^{A highly automated Midea factory}

The most revealing AI figures don’t necessarily come exclusively from AI companies. They are also starting to surface deep inside the earnings of manufacturers whose core businesses aren’t generally considered connected to artificial intelligence.

Chinese appliance giant Midea offers an unusually measurable example.

Its first-half results led with numbers analysts expected from a US$60-billion appliance maker: revenue rose 3.5 percent to 261 billion yuan (US$38.8billion) and net profit edged up 1.7 percent to 26.4 billion yuan. The company is benefiting from the AI infrastructure boom through its cooling and thermal-management businesses. That’s obvious AI connection.

Several paragraphs down, however, was a more telling number. Midea said its employees had built more than 20,000 AI agents across research and development, manufacturing, supply chain, sales and after-sales services, cumulatively saving more than 9 million work-hours and cutting costs by over 450 million yuan.

That figure, however, begs perspective. The claimed AI savings barely cover the increase in profit. If the estimate is accurate, that suggests AI may so far be doing more to offset other cost pressures than to generate excess returns.

That caveat matters, but so does the signal. The AI deployment by a manufacturer is now becoming quantified. It suggests the AI race needs two scoreboards: one for computing scale and one for deployment scale, which measures how extensively intelligence gets absorbed into factories, banks, warehouses and machines enough to lower costs, shorten cycles and raise asset utilization.

The US leads decisively on computing scale. China’s distinctive strategy is to expand computing power while pushing diffusion across a much broader industrial base.

Midea is unusually granular, but not alone.

Giant appliance maker Haier Smart Home also highlighted AI agents in its interim results, reporting 93 percent year-on-year growth in monthly active users of agent functions in its Smart Home app, with no attached cost figure.

Ping An insurance said average daily AI token consumption rose from roughly 30 billion at the end of 2025 to more than 120 billion by June, alongside faster underwriting and claims processing.

Midea estimates economic output, Haier measures adoption and Ping An measures usage. None proves that AI drove overall earnings, but together they suggest AI is moving into ordinary corporate operations beyond the tech sector.

The same is happening in the US.

GE Vernova told investors on an earnings call that it had 13 AI process transformations underway, planned to double that number to save tens of millions of dollars annually. That is forward guidance, not a realized figure, and deserves some modicum of skepticism.

US automotive supplier Lear’s broader IDEA transformation program, built partly on Palantir’s Foundry and AIP platforms, generated more than US$30 million in savings in the first half of 2025, though Lear does not attribute the full amount specifically to AI.

These disclosures expose a measurement problem. There is no standardized accounting definition of an “AI saving.” Companies use different baselines, mix AI with broader automation programs and rarely disclose whether implementation, software and computing costs have been deducted.

That makes cross-company comparisons difficult. But dismissing the numbers entirely would create the opposite problem: missing a trend simply because accounting standards haven’t yet caught up with it.

Computing is not deployment. The financial story of AI remains dominated by infrastructure spending. Amazon, Microsoft, Alphabet and Meta have issued guidance pointing to over US$700 billion or more in combined 2026 capital expenditure. Reuters calculated that Microsoft, Meta, Oracle, Amazon and Alphabet had disclosed about US$1.1 trillion in future AI leasing commitments, much of it related to data-center capacity.

The American AI boom can, therefore, look like one enormous infrastructure cycle: chips to servers, servers to data centers and data centers to power.

Those numbers tell us a great deal about how much capital is being committed to creating intelligence. They tell us less about how efficiently the rest of the economy is absorbing it.

China has its own computing race. Alibaba has committed more than 380 billion yuan over three years to cloud and AI infrastructure, while intelligent-computing capacity is expanding through both corporate investment and state-coordinated infrastructure.

But Beijing is also explicitly treating diffusion as part of its AI policy.

The 2026 Government Work Report calls for expanding intelligent-computing clusters while accelerating intelligent agents and commercial-scale AI applications across key industries.

That does not mean the government created Midea’s 20,000 agents. The effect is more indirect: build infrastructure, open operating systems and reduce barriers to adoption.

Nor is the distinction simply American computing versus China deployment. GE Vernova and Lear clearly show that American companies are deploying, too.

The more interesting question is whether China’s enormous industrial base gives it another route to AI scale. Once AI moves beyond data centers, the definition of AI infrastructure expands to

chips → computing → electricity → cooling → networks → sensors → robots → industrial equipment → factories.

A data center needs cooling systems, transformers and energy storage. An AI-enabled factory needs machine vision, industrial networks and robotics. Autonomous driving in mining still requires vehicles and communications equipment. AI logistics still requires warehouses.

China provides large chunks of that physical chain.

Midea itself captures the interaction unusually well. On the one hand, it sells cooling and thermal-management equipment into the data-center boom. On the other, it is deploying thousands of AI agents inside its existing operations.

AI therefore creates two potential dividends for industrial companies: new demand for the physical equipment needed to build an AI economy, and productivity gains from applying AI to businesses that already exist.

The second dividend is easier to miss because it does not necessarily appear as “AI revenue.” It may show up as lower inventory, fewer work-hours, faster product development or better utilization of an existing factory.

That’s why Midea’s 450-million-yuan figure matters even if it barely matches Midea’s profit increase.

The numbers are still too inconsistent and small compared with companies’ overall revenue to declare a winner, but they may tell us about where AI is now and how its economic impact will eventually be measured.

The first AI scoreboard was model performance, the second was capital expenditure and a third is emerging: how effectively companies convert intelligence into productivity.

For now, every company is keeping score by its own rules and quantitative figures for AI productivity remain elusive, but the trend of calculating this emerging AI facet is underway.

At one end is Sino-US BeOne Medicines. It earned 3.27 billion yuan (US$455 million) in the first half, with drug-product sales contributing 21.8 billion yuan, or 98 percent of revenue. Sino Biopharmaceutical presented a more blended model: Drug sales reached about 7.8 billion yuan, while overseas licensing payments added 980 million yuan, equivalent to 5 percent of revenue. At CSPC Pharmaceutical Group, the balance tilted further toward licensing deals, with payments of 5.9 billion yuan accounting for nearly a third of revenue.

Then there is RemeGen. The company turned to a 4.7-billion-yuan profit in first-half from a loss a year earlier, after recognizing about 4.5 billion yuan in technology-licensing income from its agreement with AbbVie. That income accounted for more than three-quarters of its revenue for the period.

The results capture a shift in China’s pharmaceutical industry.

For years, investors asked when it’s the nation’s drugmakers would become profitable. Now they are asking what lies behind the profit. Repeat prescriptions to patients? Licensing fee supported by a continuing pipeline of drug candidates? One large transaction that may not recur the following year?

None of the models is inherently better.

Selling drugs allows a company to retain more of the future value of its medicines, but requires years of spending on trials, manufacturing and commercial teams. Licensing brings cash forward and transfers much of the cost and risk to a partner, but can make one reporting period look far stronger than the underlying business. As China’s licensing boom reaches company accounts, the source and repeatability of earnings are becoming as important as the profit itself.

The clearest sales-led model is BeOne. The company, founded in 2010 as BeiGene and initially operating mostly in China, specializes in cancer drugs and has headquarters in Beijing and Cambridge, Massachusetts. The company is listed in Hong Kong and on the Nasdaq in New York. Its first-half product revenue rose to 3.3 billion yuan from 450 million yuan a year earlier. Its Brukinsa blood-cancer drug and main global product, generated 16.1 billion yuan – almost three-quarters of all product sales.

Hengrui Pharma, based in Jiangsu Province, is making a transition inside an older generic-drug business. Innovative-drug sales rose 16 percent in the first half to 8.8 billion yuan and contributed 63 percent of pharmaceutical sales. Generic-drug revenue fell 16 percent to 5 billion yuan, while licensing added 1.4 billion yuan. Total revenue slipped 1.9 percent and net profit was almost flat at 4.5 billion yuan.

Suzhou-based Innovent Biologics has added a new growth engine, and it is starting to look materially different from the oncology-focused biotech company of only a few years ago. First-half product revenue jumped 57 percent to 8.2 billion yuan, roughly matching its full-year level earnings in 2024. The increase came from its established cancer-drug portfolio and newer non-oncology products led by mazdutide, an obesity drug launched in China last year. Net profit rose 50 percent to 1.25 billion yuan.

