhou explained.

But not everyone saw the vision. In 2019, Unitree had not yet achieved commercial scale, so investors had to rely entirely on their judgment of the founder and the product.

One industry insider said that an investment firm, recommended by a hardware titan, conducted due diligence on Unitree but ultimately walked away. Spooked by the company’s precarious cash position, they feared the investment would turn into bad debt. Another institutional leader was equally skeptical, asking, “A DJI on the ground? Shouldn’t that have four wheels? Why does it need four legs?”

For a company this unconventional, the reasons to pass were endless.

The Niche Market

To win broader recognition, commercial revenue is the most universal proof. But Unitree’s commercialization path didn’t fully convince investors either.

The business seemed to have limited upside. For a long time, Unitree’s most reliable cash flow came from universities and research institutions, which bought robot dog demos for secondary development. Previously, these institutions had to buy Boston Dynamics’ Spot, with the cheapest version starting at $75,000. Unitree’s Aliengo was one-third cheaper. After meeting these customers at an international exhibition in 2019, Unitree made them its base.

Other commercial paths seemed more imaginative. Several quadruped robot companies pursued consumer-facing home companions, securing investment on that story. But the vision proved premature—industry consensus later judged it “at least five years too early.”

The academic market had a low ceiling but was grounded and hassle-free. Research institutions had long sales cycles but spent generously. Unitree initially priced its products between $20,000 and $30,000, then discovered customers were price-insensitive and raised prices to $45,000. Standard products could be rebranded as “EDU” versions with modified code and an added chip module, fetching tens of thousands more.

Best of all, these customers required little maintenance. Unitree sold standardized hardware; users did their own DIY. The company provided fully self-developed open interfaces so professors could modify algorithms at will. Wang even personally authored a textbook, Quadruped Robot Control Algorithms, for university use. This avoided the trap of B2B business: high unit prices but high operational costs, where “every project loses money.”

Wang understood his own boundaries. Early on, Unitree briefly considered expanding into power grid inspection but quickly retreated. They found that serving large clients like State Grid required scenario adaptation, with engineers stationed long-term, heavy delivery and heavy integration.

“If you have a few people per project and each project is only a few hundred thousand yuan, how many projects do you need to stack to grow revenue? For a company of a few hundred people, it’s simply impossible,” one Unitree insider analyzed.

After launching the A1 robot dog, shipments increased, and Wang pivoted decisively toward standardized products. One media outlet previously reported a telling detail: whenever the sales team promised clients customized development to win orders, Wang would “furiously curse them out.”

Even in education, Unitree focused only on universities and research institutes. Meanwhile, rival robotics firms chased the broader “global education” dream, covering everything from kindergarten to university. By offering curricula, textbooks, and on-campus operations, they presented a seemingly attractive business model with steeper entry barriers. But the real returns paled next to Unitree’s approach.

Wang’s strategy was pragmatic. In one interview, he described it as “encircling the cities from the countryside”—finding simple, acceptable markets outside the perimeter of giants like Boston Dynamics.

These niche markets enabled Unitree to establish its early commercial viability. They honed product capabilities, built a complete architecture from R&D to sales, and surfaced countless problems—enabling a major robot iteration every year and achieving mass production.

The underlying reason was perhaps more blunt: “Our company was relatively small from the start, perhaps intentionally, or perhaps we simply had no other choice,” Wang once admitted. In a way, Unitree’s early limitations steered the company onto a solitary, yet relatively smooth trajectory.Being a hardware company at its core, Unitree’s playbook was always crystal clear: keep pushing performance boundaries while aggressively cutting costs. “First make the hardware good and bring costs down, so people can see products being sold and feel secure,” one employee assessed.

In one interview, Wang said: “Many mobile robot products are sold by the kilogram—heavier means more expensive, because components are just material costs plus some mechanical processing. We sell by weight: smaller ones cheaper, larger ones more expensive. Not much different from pork.” As early as 2019, he promised investors: “In five years, robot dogs [and humanoids] can absolutely reach around 20,000 yuan.”

Wang’s cost control bordered on obsession. He would calculate how many grams of metal, copper and plastic went into a motor, then draw a fixed bottom line that everyone had to meet.

The approach evokes Elon Musk’s work at SpaceX—similarly deconstructing costs, bringing seemingly unattainable rockets down to earth, breaking them into spreadsheets of reducible parts, materials, and process costs.

The largest cost in a quadruped robot is its motors. Using MIT’s direct-drive approach with planetary gear joints, motors could account for over 50% of total cost.

A Unitree robot dog requires 12 motors. If outsourced, each motor would cost 3,000–5,000 yuan. Unitree’s self-developed motors sell on its website for 1,300 yuan. “If an integrator wants to control its destiny, it must self-develop motors and transmission systems,” one industry analyst noted. “Get the core motors and joints right, and everything else can be reduced. Costs can be controlled tightly.”

By then, Unitree had mastered a systematic approach to driving down costs. They achieved this by sharing internal components across their self-developed products; transitioning between different series required nothing more than swapping out exterior shells.

As the product lineup grew, development costs were effectively spread out. The results were striking: the 2020 A1 launched at under $10,000, and in 2021, the Go1 hit the education and research markets with a base price of just 16,000 RMB, a figure lower than what competitors were paying just to build their own units.