These companies offer the most visible form of recurring revenue: medicines sold across successive quarters. But the figures also show reflect years of commercial spending, exposure to price cuts and, in BeOne’s case, Xiangcai Securities’ Jiang cited BeOne as a different model – one that retains global commercial rights and thus a larger share of the drug’s future sales. The company has borne the cost of overseas expansion.

Licensing produced a faster change at CSPC. First-half revenue rose 40 percent to 18.6 billion yuan and net profit attributable to shareholders climbed 139 percent. Licensing fees surged more than fivefold to 5.9 billion yuan, contributing 32 percent of revenue. Excluding those fees, finished-drug sales rose 10.8 percent to 10.2 billion yuan.

Sino Biopharmaceutical used deals as a smaller second growth engine. Its licensing income rose from 40 million yuan a year earlier to 980 million yuan, or 5 percent of revenue. The increase followed its March agreement granting Sanofi global rights to rovadicitinib, a treatment being developed for autoimmune diseases. Sanofi agreed to pay US$135 million upfront and as much as US$1.5 billion when milestone payments are included.

The effect was most pronounced at RemeGen. First-half revenue rose more than fivefold to 5.9 billion yuan, while net profit attributable to shareholders reached 4.7 billion yuan, swinging from a 450-million-yuan loss a year earlier. Technology-licensing income contributed about 4.5 billion yuan, or 76 percent of revenue, while product sales added a further 1.4 billion yuan.

AbbVie acquired rights to RemeGen’s RC148 outside China in January, paying US$650 million upfront and offering as much as US$4.95 billion in milestone payments. RemeGen retained responsibility for the drug’s development and commercialization in China, while AbbVie took responsibility elsewhere.

Upfront payments are fixed. Downstream potential payments are not. They depend on a drug meeting clinical, regulatory and sales targets over years, if it reaches them at all.

InnoCare Pharma shows how product sales and licensing can work side by side. First-half drug sales rose 43 percent to 918 million yuan, accounting for 81 percent of its 1.1 billion yuan in revenue. Collaboration income added 213 million yuan, mainly from milestone payments under its agreement with Zenas BioPharma.

The company earned 240 million yuan in the first six months, compared with a 36-million-yuan loss a year earlier, while operating cash flow turned positive at 81 million yuan. Sales growth was driven by orelabrutinib, including broader reimbursement coverage for its first-line CLL/SLL indication, and the launches of tafasitamab and zurletrectinib. The company said it expects further product growth and additional milestone income to keep it profitable for the full year.

That combination may become more common. Chinese developers signed 81 overseas licensing agreements with a potential value of US$110 billion in the first half, according to the National Medical Products Administration. As the payments enter financial statements, companies will need to show what remains when a deal-heavy period ends: product sales, further milestones or another drug ready for licensing.

A drug can lift earnings once when its rights are sold. A durable business must find a way to be paid again.

European giants are caught in a profit squeeze driven by cutthroat price wars in China and a slowdown in home-market demand for electric vehicles. Japanese automakers have seen a massive surge in paper profits riding the tailwinds of a depreciating yen and robust demand for hybrid vehicles in the US.

Yet, beneath top-line figures, a common thread emerges: operations in China are undergoing severe growing pains. The Chinese auto market, largest in the world, is no longer the “cash cow” it once was, evolving rather into an ultimate testing ground of resilience in face of changing consumer trends and stiff domestic competition.

Mid-year reports from European automakers show a deep chill. Volkswagen Group, for example, said first-half deliveries fell 6 percent to 4.1 million units. Operating profit dropped 12 percent in the second half from the prior quarter to 5.9 billion euros.

The profit contribution from VW joint ventures in China plummeted from 506 million euros in the first half of 2025 to just 184 million euros in the same period this year.

Luxury brands are facing the same situation. In its mid-year report, Mercedes-Benz took an 844 million euro impairment charge on its Chinese joint venture assets.

The core driver behind the poorer performance by European automakers is the extremely fierce competition from Chinese brands in mainstream segments, placing foreign brands in a position where even drastic price cuts cannot protect their market share.

At the meantime, growth in demand for pure electric vehicles in Europe has stagnated, leading to underutilization of capacity at local European factories, where high fixed costs are further erode profit margins.

In contrast to their European peers, Japanese giants posted eye-catching results for the first half of 2026. Toyota Motor, for instance, reported its net profit leapt to 1.48 trillion yen (US$9.4 billion), far exceeding the 841 billion yen recorded in the same period last year. Sales also grew by 10 percent to 13.5 trillion yen.

Amid the current political tensions in Sino-Japanese relations, FAW Toyota’s Corolla Cross was the only remaining fuel-powered model. Yet its sales for the first half of year still dropped year-on-year in China.

However, the engine behind this growth was primarily the depreciation in the yen, which makes exports cheaper. As the exchange rate shifted from around 145 to near 160 yen against the US dollar, Toyota’s massive overseas profits translated into a book-value surge.

Furthermore, Toyota’s strategic bets on hybrid models in the US successfully captured consumers retreating from pure electric vehicles. But if the currency windfall and the strong North American market are excluded, Toyota’s operating profit in the Asian region, which is centered on China, actually declined by 11.2 billion yen.

Honda and Nissan are similarly grappling with downward sales trajectories in China. Nissan barely managed to maintain a positive net profit of 3.8 billion yen through extreme cost-cutting measures, while its sales in China still fell by 15 percent year-over-year.

American automakers present a slightly different picture, characterized by traditional combustion engine vehicles providing a financial floor while new energy businesses continue to bleed cash.

Ford’s electric-vehicle division continued to endure heavy losses in the first half of 2026. General Motors achieved slight overall profit growth supported by the resilience of high-margin pickups and large SUVs in its home market, but its joint venture earnings in China saw a sharp year-on-year decline.

Whether it is the profit slide of European giants or the shrinking market share of Japanese and American brands in China, one reality is clear: the era when multinational automakers could reap excess profits in the Chinese market solely through brand premium and traditional combustion engine technology is over.

Western and Japanese automakers are now rethinking their “one-size-fits-all” global strategies. Instead, they are implementing a “dual-track” survival strategy, with localized production in China and a separate defensive playbook for the rest of the world.

For their China-related operations, multinational giants are shifting from simply “manufacturing in China” to “deeply integrating into the Chinese ecosystem.” Companies like Volkswagen and Toyota are handing over the research and development reins for new energy and smart-connected vehicles to local Chinese teams. By tapping into China’s mature supply chains, they are slashing new car development cycles from 48 months to under 24 months, directly cutting development costs by more than 30 percent.

Concurrently, partnerships such as Volkswagen with XPeng, Stellantis with Leapmotor, and Audi with SAIC, have fully materialized in 2026. European giants are utilizing reverse technology licensing to acquire China’s smart cockpit and electronic architectures. They are also using these joint ventures as export springboards, shipping highly cost-effective electric vehicles to Southeast Asia, Latin America and even back to Europe to circumvent trade protectionism.

Meanwhile, on the global front outside of China, these automakers are aggressively streamlining operations and fortifying their foundational businesses. Mercedes, Volvo and Ford, for example, have recalibrated their previously aggressive targets for pure electric vehicles, reinvesting heavily in plug-in hybrids, extended-range electric vehicles and highly efficient internal combustion engines to align with the demands of consumers in Europe and the US.

SAIC-Audi is expected to become a core base of electrification transformation of the German giant in China.

To hedge against the risk of slowing global auto growth, Volkswagen has even begun evaluating the closure of factories in its home base Germany, while major suppliers like Bosch and ZF plan to cut tens of thousands of jobs across Europe.

Furthermore, multinational giants are redirecting capital expenditure toward the US, India and the Middle East. They are leveraging domestic subsidies, such as the US Inflation Reduction Act, to consolidate their highly profitable pickup and SUV bases, while laying groundwork in emerging markets to dilute their reliance on any single region.

The 2026 mid-year reports signal a profound reshaping of the global automotive industrial order. The future success and survival of the once-untouchable multinational giants now depend on their ability to balance a two-front war, which is borrowing the speed and supply chain of the Chinese market to drive down costs, while relying on hybrid technology, traditional internal combustion engines and currency advantages to maintain cash flow overseas.

The tech revolution can’t prevent violent storms from hitting Shanghai and other coastal Chinese cities, but amid mounting economic and human costs, it is supplying AI-driven tools to predict, prepare for and survive extreme weather.

This comes as the UK predicts an El Niño event developing over the Pacific Ocean is likely to be the strongest in living memory, poised to disrupt weather patterns globally and heightening the incidence and severity of typhoons in East Asia.

In the latest tropical storm to hit China, Typhoon Maysak struck the southern Guangxi Zhuang Autonomous Region, leaving 159 dead and 10 missing across the cities of Nanning and Guigang until Friday. The typhoon affected 1.65 million residents.

Super Typhoon Dolphin struck Shanghai on August 9, the 13th typhoon to make landfall on the Chinese mainland this year. It forced the city to evacuate hundreds of thousands of residents living in high-risk zones and to deploy emergency crews to flooded roads and affected Metro networks.

In July, major typhoons Maysak, Bavi and Noul hit the mainland, causing widespread destruction.

Artificial intelligence is rapidly complementing traditional forecasting methods. By analyzing massive historical datasets, large AI models extrapolate weather patterns to predict a typhoon’s track, intensity and landfall probability up to seven to 10 days in advance.

Traditional physical models rely on supercomputers to solve complex fluid dynamics equations, taking hours to run a global forecast. AI models deliver results within minutes, or even seconds and demonstrate superior accuracy over traditional numerical models when predicting typhoon paths three to five days out.

However, forecasting sudden shifts in typhoon intensity remains a global challenge.

“AI models possess strong forecasting capabilities, but currently face bottlenecks like a lack of extreme weather data and insufficient model reliability,” said Pan Jinjun, chief engineer at the China Meteorological Administration.

China’s national blueprint called Scientific and Technological Innovation in Typhoon Research aims to integrate AI with physical models to reduce 24-hour track forecasting errors by five kilometers and overall location and intensity errors by 20 percent by 2030.

During Typhoon Dolphin, Shanghai’s AI-powered early warning system broadcast alerts simultaneously to all 6,671 residential community committees. High-precision 1-square-kilometer grid forecasts were sent directly to transport authorities to assess wind probable impact on specific subway lines, while public alerts were sent to 13 major platforms, including Amap, Meituan and Hello Bike, to notify couriers, drivers and commuters in real time, according to the Shanghai Meteorological Bureau.

The domestic pioneers in meteorological technology include university students providing some of the computing muscle behind these models.

At the recent Priority Research Application contest of intelligent computing at Tongji University, organizers showcased China’s first domestically developed 1,000-card engineering-grade computing cluster. The system supports Chinese AI models, including Alibaba’s Qwen and Huawei’s Pangu models, to run on hardware built by Shanghai-listed Hygon.

“Meteorological research falls under the category of ‘AI for Science’ and has long been a prime example of the convergence between high-performance computing and artificial intelligence,” said Ying Zhiwei, vice president of Hygon.

The newest meteorological model was a major feature at the contest. It showed how open-source AI weather models can be linked to public weather services and disaster mitigation, covering a seven-day mid-range weather forecast.

Forecast results can be delivered within seconds, assisting meteorological departments in rolling updates to forecasts. The system also covers extreme weather warnings, to simulate typhoon paths and predict intensities, heavy precipitation and severe convection. It also assists in the assessment of secondary disasters, such as landfalls, flash floods and urban flooding, providing knowledge needed in decision-making related to public services and the necessity for evacuations.

Unlike Western setups where central processing and graphics processing units operate separately, Hygon’s architecture enables its deep computing units and central processing units to work in concert at the hardware, operating system and compiler levels to maximize processing power, said Ying.

Mitigating typhoon damage also requires reimagining urban infrastructure.

Build environment consultancy Arup and Imperial College London in August unveiled a structural design concept that uses a tall building’s own weight to suppress wind sway and earthquake movement.

Traditional skyscrapers are built with massive, space-consuming dampers suspended near the top. The new design decouples a block of usable upper floors from the building’s central core, connecting them via springs and dampers. The design reduces wind-induced acceleration by up to 70 percent compared to conventional towers, improving structural resilience while cutting material usage and construction costs.

Arup plans to advance the structural design to large-scale trials and real-world project demonstrations. They may cover high-density cities such as Hong Kong, Manila, Taipei, Xiamen and Shantou -- no strangers to the dual risks of typhoons and earthquakes.

The company previously collaborated on Shanghai Urban Drainage Master plan and engineered the 370-meter Xujiahui Center tower, the current highest building on Shanghai’s western shores.

China’s Ministry of Industry and Information Technology released standards for L2 combined driving assistance and L3/L4 autonomous driving last week. The former takes effect in January; the later next July.

Self-driving technology is commonly expressed in six categories, from Level 0 of no automation, to partial automation Level 2, conditional automation in Level 3, high automation in Level 4 and full automation in Level 5. As the levels rise, the requirements for driver override controls decrease.

Together, the new rules establish a complete, mandatory regulatory system, marking a turning point for China’s smart driving industry as it shifts away from a marketing-driven era into a new phase of compliance and high-quality growth.

A major change lies in who bears the liability for mishaps. In the past, whenever an accident occurred, automakers often exempted themselves from blame by arguing that an autonomous feature was merely an assistant and the driver failed to pay attention. Under the new national standards, automakers will bear some legal responsibility in L3 systems, which require driver override, and L4 modes, where driver intervention is an option.

For consumers and car owners, the implementation of these new standards is also a wake-up call to reset their understanding of smart driving. The market was previously flooded with vague promotions. Many motorists put too much trust in assisted driving, thinking they could take their hands off the wheel or even nap on highways, leading to tragic results.

In October 2025, for example, a Changan Qiyuan A07 crashed on an expressway in Jiangxi Province while operating in smart driving assist mode. The driver had taken hands off the steering wheel for about 13 seconds, and the car failed to detect a semi-trailer stalled in the travel lane due to a breakdown, resulting in a rear-end collision that killed a family of three. The official investigation report noted that the driver’s over-reliance on smart driving and distracted driving were the primary causes of the crash.

^{A car crashes into a semi-trailer after the autonomous driving system failed to detect it.}

Meeting the new requirements will demand a total overhaul across research and development workflows, cost structures and marketing operations for Chinese automakers. In recent years, automakers grew accustomed to applying fast-paced Internet product updates to cars to sell vehicles by pitching high hardware specifications and then adding features month by month via over-the-air software updates.

When it comes to hardware and development costs, the new national standards set strict baselines for minimum risk maneuvers. For example, in the past, some driver-assisted systems determined if a driver was attentive simply by sensing small pulls on the steering wheel. Many reckless drivers would hang a physical weight of a few hundred grams on one side of the steering wheel to exert continuous pull, tricking the computer into believing a human hand was holding the wheel and stopping its warning alarms.

The new L2 standard mandates that steering wheels use touch-sensitive materials that detect human skin, paired with in-cabin perception systems to monitor driver behavior. For higher-level autonomous driving, the car must be capable of automatically pulling over to the roadside or slowing down safely on its own whenever a severe system fault occurs.

The rollout of China’s new standards also clears a regulatory path for Chinese smart-driving technology and automakers expanding overseas. Chinese smart driving software and hardware solutions are accelerating penetration into markets across the Middle East, Southeast Asia and parts of Europe.

According to Gasgoo Auto Research Institute, the installation rate of L2 and higher-category systems on new cars sold by Chinese automakers, such as Chery, BYD and SAIC, in the European market has surpassed 70 percent.

^{Unitree released “superrobot” ahead of its debut on 19th.}

China’s hottest IPO of the year may also be its clearest test yet of how much investors are willing to pay for the potential of humanoid robots.

Unitree Technology, which initially grabbed global attention for humanoids dancing and performing martial arts at during China’s TV Spring Festival Gala for two years, is scheduled to begin trading on Shanghai’s STAR Market tomorrow after a heavily oversubscribed IPO that raised 6.1 billion yuan (US$915 million).

The Hangzhou robotics company priced its shares at 150.80 yuan, giving it a post-IPO valuation of roughly US$9 billion.