“Every time Unitree released a product, everyone was shocked: ‘We’re done. We’ve been revolutionized,’” one industry insider recalled. By combining standardized products with extreme cost control, Unitree turned a profit in its second year of commercialization, 2020.

Yet, in a venture capital ecosystem that prioritizes boundless growth potential, this achievement was viewed as underwhelming. Academia, by its nature, continued to be a constrained niche. Even after capturing 60% of the global research market, Unitree’s revenue was only a little over 1 billion yuan. That year, the robotics companies raising the most money were industrial robots, surgical robots, and warehouse robots. Robot dog companies didn’t even rank in the top ten.

Unitree still hadn’t become the “consensus” pick among professional investors.

But Unitree seized a small opportunity. In 2021, after its robot dogs first appeared on the Spring Festival Gala, performing alongside celebrities, Unitree landed several major B2B clients, including inspection services. Headcount reached 100, with annual revenue in the tens of millions.

A few months after that Gala appearance, Unitree closed a Series A round of tens of millions of dollars led by Shunwei Capital, lifting its valuation to the billion-yuan level.

The Gala put Unitree on the map in primary markets, but robot dogs were an old product—hardly a breakout. One investor recalled that before Unitree’s Series B, they introduced the company to nearly 20 institutions. Not one placed a bet.

For those years, Unitree lived small but beautiful, yet inevitably fell into the trap that ensnares many technology companies: holding patents, exclusive technology, and mass-production capability, but unable to cross the chasm from niche to mass market.

Making Profits

In October 2022, when Elon Musk unveiled Optimus at Tesla’s AI Day, he likely didn’t anticipate that he would catalyze a phenomenon across the Pacific. Unitree had been operating for six years, still tepid. It was self-sufficient, with orders and reputation, but lacked a story compelling enough to unlock valuation ceilings.

Beyond robot dogs, Wang seemed to have little ambition. He was no risk-chasing fanatic like Musk, but a pragmatist. In a media interview, Wang mentioned that as early as 2021, investors had asked whether he wanted to build humanoid robots. He declined, believing “humanoid robots had already been hot for many years and were in a trough before Musk entered. I hadn’t heard of anyone wanting to buy one. The circle was pessimistic.”

It wasn’t until Tesla unveiled Optimus and ChatGPT exploded globally, with AI intelligence emerging, that Wang formally greenlit humanoid robots. “If he didn’t choose this path, his investors would force him. His employees would force him,” one investor said.

Initially, Wang couldn’t be certain of the returns. He assigned just three full-time employees to humanoid R&D, with the robot dog team pitching in occasionally. The story has been told countless times since: this tiny team took just half a year to launch the first-generation humanoid H1 in August 2023.

H1 stood 1.8 meters tall and was priced at 650,000 yuan. However, a closer look at its teardown reveals that it was essentially a lightweight test product. H1’s joint modules and motors were directly reused from quadruped robots, with only parameter adjustments. No heavy resources were poured in. The positioning was academic: “Use the robot dog architecture for a humanoid skeleton—able to walk and support secondary development. Doesn’t need to be perfect.”

This product’s role was simply to stake a claim in the humanoid race. Unitree perhaps didn’t expect market sentiment to shift so rapidly.

As the tailwinds of Elon Musk’s Tesla Optimus and OpenAI’s GPT swept in from Silicon Valley, a wave of domestic humanoid robot startups emerged, including Agibot, Galbot and Booster Robotics.

The sector reached a tipping point for massive capital deployment in October 2023, when China’s Ministry of Industry and Information Technology released the Guiding Opinions on Humanoid Robot Innovation and Development. Public data shows that funding for humanoid robots that year surged to 5.47 billion RMB, a decade-high.

Many of these new embodied-AI founders boasted dazzling résumés. Even among established firms, Fourier Intelligence released a humanoid robot just a month before Unitree. Yet, Unitree held a distinct advantage that latecomers could not easily replicate: it was already profitable.

One investor said that in 2023, after surveying the entire humanoid robot landscape, Unitree was the only company making money. According to its prospectus, the company generated 393 million RMB in revenue in 2024, with a net profit of 95.47 million RMB.

“This wave of humanoid robot companies is just getting started, and the previous generation of robotics firms was bleeding cash,” the investor noted. “Where else could you find a company with such high gross margins and revenue?”

Capitalizing on the surging momentum in the humanoid robot sector, Unitree secured a 1 billion RMB funding round in 2024, achieving a valuation of approximately 3 billion RMB in one of the year's largest transactions.

Unitree was still Unitree. Its DNA hadn’t changed. The rush of investors had finally recognized it as the most scarce and premium target on the market. Its cost control was ferocious. In August 2024, Unitree shocked the industry again with the G1, a compact humanoid priced at just 99,000 RMB. Competitors’ manufacturing costs alone often exceeded 100,000 RMB, yet Unitree had undercut the industry’s floor price while maintaining healthy margins.

At this aggressive price point, Unitree’s hardware reliability was clearly ahead of its peers. Zheng noted that bringing the hardware cost down to the tens of thousands was akin to IBM creating the personal computer: only then could a robust software ecosystem and practical applications flourish on top of it.