The first indication of the unusual frenzy came in a subscription lottery for shares. Unitree’s final online allotment rate was just 0.0181 percent, roughly one allocation for every 5,525 applications. Nearly 9.8 million retail accounts participated. By comparison, this year’s other Chinese IPO star CXMT’s lottery had an online allotment rate 26 times higher.

Analysts estimate each allocation of 500 share-lots could generate at least 200,000 yuan in first day paper profits.

The scarcity of shares is being matched by an unusually aggressive valuation, with a price-to-earnings (PE) ratio of over 219 times. That implies a market capitalization roughly 45 percent above the valuation indicated by the company’s original fundraising plan.

Unitree generated about 1.7 billion yuan of revenue in 2025, more than quadruple the previous year, with humanoid robots contributing about half, overtaking its four-leg robot business. Unlike most humanoid startups, the company is already profitable, but first-quarter profit fell 53 percent as research and development and marketing expenses surged on growing competition.

Investors are paying not only for growth, but also scarcity. Unitree will be the first publicly traded humanoid robot manufacturer on the Chinese mainland, giving investors direct exposure to one of China’s emerging industries rivals that of the US. Unitree’s domestic competitor UBTech Robotics was listed in Hong Kong in December 2023. In the US, electric carmaker Tesla is also moving into robotics, and Agility Robotics is preparing a merger to become listed at a reported US$2.5 billion valuation.

One intriguing detail of the offering is the presence of DeepSeek founder Liang Wenfeng on both sides of the institutional allocation structure. DeepSeek received 933,399 shares, worth about 141 million yuan, through Unitree’s strategic placement, with a 36-month lockup. Separately, Liang-linked investment funds High-Flyer Quant and Zhejiang Jiuzhang Asset Management emerged as two of the largest private-fund recipients in the offline placement.

The broader strategic-investor roster reads almost like a snapshot of China’s technology-industrial establishment. It includes the National Council for Social Security Fund, DeepSeek, Tencent-linked capital, PetroChina’s Kunlun Capital, China Southern Power Grid, China Telecom and Citic Securities-related capital, among others. Strategic placements accounted for one fifth of the offering.

AI startup DeepSeek’s investment is particularly notable when reading alongside Unitree’s spending plan. Under its original 4.2-billion-yuan investment plan, the company’s single largest proposed use of IPO proceeds -- over 2 billion yuan -- is not another robot factory or a new generation of humanoid hardware. It is intelligent-robot model research -- in other words, the AI brain.

That spending points to what may ultimately determine whether today’s humanoid boom becomes a real industry.

China has become extremely good at building a robotic body from motors, actuators, and batteries to sensors, joints and increasingly sophisticated mechanical systems. But producing robots capable of running, dancing or performing complex tasks is different from producing machines that can autonomously understand unfamiliar environments, reason about physical objects and reliably perform easy work.

That requires more of AI brain than hardware.

Unitree is therefore investing heavily in embodied-AI models, including world model action and vision language action approaches designed to connect perception, reasoning and physical movement.

In that sense, the biggest question behind its IPO is no longer whether China can manufacture sophisticated humanoid hardware, but whether AI can advance quickly enough to make that hardware generally useful, industrially deployable and commercially economical.

The other risk Unitree is facing is Washington.

The IPO arrives just weeks after the US opened a new front in the tech race with China. On July 28, the Trump administration announced restrictions on imports of new models of certain foreign-made humanoid and quadruped robots.

Existing authorized Unitree models are not automatically banned, but future products could face substantial barriers. This matters because Unitree is unusually international for a young Chinese hardware company.

More than 40 percent of its revenue comes from overseas markets, while the US alone accounted for 13 percent of 2025 revenue.

The new US action isn’t a blanket ban on Chinese robot parts, but a measured one intended to reduce future US dependence on Chinese robotic supply chains.

All in all, Unitree’s market debut presents an unusual combination of narratives.

It is a rare profitable humanoid manufacturer entering an industry that barely exists at commercial scale; it is a company with a high PE that may not deter investors who have few opportunities to get a piece of the industry; it is a hardware champion spending its largest single IPO investment on AI software; and it is an increasingly global Chinese technology company confronting a US market that may become progressively harder to access.

The 1-in-5525 lottery success rate captures the excitement, while the price-to-earnings ratio captures the expectations. Unitree’s IPO is ultimately not a bet on how many robots it can sell today, but on whether AI can finally give those robots a brain capable of turning impressive machines into a mass industry.

Disclaimer: The writer holds shares in Unitree.

The fund is structured as an eight-year equity partnership rather than a traditional subsidy program. Days later, state-backed China Reform Holdings President Xu Siwei published an op-ed arguing that long-duration, risk tolerant capital, rather than short-term subsidies or speculative venture funding, would be China’s competitive advantage in the years ahead.

By late July, the cities of Shenzhen and Wuhan also announced their own multi-billion-yuan industrial funds built on similar structures, while Shanghai has long had its own dedicated semiconductor funds, including the Shanghai Integrated Circuit Industry Investment Fund III, which has increased in value this year to 6 billion yuan.

Taken together, these developments suggest more than a series of isolated announcements. They point to an industrial policy increasingly being formalized at the national level.

The debate over China’s industrial rise often returns to the same assumption: Beijing is simply spending more than everyone else. More subsidies, more distortion, more state money chasing market share. That explanation captures part of the story, but it increasingly misses a more important institutional shift.

The evidence points not primarily to the amount of state involvement, but to the design of the architecture of support.

Rather than relying solely on larger subsidies, China is increasingly structuring the state as an equity investor. That’s changing not only how much public capital is deployed, but also how it is deployed.

The China Life fund is the freshest example, but it is also perhaps the least novel part of the story. The underlying model —the government-backed equity investment versus a straightforward subsidy dispenser — has been evolving for years at the municipal level.

China Life’s June announcement is better understood not as the beginning of a new approach, but as the national scaling up of one that has already been tested locally. Three cases make that architecture visible.

When ChangXin Memory Technologies (CXMT) went public on Shanghai’s STAR Market two weeks ago, most coverage focused on the US$8.6 billion IPO and what it meant for rivals Samsung and SK Hynix. The more consequential story lies upstream: how the Hefei-based company was financed long before it reached public listing.

The Hefei city’s government investment vehicle began with just 10 million yuan back in 2016, when CXMT was merely a fledgling dynamic random-access memory chip startup with no products, no customers and little investor expectation that it could someday challenge industry giants like Samsung, SK Hynix or Micron.

Hefei didn’t support CXMT’s development through conventional grants, tax reductions or infrastructure incentives. Instead, its state-owned investment vehicle took equity stakes, assuming the same downside risk as private investors while sharing in any future upside.

For a chip startup, manufacturer progress requires years of capital infusion before meaningful revenue appears. Government, as an investor, could afford to be patient because its involvement looked to long-term value rather than annual budget cycles.

Hefei’s bet wasn’t just on CXMT. It was on building a whole ecosystem around it.

The contrast with Foxconn’s Wisconsin project is instructive. The US state of Wisconsin pledged more than US$4 billion in state and local incentives for a promised LCD manufacturing complex that was repeatedly downsized, repurposed and ultimately delivered only a fraction of the jobs originally projected. There was no ownership stake to appreciate if the investment succeeded nor one to absorb losses if it failed.

Hefei is not an isolated experiment. It has increasingly become a template. The Tianjin-registered investment vehicle behind China Life’s June commitment is mandated to invest in semiconductor companies with demonstrated technological capabilities and established research and development systems.

The language surrounding the fund is notable because it implicitly acknowledges the limitations of earlier industrial-policy approaches. Rather than emphasizing broad financial support, it prioritizes concentrated, long-duration equity investment in companies that have already demonstrated technical capability.

If Hefei demonstrated the model at the company level, Shanghai is applying it across an entire innovation ecosystem. The city’s state-owned assets regulator oversees one of China’s largest government-backed guidance funds, organized as a fund-of-funds rather than a direct subsidy vehicle. Instead of selecting individual companies itself, the fund allocates capital to professional investment managers focused on sectors including artificial intelligence, biotechnology and semiconductors.

The scale is substantial. Shanghai’s equity investment funds raised roughly 250 billion yuan in 2025, with approximately 135 billion yuan directed toward technology sectors. City officials have described this approach as “patient capital,” a concept that has repeatedly appeared in discussions surrounding China’s 15th Five-Year Plan. Shanghai has also sought co-investment from Middle Eastern sovereign wealth funds in many of the same strategic industries.

That co-investment is significant. Sovereign wealth funds typically seek governance structures capable of generating long-term financial returns rather than mechanisms designed primarily to distribute public subsidies. Their participation does not validate Shanghai’s model on its own, but it suggests these vehicles are increasingly being viewed as investment platforms rather than conventional subsidy programs.

The equity model explains how capital is supplied. An equally important, though less discussed, mechanism operates on the demand side: government acts as an early customer, absorbing commercial risks that private markets are often reluctant to bear.

Shanghai’s medical robotics sector illustrates the approach. Municipal policy documents emphasize not direct cash subsidies but rather accelerated regulatory approval, government procurement of innovative products and inclusion in medical insurance reimbursement schedules.

For a domestic surgical robot, commercial viability depends less on receiving a subsidy than on gaining access to public hospitals and reimbursement systems. The state is reducing early-market risk through demand creation rather than direct financial transfers.

Zhangjiang Science City demonstrates the scale of this ecosystem. The district now generates more than 1 trillion yuan in annual revenue across its pillar industries and accounts for over half of the Yangtze River Delta’s biomedical research talent. Multinational companies, including Roche and AstraZeneca, continue expanding manufacturing and research operations there, suggesting that global firms increasingly view these ecosystems not simply as beneficiaries of state support but as commercially attractive innovation clusters.

The prevailing narrative that China’s industrial success is primarily the result of larger subsidies is appealing because it is straightforward, but the reality is more complex. Many countries, including US, Germany, South Korea and Japan, provide some sort of support for specific industries, with varying results. And it’s interesting to note that the Trump administration since taking office has invested at least US$27 billion in 37 private companies, breaking with the long-held tradition of government support through loans, grants and tax incentives to take equity stakes in companies in AI, quantum computing and rare earths, among other industries it considers strategic.

If the crux of China’s support for its industries were simply spending, then the obvious competitive response would be to spend more. But if the more significant shift lies in institutional design, such as using equity investment to align incentives and procurement policy to reduce commercial risk, then simply matching China’s subsidy levels may address the wrong problem.

None of this suggests that China’s model is cost-free. State-directed capital allocation carries well-documented risks, including political favoritism, inefficient investment, and the misallocation of resources. Indeed, the current emphasis on “patient” capital follows years of uneven outcomes from earlier industrial policies, including failed semiconductor investments and corruption scandals surrounding some state-backed funds.

Even so, the cases of CXMT, Shanghai’s guidance funds and Zhangjiang’s procurement model point to an important institutional evolution. Increasingly, the Chinese state is acting not only as a regulator or subsidy provider, but also as a long-term shareholder, a limited partner and an early customer.

Whether that model ultimately proves more effective remains an open question. But it is likely to become increasingly important as governments around the world rethink industrial policy.

_{Commencement for PKU Mathematics in 2011 — Wang Hong is first on the right in the first row.}

When Wang Hong and Deng Yu became the first Chinese-trained mathematicians to win the Fields Medal at the International Congress of Mathematicians last week, Chinese social media produced a story of Wang’s depressing undergraduate years at Peking University that went viral within hours, drawing the stereotypical conclusion that China is no good soil for talents – they only shine abroad.

Within two days, the viral story was overturned with interviews and posts from prominent scientists close to the matter, shedding light on the “golden generation” of PKU mathematics department and the “platinum generation” that followed.

Wang ‘s research is in Fourier analysis and geometric measure theory, and her landmark result is on the three-dimensional Kakeya conjecture that centers on measure-zero sets: she and Joshua Zahl showed a Kakeya set in three-dimensional space can occupy zero volume and still have dimension exactly three.

Measure zero set can still be infinitely complex and take up a lot of space while containing no actual substance, just like the viral story that spread fast and furious, but empirically empty.

Wang laid her foundation in PKU, before heading to France for further studies and later, headed to the US. People in each of the three countries, including French President Emannuel Macron, wanted to claim the credit, while other people in the three nations reflect on what it means for the education system in the country.

The “missing” recommendation letter

The viral version held that PKU’s mathematics department refused to write Wang a recommendation letter for graduate study in France, and that her only letter came from Lei Jun, a class advisor in the unrelated Department of Earth and Space Sciences, where Wang had originally enrolled before transferring into math.

The story drew on a genuine interview in which Wang described feeling like she had merely survived, not excelled, as an undergraduate. It’s possible that she was describing a common feeling when enrolled in a top institution like Peking University or Harvard – suddenly you are no longer the best. She didn’t specify in the video interview, but it fit a familiar narrative: a brilliant student overlooked by a rigid, credential-driven system, vindicated only after leaving it

Biologist Rao Yi, a prominent public critic of Chinese academia with no institutional stake in defending PKU — stated on the record that Wang had received multiple letters, not one. Two came from PKU math faculty; Lei Jun’s, from Earth and Space Sciences, was a third. PKU’s mathematics and earth-and-space-sciences schools subsequently published their own accounts broadly corroborating Wang’s history there.

Other reports say the two math professors who wrote Wang recommendation letters kept low-key because she later switched her research subject – they don’t want to take the credit.

The more interesting question isn’t whether Wang had recommendations. It’s why the false version was instantly believed by so many, and the true one wasn’t dramatic enough to travel as far.

A stereotype past its expiration date

The rumor took hold because it matched a story both Chinese and Western audiences have told for years: Chinese STEM education optimizes for exam performance, not the independent thinking that produces Fields Medalists. That was never entirely fair, but it was closer to true a generation ago than it is now.

PKU’s own math department is a counterexample by its own recent record. It trained Wang and Deng, who shared the medal, and also Wei Dongyi— the ascetic, socially withdrawn “PKU math monk” whose public image is the opposite of exam-optimized careerism. Whatever the department’s reputational baggage, its output at the frontier is no longer in serious dispute.

Tsinghua’s Yao Class, founded in 2005 by Turing laureate Andrew Yao, was built explicitly against rote pedagogy: English-language instruction from year one, an MIT-derived curriculum from year three, direct classroom access to Turing and Gödel Prize winners, and an explicit “teach to aptitude” philosophy over standardized progression. Nearly 750 students have graduated since 2005. Close to 60 alumni now hold faculty positions, more than 30 at top domestic universities; alumni have founded Pony.ai and Megvii.

Most recently, a screenshot that shows DeepSeek hiring interns from Tsinghua Yao class at a daily wage of US$810 sent the class on top trending again. The claim was not verified by either DeepSeek or Tsinghua, but it reflects how valuable its graduates are viewed generally.

Another noticeable case is Moonshot founder Yang Zhilin, a Tsinghua computer sciences undergraduate who went to US for further studies and returned to found Moonshot. When Kimi K3 was unveiled two weeks ago, and proved to beat US frontier models in some benchmarks, American netizens kept asking “why couldn’t we keep Yang in the US?”

Where talent chooses to go

Individual stories can be dismissed as exceptions. Career decisions, taken in aggregate, are harder to wave away — and the direction of flow has shifted in measurable ways.

A 2023 Proceedings of the National Academy of Sciences study (Yu Xie, Xihong Lin, Ju Li, Qian He, Junming Huang) tracked a total of 19,955 scientists of Chinese descent who began their careers in the US but left for other countries, including China, between 2010 and 2021, with annual departures rising from 900 in 2010 to 2,621 in 2021. The study identifies the Department of Justice’s 2018 China Initiative as the clearest inflection point: departures increased 75% after the policy began, with two-thirds of relocated scientists moving to China.

A Stanford survey of 1304 Chinese-descent scientists still in the US found 61% considering leaving and 45% avoiding federal grant applications — a chilling effect on those who haven’t left yet.

More recent tallies suggest the trend is accelerating, not plateauing. CNN counted at least 85 scientists moving from U.S. institutions to Chinese research institutions full-time since the start of 2025, more than half of those moves occurring in 2025 itself. Princeton researchers separately tracked around 50 tenure-track scholars of Chinese descent leaving U.S. universities for China in just the first half of 2025, adding to more than 850 such departures since 2011 — the 2025 figure isn’t a standalone data point, it’s a sharp acceleration against a decade-long baseline.