A former Unitree employee said that the core of their product stability lay in motor performance. Take instantaneous power, for example: the stronger the performance, the longer a robot can sustain violent, high-stress movements without overheating, enabling extended operation in extreme conditions.

In one instance, Unitree employees watched a competitor’s robot compete without arms. They studied it for a long time, assuming it was an intentional design choice, only to later realize the arms had simply fallen off during a crash. In contrast, Unitree’s self-developed motors ensured that even when robots crashed during extreme scenario testing, the components remained securely attached.

Despite not being a household name, once the investment community reached a consensus on the humanoid robot sector, Unitree found itself at the center of fierce primary-market bidding wars.

In 2024, due to external factors, some early investors partially exited, briefly releasing old shares onto the secondary market. Those shares were pure gold. One insider said that when a single institution announced its partial exit, phones rang incessantly. Other investors were willing to pay a billion-yuan premium just to acquire those shares, and every colleague at that firm was harassed for days with the same question: “Do you have any old shares?” By year-end, Unitree’s valuation had skyrocketed to 8 billion RMB.

“2024 was the last year people had doubts about Unitree. Looking back, that was exactly the inflection point before takeoff, and some institutions exited below 5 billion RMB. You should have grabbed that bargain with everything you had!” the insider remarked.

A few months later, all doubt and hesitation evaporated. Unitree was no longer merely a startup; it was poised to become the brightest totem on China’s hard-tech map.

Household Name

Despite his outwardly rigid and serious demeanor, Wang knew exactly how to ride a wave of momentum. In an interview, he articulated this philosophy: “When what you’re doing aligns with the era’s main theme, good resources will naturally tilt toward you.”

When the 2025 Spring Festival Gala organizing committee invited several humanoid robot companies at the end of 2024, Wang was the only one to agree without hesitation. The others declined for various reasons.

One company refused primarily because it lacked a viable humanoid robot. Despite having announced one years earlier, mass production had not kept pace, causing the firm to miss the Gala window. Another company had a suitable robot and a fitting scenario, but the math simply didn’t work.

Robot companies receive no payment for appearing on the Gala; all costs are borne by the enterprise. This meant committing not only a large fleet of machines but also stationing R&D personnel in Beijing for three full months. A 1.8-meter humanoid robot required two people just to carry it on and off stage. In 2024, humanoid robots were still relatively crude, and their algorithms were not yet generalizable—meaning each robot required a dedicated engineer. At a minimum, the human resource investment exceeded 20 people.

More daunting was the timing: Gala rehearsals coincided with the year-end rush, precisely when robotics companies were scrambling to fulfill project deliveries. For an industry heavily dependent on meeting deadlines, pulling core personnel to coordinate an unpaid performance was essentially a high-stakes brand gamble.

According to 36Kr, one company internally estimated its total costs—including parts attrition, labor, and miscellaneous expenses—at 20 million RMB. Participating in the Gala meant spending one-tenth of its annual revenue on a five-minute show. “In a market that hadn’t fully taken off, ordinary people just don’t have that kind of nerve,” one insider noted.

While others questioned the Gala’s utility, Wang pushed all his chips to the center of the table.

Unitree had already capitalized on the massive exposure generated by its stage appearances. Following its 2021 Gala debut and its billion-yuan valuation milestone, the company participated in several major official events, including the 2022 Winter Olympics and the 2023 Asian Games. In 2024, twelve H1 units took the stage at Macau’s MGM Theater, performing yangko dances alongside humans. Directed by Zhang Yimou’s team, this performance served as a rehearsal for the later Gala Yang BOT program.

An insider revealed that the usual Wang was a boss who would message administration if he found a South Battery in a conference room remote, demanding a generic brand instead. Yet, when it came to Gala spending, he didn’t hesitate. “Some important decisions require business intuition and boldness,” said Yu from Shunwei. “Even if the ROI doesn’t compute on paper, if the matter is important enough and the long-term direction is correct, he will fully support the team in executing it flawlessly.”

Wang took the performance incredibly seriously. To ensure the Gala robots were supplied, other orders were deliberately deprioritized. He stayed on-site as his team rehearsed for days without sleep.

The result exceeded all expectations. Participants described the impact as “explosive.” On the night of the 2025 Gala broadcast, the founder of one company that had declined the invitation was still sending red envelopes in a group chat. After seeing Unitree’s performance, the red envelopes stopped. “The emotions were complex,” an employee recalled. “They didn’t expect Unitree could pull it off so well.”

Unitree’s five minutes of destiny in 2025 were born of both luck and a decade of accumulation. “Ten years of sharpening one sword—all that accumulation just happened to explode at this precise moment of opportunity, presenting itself fully before everyone’s eyes,” one investor marveled.

The performance struck the exact emotional chord of great-power tech competition. China needed to prove its technological strength, and Unitree arrived at precisely the right moment.

After the Gala, orders flooded in. Daily shipments, which previously hovered around a dozen units, surged to 40 or 50, peaking at over 100 robot dogs and humanoids shipped daily. Ever conscious of crises, Wang messaged the company-wide group chat: If Unitree can’t ship, competitors will steal our customers. Everyone must improve efficiency.