Fear is one thing, where they are landing matters too and suggest improved salary, and work environment for scientists in China.

A population-geography study of roughly 1,250 returning researchers over 12 years found Beijing took more than a third of returnees, Shanghai 17%, and — notably — Cambridge, Massachusetts was the single largest source city, and the primary origin point for returnees to both Beijing and Shanghai. That’s a flow out of Harvard and MIT’s backyard, not out of institutions with nothing to offer.

The countervailing evidence matters equally. Over 83% of Chinese graduates who earned science and engineering doctorates in the US between 2017 and 2019 were still living there in 2023. Most Yao Class graduates still leave for doctoral study: of the 2025 cohort’s 81 graduates, 47 entered overseas PhD programs against 30 who stayed domestic. Studying abroad isn’t emigrating. The real shift isn’t in the outflow for training — it’s in the stay-or-return rate afterward.

By 2024, roughly 44.6% of Yao Class alumni were choosing to remain in China for further study or employment, a marked change from a decade earlier when overseas careers were the default for the program’s strongest graduates.

What this argues — and what it doesn’t

This isn’t an argument that China’s education system has overtaken America’s. It’s narrower: the forces that once pushed China’s best scientific talent permanently westward — better funding, better infrastructure, freedom to just do the work — have weakened faster on the American side than they’ve strengthened on the Chinese side.

Visa tightening, funding volatility, and a documented chilling effect on Chinese-descent researchers already in the US are pushing people out. Meanwhile China’s own research ecosystem — compute access, lab funding, internationally-taught programs like Yao Class — has made staying, or returning, a less obvious sacrifice than it was fifteen years ago.

The Wang Hong episode isn’t evidence of that shift. It’s evidence of something adjacent: the assumption that Chinese universities can’t produce independent, frontier-level researchers is now embedded enough — on both Chinese and Western social media — that people reached for a fabricated detail rather than sit with an ambiguous one.

The rumor needed invented specifics because the real specifics no longer cooperated with the story it was trying to tell.

China’s robotaxi industry buzzed recently with claims of the “first year of robotaxi profitability” after Pony.ai announced monthly per-vehicle operational profit in Guangzhou and Shenzhen, and Baidu’s sixth-generation self-driving vehicle drove down per-vehicle costs drastically.

However, do “positive unit economics” equate to “comprehensive corporate profitability?”

To measure the commercial health of robotaxis, the industry has long relied on two distinct metrics: single-vehicle unit economics and overall corporate revenue.

From the perspective of unit economics, Chinese companies have indeed crossed the golden crossover point. This is largely driven by a decline in hardware costs.

Take Baidu’s Apollo Go as an example. The bill for materials cost for its sixth-generation vehicle dropped to around 204,600 yuan (US$30,305), with costs for light detection and ranging sensors slashed by over 70 percent.

Simultaneously, today’s robotaxis have completely eliminated in-car safety operators in favor of remote cloud monitoring, further reducing labor costs and turning the marginal revenue per vehicle positive.

However, the reality at the corporate level may be much harsher.

Take ride-hailing service WeRide as an example. Its total revenue for 2025 reached 685 million yuan, representing a year-over-year increase of 90 percent. Of that robotaxi revenue tripled to 148 million yuan. Yet, its research and development expenses last year shot up to 1.4 billion yuan, resulting in an adjusted net loss of 1.25 billion yuan.

Even though robotaxi revenue is multiplying, research and development costs in the emerging, highly competitive market are eating away at overall profitability.

Apollo Go seems to be faring slightly better, having accumulated over 22 million service trips by April. Still, it relies on continuous financial lifeblood from parent Baidu’s other core businesses.

Thus, the “profitability milestone” currently proclaimed by the Chinese industry essentially means that technological maturity has successfully overcome hardware production and operational costs. But for an entire company to become profitable, operational scale must successfully overcome research and development expenses and capital amortization.

So, what paths can these companies take moving forward? Are there any precedents to follow?

Let’s turn our attention across the ocean to robotaxi operator Waymo in the US.

Alphabet-owned Waymo represents a typical “high-pricing, asset-heavy” model. It deploys high-density fleets in places such as San Francisco, Los Angeles and Phoenix, cities that have high personal incomes, strong travel demand and generally warm climates.

By the end of 2025, Waymo’s weekly paid orders had reached 450,000, which is 100,000 more per week than Apollo Go’s figures in the first quarter of 2026.

Waymo’s average revenue per trip ranges between US$15 and US$17. While slightly lower than local Uber rates, the absolute value is far higher than that in the Chinese market.

By deeply penetrating select core cities, Waymo directly pursues high per-vehicle profits to support its massive research and development efforts. However, this model cannot be easily replicated in China where road networks and a distinct consumer environment are different.

But Chinese companies do not necessarily need to copy Waymo’s playbook.

Compared with Waymo’s premium pricing, domestic companies are more likely to compete on price and a strategy of market penetration and mass deployment.

Baidu’s Apollo Go in Wuhan is a prime example. It quickly built up customer volume by pairing extremely low vehicle costs with fares cheaper than traditional ride-hailing services. The challenge of this model lies in the fact lower charges require higher fleet sizes to offset back-end costs such as cloud computing power, algorithm training, real-time updates for high-definition maps and the construction of remote cloud dispatch centers.

Today, the base fares for robotaxis operating in trial cities like Beijing, Shanghai, Guangzhou and Wuhan are comparable to standard ride-hailing and taxi services, but platforms frequently offer discount coupons to drum on trade. In pilot zones in Wuhan and Beijing, coupons for new users or regular discounts can drive the fare for trips of less than 10 kilometers down to between 5-12 yuan, significantly lower than the 15-30 yuan typically charged by comparable ride-hailing platforms.

According to estimates, robotaxi fleet size in a single city must exceed 10,000 vehicles, and the national total must surpass 100,000 vehicles. Only through such economies of scale can the massive AI research and development, and data center costs be diluted to just a few cents per kilometer.

Therefore, once a robotaxi network is large enough and user habits have been cultivated, pricing must gradually return to the levels of ordinary ride-hailing services.

Meanwhile, expanding overseas has become another inevitable path for Chinese robotaxi companies to generate profits. Today, both WeRide and Baidu are deploying robotaxis in the Middle East and Europe.

WeRide’s Middle East subsidiary already achieved operational profitability in 2025 by relying on an “asset-light” model. It delegates vehicle operations to local companies, such as the UAE’s Tawasul, and lists its self-driving taxis on the most popular local Uber app to pick up passengers, charging only a technology service fee.

Furthermore, robotaxis lack drivers, turning the interior of the car into a space that can be tapped for profit by offering amenities such as paid food and beverages, or massage seats or selling advertisements.

Xpeng’s robotaxi business is already offering super-VIP memberships that include QQ Music and NetEase Cloud Music as standard features, with every vehicle including a vast library of copyrighted content for free.

_{Visitors experience an AI-powered traditional Chinese medicine diagnostic robot at the 2026 World Artificial Intelligence Conference in Shanghai./IC}

For years, medical artificial intelligence has been shadowed by a blunt question: will algorithms replace doctors?

In China, a new physician survey points to a different reality. Doctors are not avoiding AI. They are using it widely but mostly on terms that keep final medical decisions in human hands.

The 2025 China Digital Life Physician study, based on a nationwide survey of 5,000 doctors across more than 360 cities, 2,200 medical institutions and 30 clinical departments, found that 91 percent of physicians had used or were familiar with AI tools, while 93 percent said AI-driven products and services brought more benefits than drawbacks. The study was conducted on the Gooddr platform, which says it has more than 4.6 million verified registered physicians in China.

But the findings also show the limits of AI adoption. The harder hurdle is whether AI can move from doctors’ phones into the formal systems where diagnoses are made, records are written and treatment decisions are reviewed.

Doctors can use Doubao or DeepSeek to search for information with almost no barrier, Li said. But once AI enters diagnosis, treatment or hospital workflows, it must clear four hurdles: clinical validation, access to hospital information and imaging systems, responsibility allocation, and hospital management procedures.

That helps explain why doctors can be enthusiastic about AI without wanting to surrender authority to it.

In the survey, 83 percent of respondents agreed that AI is a powerful assistant but said final diagnosis and treatment decisions should remain with doctors. Seventy-nine percent said AI would improve healthcare quality and efficiency without diminishing the professional value of physicians. Sixty-eight percent said they expect AI to change the way they work within five years, while 40 percent believed it would replace their current work.

Dr. Zhang Zhihan, an attending pathologist at Ruijin Hospital, said the change is already visible in pathology.

For Gooddr’s Li, the data show that doctors are not rejecting AI; they are defining its place.

AI is being accepted first in the parts of medicine treatment process where outputs are easier to check: information gathering, documentation, image pre-screening, patient education and follow-up. It is less welcome where judgment, responsibility and patient trust converge.

Li described the divide as one between the “information and execution” layers of the process and the “judgment and responsibility” layers.

That is why the first wave of deeper AI use is emerging in tasks with clearer boundaries, such as assisted reading of medical images and lab results, automated medical-record generation and structured patient management and follow-up. These tasks are standardized enough for doctors to review and narrow enough to fit into hospital quality-control systems.

For now, much of AI’s impact is more prosaic than revolutionary. It is not replacing doctors. It is starting to replace some of the repetitive cognitive work around them.

Current AI use remains concentrated in information searches, content generation and literature retrieval, Li said. But she said those tasks should not be dismissed as superficial because doctors spend large amounts of professional time searching medical knowledge, reading papers and drafting documents.

The result is not yet a clean saving of doctors’ time. Li calls it “time substitution.” AI reduces some low-value searching, but doctors still need to check outputs, verify sources and correct mistakes. Search is being replaced by review.

The study also challenges the assumption that medical AI will spread through China’s health system the way expensive medical equipment has. CT scanners, surgical robots and other capital-intensive technologies tend to appear first in large urban hospitals. General AI tools are different. They are cheap, accessible and already on doctors’ phones.

That may help explain one of the study’s more striking findings: doctors in lower-tier cities reported higher AI familiarity than those in top-tier cities, rising from 89 percent in top-tier cities to 97 percent in lower-tier cities.

The finding might not be read as evidence that lower-tier doctors are more technologically advanced. Rather, they may have fewer alternatives. “Where expert networks, multidisciplinary consultations and specialist resources are thinner, AI can become a readily available second opinion,” Li said.

In major hospitals, AI may start as an efficiency tool. In lower-resource settings, it may function more as a capability supplement.

Broader access, however, raises a harder question: what kind of AI can doctors trust? Tools such as Doubao and DeepSeek can help doctors search literature, draft documents or turn complex medical information into patient-friendly language. But high-risk clinical decisions require a different threshold.

The weakness of general models, she said, is not simply knowledge. It is control.

“They can produce confident errors,” she explained. “They may sound professional, but the answer can be outdated, misplaced or impossible to trace.”

That is why the next commercial race may not be about who has the largest model. It may be about who owns the clinical workflow.

For drugmakers, that shift could be significant. AI disease-management platforms may move competition from “product versus product” to “pathway versus pathway.”

That opens a new role for pharmaceutical companies, but also a new compliance risk. A disease-management platform can help patients stay on treatment and generate real-world evidence. It can also become a disguised promotional tool if its recommendations favor a sponsor’s own drug.

AI recommendations, she said, must be based on guidelines and evidence, fairly cover standard treatment options, keep patient data tied to care rather than marketing, and separate physician decision support from product promotion.

China’s biotech companies have learned how to sell drugs to the world. The harder part is proving the deals will create lasting value.

In the first half of 2026, Chinese drugmakers signed 81 overseas licensing deals valued at a potential US$110 billion, or 80 percent of the total value in 2025, according to China’s National Medical Products Administration.

China-related transactions accounted for about 34 percent of the potential value of US and European biopharma partnerships last year, up from roughly 8 percent five years earlier, according to a latest report from consulting firm Ernst & Young, which now calls itself simply EY.

The boom in drug deals has pushed Chinese companies to the forefront of new global industry growth, but the headline figures are typically tied to performance milestones that may take years to attain. At the same time, a Chinese drug may find itself competing for funding, clinical resources and priority inside a buyer’s pipeline.

EY found that only 32 percent of 225 life-sciences transactions it examined met their original growth targets. The next phase of China’s biotech expansion will therefore be judged less by how many deals are signed than by how many drugs make it through global development and generate revenue.

“The biggest point of failure is still clinical development – clinical, clinical, clinical,” said Felix Fei, co-leader of the health sciences and wellness for EY China.

Once a Chinese drug enters a multinational drugmaker’s portfolio, Fei said, it is no longer judged in isolation. It must compete with dozens of internal and external programs for budget, trial capacity and management attention. A promising molecule can still be delayed or pushed down the queue if it doesn’t fit a buyer’s broader pipeline strategy.

That is also why a large licensing headline doesn’t automatically translate into a lasting valuation boost. Investors are increasingly looking beyond the maximum deal value to the upfront payment, the conditions attached to future milestone payments and whether a company can produce another drug asset after selling its best one.

“The market is starting to treat business development as a source of recurring value, rather than a one-off story,” Fei said. “The real question is whether a company can be paid for its innovation again and again, not whether it can sell one asset at a good price.”

Global drugmakers have strong reasons to keep buying.

The world’s 25 largest biopharma companies could face a US$370 billion growth gap by 2032 as blockbuster products lose market exclusivity through patent expiry. At the same time, the industry has about US$2.1 trillion in dealmaking firepower.

“China offers what many buyers need: a large and fast-moving pool of clinical-stage assets,” Fei told us in an exclusive interview. “Its advantages lie in lower development costs, faster patient recruitment and strong engineering, chemistry, manufacturing and controls capabilities.”

He added, “Chinese companies are producing more assets in areas global drugmakers are already prioritizing, including oncology, autoimmune disease, metabolic disorders and neuroscience.”

That does not mean China has overtaken the US at the scientific frontier. Fei said the US still holds an edge in basic research, original target discovery and early scientific validation. China’s current strength is turning known or emerging science into trial-ready drugs more quickly and at lower cost.

For global drugmakers facing patent cliffs, that combination is difficult to ignore. But buying rights to a drug is easier than turning it into a global product.

A closer look at deal structures shows why investors are cautious. In 2025, upfront payments comprised only about 7 percent of the potential value of biopharma licensing agreements, according to EY. Further profits rest on development, regulatory and commercial milestones.

“The first number investors should look at is the upfront payment and its share of the total,” Fei said. “They should also ask when the money will be paid and what conditions must be met for the milestones to be triggered.”

A low upfront payment does not automatically mean the buyer lacks confidence, he added. It can also reflect the stage of a drug’s development, the amount of data available and how much risk the buyer is willing to bear at the start.

The same logic applies to the choice of partner. A smaller overseas biotech may sometimes be a better fit than a large pharmaceutical company if a drug fills a clear gap in its pipeline and will receive more management attention.

“Good assets do not always go to the biggest buyer,” Fei said. “What matters is whether the program fits the partner’s strategy and whether the team has the ability to move it forward.”

That fit will become more important as Chinese companies sign more deals. Global drugmakers are not short of options, and a licensed drug can still lose priority if trial costs rise, data disappoint or another program moves ahead of it.

The next question is whether Chinese drugmakers can shape drugs for global development before a deal is signed, rather than trying to retrofit them later.

Fei said companies that design trials around international regulatory requirements from the start are easier for overseas buyers to assess and advance.

“Assets developed first for China and adapted later may face questions over trial design, patient population and whether additional studies will be needed for regulators such as the US Food and Drug Administration,” he said.

That can affect both valuation and execution. A buyer may like the science but still discount the value of a new drug if the development package is not ready for a global program. It may also have to repeat work, add trial patients or redesign parts of the clinical plan before moving forward.

The challenge is increasingly organizational as well as scientific. Chinese biotech companies have become faster at generating candidates and early clinical data, but global development requires teams that can manage overseas trials, regulatory strategy, manufacturing standards and negotiations with multinational partners.

“Capital was the first-generation bottleneck. The second generation is people,” Fei said.

He said senior business-development executives who can understand the science, negotiate international deal terms and work within US or European regulatory frameworks remain scarce. For some biotech companies, chief executives still lead overseas negotiations themselves.