During that period, shipping and assembly headcount doubled. All hands were on deck, working from 8 a.m. to 10 p.m. for two straight weeks with only one day off. “I worked until I wanted to vomit,” one former employee recalled.

Unitree quickly became a company ordinary investors simply couldn’t afford. Its valuation climbed from 8 billion RMB pre-Gala to over 12 billion RMB by mid-2025, when it closed a 700-million-RMB Series C+ round. The investor list featured just six names: China Mobile, Tencent, Alibaba, Ant Group, Geely, and Jinqiu (backed by ByteDance).

It was a meticulously curated roster. One major automotive group reportedly failed to make the cut, while another investor only secured an allocation after their chairman personally called Wang.

Everyone understood implicitly: Unitree’s A-share IPO was imminent, and this deal was a sure bet.

The Blockbuster IPO

Unitree’s public listing represents far more than a single company’s milestone; it will set the tone for the entire embodied intelligence industry.

In the six months leading up to the IPO, China’s embodied intelligence sector experienced an unprecedented surge of activity. Total funding in the first half of 2026 approached 100 billion yuan, a fivefold increase within a single year.

Rarely in China’s history has a hardware industry attracted capital at such a remarkable pace. Embodied intelligence robots have become the third-largest destination for early-stage investment in 2026, trailing only large language models and AI chip supply chains.

Within half a year, the sector produced 23 robot companies valued in the tens of billions, with eight surpassing 20 billion yuan. “Last year, people were still counting the hundred-billion-yuan companies. This year, nobody bothers,” one robotics investor noted bluntly. He added that if an investor missed checking in on a top-tier company for just a month, its valuation would immediately surge or double. “People used to advocate investing small and early, but in 2026 China’s embodied intelligence, there is only early—no small.”

With robotics integrated into the national five-year plan, state-backed institutions entering the market, and listing channels opening, the entire sector has been propelled upward. Yet, on the eve of the public offering, investors face a critical question: What will these companies be worth in the public market? Will public investors pay a premium for these future aspirations?

For years, capital markets valued Chinese startups by benchmarking them against American peers. However, embodied intelligence is largely flourishing in China, and American benchmarks no longer apply. As the highest-revenue and highest-margin humanoid robot company in the country, Unitree serves as the true indicator of the industry’s vitality.

Investors hope Unitree’s listing will stabilize industry sentiment. Previously, optimism regarding humanoid robots swung between extremes—either robots would enter every household tomorrow, or their practical application remained a distant dream.

Unitree’s listing proves that the industry already possesses substantial market demand and can sustain at least one publicly traded company. Furthermore, Unitree establishes a pricing benchmark. Once other companies can replicate its production scale and commercial success, they can expect similar recognition from the capital markets.

Unitree’s initial public offering valuation of 60.9 billion yuan is substantial. Yet, most investors believe the company will likely reach 150 billion yuan post-listing, with peaks potentially hitting 300 billion. When asked for the basis of this judgment, one investor answered candidly: “A gut feeling.”

Regardless, early investors are poised to reap extraordinary returns. “This is a once-in-20-years opportunity,” marveled one investor, who secured a massive return from this single deal. He concluded that being able to invest in Unitree back then felt like a gift, allowing him to encounter a founder destined for greatness.

But after the initial wave of market euphoria subsides, what valuation can Unitree sustain?

Unitree is not without its vulnerabilities. Currently, its greatest pressure comes from developing the artificial intelligence “brain.” The company’s prospectus frankly states that early research and development focused heavily on hardware bodies and motion control, with relatively little investment in embodied large models or large-scale real-world data collection.

Of the money raised, much will be directed toward intelligent robot model development. Wang personally leads the team working on this technology. With the technology not yet fully converged, Wang and Unitree face a challenge reminiscent of the grueling battles of a decade ago.

Another pressing concern is whether a company with hardware roots can ultimately stand shoulder to shoulder with AI-native players, or if it will only be relegated to the lower end of the value chain—serving as a hardware supplier providing the physical platform for algorithm-savvy firms.

Amid abundant investment and optimistic prospects, Unitree now faces a dense forest of competitors. In 2026, China is home to 400 embodied intelligence companies, including well-funded newcomers, automakers, smartphone manufacturers, and hardware supply chain giants.

On the eve of the listing, sentiment inside Unitree is divided. Employee stock purchase thresholds were set high: a minimum subscription of 1 million yuan at the IPO price without a discount, accompanied by lock-up periods of one to three years. Despite these stringent conditions, nearly half of the research and development staff subscribed. Wang personally invested 15 million yuan, and three core technical leaders each subscribed 9 million yuan with a three-year lock-up.

Others watched coldly, citing concerns over the high threshold, lack of discount, or the fear that the industry might enter a downturn during the lock-up period.

In just eighteen months of mainstream fame, Unitree’s valuation has skyrocketed nearly five times over, propelling the company rapidly toward a public listing. Yet, in a recent interview, Wang acknowledged that this sudden fame has brought enormous pressure to both Unitree and the broader industry.

“Everyone expects Unitree’s technology to advance extremely rapidly, but in reality, foundational technology cannot be achieved overnight,” Wang observed. To cross the difficult technical threshold of achieving true adaptability, he estimates it will take at least another two to ten years.

In this sweltering summer, this quiet truth may soon be lost in the frenzy, waiting for the day when China's embodied intelligence industry finally faces its true test.

Share

Leave a comment

In October 1992, the 14th National Congress of the Communist Party of China set the nation on a bold new course: building a socialist market economy. Just weeks later, on the 19th, Zhu made a meteoric rise at the First Plenary Session of the 14th Central Committee. Leaping several ranks, he vaulted from an alternate member of the Central Committee to a Standing Committee member of the Politburo. He was promoted to being premier in 1998. In an era still heavily bound by seniority, Zhu’s ascension was nothing short of a shockwave.

As the old saying goes, extraordinary achievements require extraordinary people. Building a socialist market economy was uncharted territory, demanding a leader of iron resolve. Deng Xiaoping had spoken of him with profound gravity: “Our Party needs, now and in the future, a cadre of leaders with strong party spirit, economic expertise, open minds, innovative spirit, and a willingness to toil without complaint. Such comrades are rare today. Zhu is precisely one of those.”

The Empty Treasury and the March for Tax Reform

Fiscal management has been a perennial governance challenge throughout China's long history. Upon taking office, Zhu was immediately confronted with a classic dilemma: a brilliant cook staring into an empty pantry. Central government coffers had been bleeding red ink for years.

Then-finance minister Liu Zhongli often lamented the “empty treasury,” recalling a 1993 symposium in Hainan: “When Comrade Li Xiannian was Finance Minister, he had a jacket and trousers to wear. When Wang Bingqian took over, he still had a shirt. By the time it was my turn, I was left in just a tank top and shorts.” In 1992, central government revenue stood at a meager 97.9 billion yuan, accounting for a mere 28% of total national fiscal revenue. Beijing was repeatedly forced to borrow from local governments.

In April 1993, the Politburo Standing Committee tasked Zhu with spearheading several major reforms, chief among them the fiscal and tax system.

Three months later, at a national fiscal and taxation conference, Zhu issued a stark warning to the attendees: “If we don’t begin reforming the fiscal and tax system now, next year will be a disaster.”

The existing system was a “fiscal contracting” model, where local governments remitted a fixed portion of revenue to Beijing and kept the rest. Zhu’s vision was a “tax-sharing system” (分税制): revenues would be divided by tax category, clearly defining what belonged to the center, what to localities, and what was shared. For instance, consumption taxes would go 100% to the center, while the lion’s share of value-added taxes (VAT) would be split 75-25 in Beijing’s favor.

To push this through, Zhu led a caravan of over 60 officials on a grueling northward tour starting in Hainan and Guangdong in September 1993, engaging in massive persuasion, mobilization, and negotiation.

He drew a hard line: the reform would be implemented uniformly nationwide; no province would get a bespoke rate. Liu recalled the intense negotiations in Guangdong: “We negotiated in tiers. Zhu spoke directly with Guangdong Party Secretary Xie Fei, while I handled Governor Zhu Senlin and Vice Governor Lu Ruihua, who oversaw finance. Bureau heads of finance ministry and provincial finance chiefs crunched the numbers together, and national and local tax officials haggled at their level. We explained repeatedly: yes, under the new system, the center takes more, but the pie will grow so much that local wealth will surge anyway. It won’t stop Guangdong from chasing Asia’s ‘Four Little Dragons.’”

Calculations showed that from 1994 to 1997, the center took only 4% more from Guangdong than before, rising to just 5% by 2000. “Because Guangdong’s pie grew so much,” Zhu argued, “I don’t believe this will hinder its development; it will actually promote it.” After several clashes, Guangdong, seeing the bigger picture, agreed.

Resistance also flared in the country’s central and western regions. Provinces like Yunnan and Guizhou relied heavily on tobacco and alcohol taxes, but under the new plan, consumption taxes on these goods went entirely to Beijing. Some regions begged for exceptions: Could we keep 100% of the VAT? Could the center leave the consumption taxes alone?

In Xinjiang, Zhu was uncompromising: “The spirit of the tax-sharing system boils down to two points. First, the center must have a stable revenue stream. Second, central revenue must grow faster than local revenue. Otherwise, how can the central government bridge the wealth gap between regions? The comrades in Altay just mentioned building a reservoir on the Irtysh River, developing mines, and paving rural roads. None of that happens without national support. If the central government doesn’t collect more from affluent areas, where does that money come from? This system actually favors poorer regions. Xinjiang should be applauding this with both hands! The national plan is non-negotiable. No special treatment for anyone.”

That year, Zhu spent over 70 days traveling across 13 provinces. He later joked that it was a campaign of “marching south and north,” relying on “sweet talk and bitter persuasion,” sometimes swallowing his pride, other times wielding a mix of carrots and sticks.

“The reform was a massive success,” he reflected at a 2002 national fiscal conference. “It initially standardized the distribution of wealth among the state, enterprises, individuals, and between the central and local governments. It mobilized both sides and established a mechanism for stable fiscal growth.”

Over two decades later, even as national and local tax bureaus merged in 2018, the fundamental tax-sharing system remained intact, widely recognized globally as the most effective way to manage fiscal relations across multiple tiers of government.

“Three Years to Turn Around SOEs, or I’ll Step Down”

Today, Zhu Rongji is primarily remembered—both praised and criticized—for the state-owned enterprise (SOE) reforms he implemented.

A portion of fiscal revenue came from SOE profits, but by the mid-1990s, SOEs were dragging down the economy. Of 11,000 SOEs, 63% were operating at a loss.

In November 1993, the Party formally proposed establishing a “modern enterprise system” characterized by clear property rights, well-defined rights and responsibilities, separation of government from enterprise, and scientific management. “Clear property rights” topped the list, sparking fierce debate.

Starting in 1992, Chen Guang, mayor of Zhucheng (诸城) in Shandong province, had spearheaded a radical reform, converting 282 SOEs and collectives to a “shareholding cooperative” model where employees held stakes. The “Zhucheng Experience” drew both applause and brickbats; critics branded it a capitalist road.

In March 1996, Zhu visited Zhucheng. While praising the city’s exploratory spirit, he warned that a shareholding cooperative wasn’t a magic bullet. “It’s not true that ‘selling off solves everything’ or ‘joint-stock reform solve everything,’” he noted. He pointed to a local tire factory where the manager could buy 60,000 yuan in shares while workers were capped at 3,000 yuan—a 19-to-1 gap. With a 22% dividend payout, the manager earned over 10,000 yuan annually while workers got a mere 600.

“If this continues, it won’t be a ‘community of shared interests.’ Workers will feel they’re just toiling for the boss.”

He suggested suspending cash dividends, instead reinvesting both worker and state dividends back into the company to build reserves. He also advised using worker funds to pay down bank loans, thereby lowering the company’s debt ratio. Zhucheng subsequently adjusted its equity structure to further unleash corporate vitality.

But Zhucheng’s SOEs were mostly small. Large and medium-sized SOEs nationwide were still bleeding money. In March 1998, at a press conference during the National People’s Congress, Zhu declared: “We have set a goal to help most large and medium-sized loss-making SOEs turn around within three years.”

He diagnosed three fatal flaws: redundant construction, blind investment, and bloated workforces. Redundant construction cannibalized existing SOEs; blind investment was driven by officials chasing “political achievements,” building projects with zero capital and 100% bank loans that could never be repaid.

“From now on, we must decisively halt redundant and blind construction. The central government will not contribute a single cent in capital for general industrial projects.” He identified the most glaring issue: surplus personnel. “Layoffs are the key.”

China was also facing overcapacity and downward economic pressure. After years of double-digit growth, GDP slowed to 9.9% in 1997 and 9.2% in 1998. In 1998, the State Council set a 8% growth target.

Against this grim backdrop, Zhu launched sweeping reforms. Li Rongrong, former head of SASAC (The State-owned Assets Supervision and Administration Commission of the State Council), recalled: “Premier Zhu poured immense energy into this. He literally said, ‘If we don’t turn a profit in three years, I’ll resign.’”

To curb blind investment, massive numbers of enterprises were closed, merged, or restructured. In Liaoning, for example, every non-ferrous metal mine was ordered shut, except for a single copper mine.

The word xiagang (laid off) entered the national lexicon. During the 1999 Spring Festival Gala, comedian Huang Hong quipped in a sketch: “The boss favors me and was about to promote me to deputy team leader. But then he sat me down and said we need to downsize and merge factories. I immediately told him: ‘Workers must think of the country. If I don’t take the layoff, who will?’”

Within a few years, tens of millions of workers were laid off from SOEs and collectives. “We didn’t want to lay off workers,” Li Rongrong explained. “But to save the enterprises, layoffs were mandatory. Otherwise, we’d all go down together.”

He stressed that laid-off workers must be channeled into other industries and that basic living standards must be guaranteed. For those entering reemployment service centers, living allowances were split equally among the enterprise, society and government.

Amid the “layoff wave” came a “startup wave.” Chen Xiaoping, a laid-off woman in Fuzhou, founded a steamed bun company that hired other laid-off female workers. In January 1991, Zhu visited her shop and ordered a bun. CCTV later produced a drama based on her story, Endless Blossoms. Her company eventually grew into a major Fuzhou restaurant chain.

By November 2001, at a central economic work conference, Zhu acknowledged: “The three-year SOE turnaround goal has been basically achieved. But the bigger problem is that a modern enterprise system hasn’t been fully established.”

Shao Ning, former deputy head of SASAC, noted that Zhu’s SOE reforms “fundamentally reshaped the layout of the state-owned economy and resolved contradictions during the transition. Without those reforms, we wouldn’t have the relatively strong state-owned economy we have today.”

The SOE reforms Zhu championed were a painful yet inescapable necessity, a process that also left behind countless heartbreaking stories. It was precisely through this crucible of suffering that China emerged as the world’s second-largest economy.

“RMB Devaluation: The Harms Outweigh the Benefits”

In 1997, speculators like George Soros aggressively shorted the Thai baht, causing it to plummet 30% against the dollar, eventually falling 60%. The storm swept through the Philippines, Malaysia, Singapore, South Korea, and Indonesia, triggering massive currency devaluations, corporate bankruptcies, and toppled governments.

Foreign leaders, including Singapore’s Senior Minister Lee Kuan Yew, asked Zhu: Will the RMB devalue? Zhu’s response was resolute: “I have announced in the press that China’s RMB will absolutely not devalue.” This sent a clear signal to speculators.

Zhu explained that devaluation would do more harm than good. Exports might tick up slightly, but not enough to matter. “Chinese people fear currency devaluation and inflation most of all,” he said. “Devaluation would signal to the public that China’s economy is failing. It would also trigger further devaluations across Southeast Asia, especially devastating Hong Kong and causing financial chaos.”

In 1998, the Hong Kong dollar also came under attack. At the Two Sessions press conference, a Hong Kong reporter asked what Beijing would do if Hong Kong faced trouble. Zhu replied: “If the SAR government requests it, the central government will spare no expense to maintain Hong Kong’s prosperity and stability and protect its linked exchange rate system.”

“It is wrong to pin the entire Asian Financial Crisis on international speculators or call it a political conspiracy,” Zhu said at a 1997 central economic work conference. “The root cause is your own economic problems. Southeast Asian countries borrowed heavily in foreign currencies and dumped the money into real estate. When the property market collapsed, the capital died, banks went bankrupt, and the entire economy ground to a halt. Thailand and South Korea were the most obvious cases. Their prosperity was an illusion, riddled with bubbles that couldn’t withstand the test.”

He warned: “When doing economic work, you must pay extreme attention to economic structure. No matter what industry you develop, it must have a market, be profitable, and improve management. You can’t just spread out a massive canvas. If you can’t clean it up later and can’t repay the loans, the banks will collapse, and the entire economy will follow.”

He noted that Southeast Asian crises universally started in the banking sector. Commercial banks lacked sufficient capital and reserves, accumulated massive non-performing loans (NPLs) from bad investments, and central banks failed to issue early warnings. When the collapse came, it was catastrophic.

To build defenses, China established the China Cinda Asset Management Company on April 20, 1999, to handle $200 billion in NPLs from China Construction Bank. Soon after, Great Wall, Orient, and Huarong were created. Together, the “Big Four” absorbed $1.4 trillion in bad assets, significantly boosting the nation’s financial resilience.

To achieve the “three-year SOE turnaround” and cushion the blow of the Asian Financial Crisis, Zhu implemented a “debt-to-equity swap.” Massive corporate bank loans were converted into equity. Enterprises were relieved of crushing debt and interest payments, while banks saw their balance sheets clean up and NPL ratios drop dramatically.

Zhu also deployed an active fiscal policy, issuing bonds to concentrate resources on infrastructure: reinforcing dikes along major rivers and lakes, building railways and highways, upgrading rural power grids, improving sewage and waste treatment, and supporting key enterprise technological upgrades.

“These are long-term investments,” he said at a 2003 State Council meeting. “They yielded tremendous results, stimulating production, reviving the shrinking domestic market, and supporting the overall development of the national economy.”

Navigating the Minefield of Institutional Reform

SOE reform and government institutional restructuring were dubbed by the media as the two “minefields” Zhu had to cross. Upon being elected Premier in 1998, he told domestic and foreign reporters: “Whether it is a minefield or a bottomless abyss ahead, I will march forward without hesitation, without reservation, giving my all until my heart stops beating.”

In a 1997 conversation, Zhu revealed the sheer friction of reforming the State Council: “I’m currently orchestrating the restructuring. I’ve spoken to dozens of ministers one by one, and not a single one volunteered that their department should be abolished.”

The marathon meetings left him so exhausted that standing up was a struggle. But this didn’t deter him. During the 1998 NPC session, he declared: “I am approaching this with the resolve to be smashed to pieces.” On March 24, at a State Council plenary meeting, he announced: “The total headcount of the State Council will be cut in half. The decision is final.”

The number of Vice Premiers dropped from six to four; State Councilors fell from eight to five (two part-time); and Deputy Secretary-Generals were halved from ten to five. “Without cutting personnel, there is no efficiency. We keep meddling in things we shouldn’t, leading to endless buck-passing.”

Once, Zhu saw a TV report about a winery producing “Sanmei Wine (三梅酒).” After the winery was acquired by the Sanjiu Group (三九集团), the trademark should have been transferred. Instead, the local Party Secretary ordered the municipal finance bureau to buy the trademark, turning it into a government asset. “This is overreach!” Zhu fumed. “If our provincial secretaries, governors and mayors all interfere in the economy like this, how can China’s economy thrive? How can we build a socialist market economy?”

At a provincial and ministerial seminar on government reform in March 1998, Zhu explained the logic behind dismantling industrial ministries: “Why are we abolishing industrial departments and converting them into state bureaus under the State Economic and Trade Commission (SETC)? Facts have proven that managing enterprises through industrial ministries doesn’t work. In the past, we had eight ministries of machine-building, yet we still couldn’t manage the machine-building industry well... This government has decided that industrial departments will no longer manage enterprises. They will become state bureaus under the SETC.” Fifteen ministries, including Electric Power, Machine-Building, and Electronics, were abolished. State Council constituent departments shrank from 40 to 29.

The restructuring didn’t stop there. In April 2001, the State Council merged the State Bureau of Quality and Technical Supervision with the State Administration for Entry-Exit Inspection and Quarantine to form the General Administration of Quality Supervision, Inspection and Quarantine (AQSIQ). This fulfilled a 20-year dream for Zhu. “When I worked at the State Economic Commission, the SEC handled product quality while the foreign trade department handled export commodity inspection. The system was disjointed. I wanted to merge them back then. Now, finally, they are united. This aligns with international norms and facilitates work.”

At a symposium for AQSIQ officials, Zhu recounted a story: “I heard that in some cities and counties in Henan, they actually barred Shuanghui (双汇) Meatpacking Factory from setting up sales outlets. Shuanghi’s product quality meets international standards, and its large-scale production keeps costs low. Why block them? They should be able to sell nationwide. If outside enterprises want to sell in your city or county, you can collect their business tax. Why engage in local protectionism? Why protect backwardness?”

The State Administration for Industry and Commerce and the Press and Publication Administration were upgraded to ministerial-level administrations, joining AQSIQ as the three pillars of market regulation. “To run a market economy, the market mechanism must play a foundational role,” Zhu asserted. “This requires impartial law enforcement, supervision, and management. Without a fair, open, and just competitive environment, a market economy is impossible.”

By June 2002, after four and a half years of institutional reform, 1.15 million administrative positions were eliminated across party, government, and mass organizations at all levels.

In February 2018, an article titled “Twenty Years Ago, They Left the Ministry Buildings” went viral. It recalled the great redirection (分流) of 1998. Wang Xian, who had just joined a district discipline commission in Beijing, watched Zhu’s press conference live in his unit’s reception room, praising him as “a true man.”

Joining the WTO

In retrospect, if one were to single out the most pivotal event in China’s meteoric economic rise, joining the World Trade Organization (WTO) would undoubtedly top the list. At the heart of this historic milestone were the grueling negotiations between China and the United States, in which Zhu Rongji played a decisive role.

On November 9, 1999, U.S. Trade Representative Charlene Barshefsky led her delegation to Beijing. At 10:30 a.m. the following morning, Chinese representatives Shi Guangsheng and Long Yongtu met with their American counterparts to commence the final round of bilateral market-access negotiations.

By this stage, only the most intractable issues remained on the table: telecommunications, insurance, automobiles, and audiovisual products. On November 14, talks hit another impasse. That evening, after the American team walked away from the negotiating table, they simply vanished. Neither the Ministry of Foreign Affairs, through official channels to the U.S. Embassy, nor Minister Shi Guangsheng, who personally called the U.S. delegation’s hotel, could locate them.

No one expected that, even at this critical juncture, the U.S. side would resort to such theatrical brinkmanship.

Eventually, the Chinese side engaged the U.S. delegation through the embassy, firmly pointing out that their abrupt departure was entirely inappropriate. At 4:00 a.m. on November 15, the Americans finally returned to the table. Once talks resumed, the U.S. side proposed meticulously proofreading the more than 500 pages of agreements reached over the years, scrutinizing everything down to the punctuation.

Sensing the moment, Long Yongtu decided to make a crucial call. At 6:00 a.m. and again at 7:00 a.m., he phoned the office of Premier Zhu Rongji, hoping to relay a vital message to the top leadership: the U.S. genuinely wanted to sign.

At 8:00 a.m., a large-scale bilateral meeting began, with three to four dozen representatives from each side gathering for the final showdown. Shortly after 9:00 a.m., Premier Zhu returned Long’s call. After confirming the U.S. intent to sign, Zhu told him, “You have dealt with the Americans for years; I trust your judgment. Let me seek further instructions from the central leadership.”

Although Premier Zhu was originally attending a central economic work conference that day, he abruptly left the session alongside Vice Premier Qian Qichen and State Councilor Wu Yi to head directly to the negotiating frontlines for the final decision.

Upon meeting the U.S. representatives, Zhu made his displeasure unequivocally clear: “Disappearing suddenly and leaving without a word the next morning is incredibly discourteous! Frankly, I am still baffled. This feels less like a negotiation and more like a game of hide-and-seek. That is why I had to come see you today—it was absolutely necessary!... You neither confirm a deal nor declare a failure; you simply vanish. Is this politically serious? Therefore, I wanted to ask you face-to-face: what exactly are you trying to do?”

In the ensuing negotiations, Zhu personally took the helm, negotiating directly with Barshefsky. Ultimately, both sides laid their cards on the table. At 4:00 p.m. on November 15, 1999, China and the United States signed the bilateral agreement on China’s WTO accession. The most arduous leg of the journey to join the WTO was finally settled.

Two years later, China officially joined the organization, bringing an end to a grueling 15-year marathon of negotiations.

“Just Call Me an Honest Official”

In March 2000, when a Danish reporter asked Zhu what he most wanted the Chinese people to remember about him after his term, he replied: “I only hope that after I leave office, the people will say one thing: ‘He was an honest official, not a corrupt one.’ That would satisfy me. If they are generous enough to add, ‘Zhu actually got some real things done,’ I would be eternally grateful.”

Born in Changsha, Hunan, Zhu became a member of the Communist Party of China in 1949, the same year the PRC was established.

From 1991 to 2003, during his 12 years as Vice Premier and Premier, China underwent its monumental transition from a planned economy to a socialist market economy. His reforms in taxation, finance, SOEs, government institutions, social security and social governance continue to influence the nation today.

This piece has been adapted from articles published in Global People magazine.

Share

Leave a comment