The shortage of savvy executive teams also may affect what companies retain after signing a deal. A drugmaker that hands over most overseas development may receive cash and reduce risk, but it also gives up experience that could help it manage the next drug discovery. Companies that retain a role in global trials or joint development may generate more long-term value, though they also take on more cost and execution risk.

A broader set of options is now emerging, from single-asset licensing and multi-drug partnerships to co-development, platform deals and self-funded overseas expansion. None is inherently superior, and the choice depends on a drugmaker’s capital, pipeline maturity and execution ability.

For most Chinese biotech companies, licensing remains the fastest route to overseas markets. But companies that command higher valuations may be those that can show the deal was not a one-time fluke and they are capable of developing other new drugs or taking on more of the global development themselves.

Six days ago, I argued in an opinion piece that Moonshot AI’s Kimi K3 wasn’t just another model release. It was evidence that Chinese AI labs had made a strategic choice the American industry was still debating: release open models, diffuse capability quickly and let the ecosystem compound.

Today, that industry debate took another twist, at least on paper.

Twenty-five US companies and organizations, including industry heavyweights like Nvidia, Microsoft, Meta, IBM and the Linux Foundation, published an open letter urging Washington policymakers not to restrict open-weight models.

Nvidia CEO Jensen Huang broke his own social media silence to post a copy of the letter -- his first message ever on X. Microsoft CEO Satya Nadella also posted the letter online. Elon Musk, whose xAI was not a signatory, posted simply: “This has my full support. Jensen is right.”

Two names are conspicuously absent: OpenAI and Anthropic.

That divide matters.

The timing of the letter is difficult to dismiss as coincidence. It followed the recent release of Kimi K3, a 2.8-trillion-parameter open-weight model that challenged frontier American models on multiple benchmarks and pierced the prevailing assumption that frontier AI capability is the exclusive domain of a handful of US labs.

Adding to the debate, US AI platform Hugging Face just disclosed that it had to resort to the analytics of open-weight Chinese AI developer Zhipu (Z.ai) to retrack an attack from an OpenAI model that went rogue. Hugging Face said that US guardrails prevented help from US models.

Whether any one or two incidents change anyone’s view of open vs closed AI, the Silicon Valley’s coalition letter underscores the need in real-world environments for models that can be inspected, controlled and run on multiple infrastructures.

This isn’t the first time similar arguments have arisen in industries beyond AI.

German automakers argued against EU tariffs on Chinese electric cars, warning that protectionism would weaken Europe’s own competitiveness.

US chipmakers argued against increasingly strict US export controls on shipments of advanced semiconductors to China, saying such measures could damage US technological leadership.

Now AI infrastructure companies are making a similar case, arguing that restrictions of open-weight models won’t slow global AI expansion and could weaken the American ecosystem building up around it.

Every generation of advanced technology eventually reaches the same political crossroads. Industries argue for openness. Governments weigh the impact on domestic economies, sometimes invoking national security concerns. Markets wait to see which side wins.

The US tech heavyweights who signed the public letter to policymakers are, of course, keeping an eye on their own bottom lines, especially as investors express increasing concerns about the massive amounts of money, often debt-financed, being plowed into building AI infrastructure.

On Friday in New York trading, AI related shares dropped again, with the Philadelphia Semiconductor Index falling around 4.3 percent and newly listed SK Hynix shares plunged 8.8 percent as investors continue to reprice how such vast capital spending will affect earnings.

AI infrastructure companies comprised more than half of the 25 signatories to the public letter. That makes sense. They need a new sweet spot to convince investors that future AI spending is a good business decision.

Nvidia benefits when more models run on its graphics processing units. Microsoft profits from hosting every major model, whether proprietary or open, and IBM benefits from corporate needs for more computing power. Meta, which owns social media platforms, monetizes user numbers rather than model licensing, making an open ecosystem commercially attractive.

OpenAI and Anthropic, which are both planning big IPOs in the US, operate under a different economic logic.

Their flagship models are not simply products but rather the very foundation of their business models. Their frontier models benefit from being scarce assets that the company relies on for profits. Hence the missing signatures.

Sam Altman, chief executive of OpenAI, posted on X: “I want the US to win in both open and proprietary models.” That was essentially a “no” to the Silicon Valley letter.

Trump administration officials are loading their weapons. Treasury Secretary Scott Bessent earlier said the government is investigating whether Chinese AI companies are thieving American intellectual property and noted that sanctions might be imposed if theft is established. White House technology officials have alleged that Moonshot AI pinched information from restricted Nvidia hardware, an allegation that has been denied.

Washington is actively considering where open-weight AI fits within its broader technology competition with China. And the US Congress is considering legislation that would tighten US guardrails against AI spinning out of human control.

My expectation is that no sweeping US ban on open models will eventuate. Instead, a more likely outcome is targeted action against specific companies tied to specific government grievances, giving the Trump administration the appearance of toughness against China while leaving the broader open-weight ecosystem intact.

Looking over the range of arguments now raging, these are my bets.

• OpenAI and Anthropic won’t change course. Closed models aren’t simply a policy preference but rather a profit driver. Open-weight models don’t mean giving up monetization altogether, but they bet on a better commercial future. An eco-system can’t be built overnight, making it almost impossible for a company that fully relies on profits from closed models to move in that direction.

  Both companies want to impress investors ahead of IPOs, and one way to do that is the promise that short-term profits are more easily quantified with closed models.

• The coalition of letter signatories will ultimately lose their fight. Industry versus government usually ends that way. US chipmakers have spent years arguing against tighter export controls on shipments to China, and Washington has largely ignored their pleas. The EU rebuffed tariff warnings from German automakers and imposed levies on imported electric cars anyway, though intense lobbying did dilute them.

  If Washington concludes that unrestricted open-weight models are a risk to the interests of the US, voices from Silicon Valley are unlikely to hold sway.

• This open letter may accelerate the very trend the coalition is trying to slow. If US open-weight models become more constrained while capable open-weight models continue emerging elsewhere, the AI trend will simply become less American, and US restrictions aimed at preserving national dominance will crumble.

  That is the irony beneath the Friday letter.

“Bhaiya,” he said, “my Claude account is gone!”

He had opened his laptop that afternoon to work on an assignment due at the end of the week. He uses Claude, the Anthropic chatbot, the way students everywhere now use these tools: to check drafts, tighten arguments and catch the grammar slips that come with writing in a second language.

Anthropic accepts only international credit and debit cards, a quiet barrier in a country where local mobile wallets dominate daily payments. My brother used a Visa card from a Bangladeshi bank, issued in his own name, to pay Anthropic’s US$20 fee. That afternoon, instead of the familiar chat window, he found himself locked out. He received an email from Anthropic’s safeguards team, with an explanation of sorts. An internal investigation had found “suspicious signals” on his account, it said, indicating a violation of the company’s usage policy, and his access had been revoked. It didn’t explain further.

Then came a second message. In the cloying language of automated customer service, it told him his subscription payment had been refunded. He checked his statement the next day and ensuing days, but no refund appeared. The message said nothing about the US$30 dollars of prepaid usage credit he had bought separately, which vanished with the account. For a student in Dhaka, US$50 is lot of money to lose.

He did what the email told him to do. He logged in, filed an appeal, explained his situation, and mentioned the school deadline bearing down on him. More than a month later, he had received no response or even any sign that his appeal had been read.

On the phone that evening, he asked me a question that lingers with me: “You live in China. Which tool should I switch to?” At twenty-something, he had already learned a valuable lesson. When an AI company cuts you off from the other side of the world, the only remedy is to find another company.

One family’s grievance is not a governance argument. I use Anthropic as an example not because it is a uniquely bad actor. In one respect it is better than most: it publishes enforcement data on its own transparency hub, whereas most technology companies do not.

In a disclosure report for the second half of 2025, Anthropic reported 1.45 banned accounts, with 1,700 of 52,000 appeals overturned. The figures include enforcement related to fraud, spam and abuse. They disclose the scale of enforcement and say nothing about the quality of any single decision: whether the customer learned the reason, whether anyone reviewed the appeal and whether the promised refund arrived.

One detail of his case deserves attention. Anthropic’s consumer terms say that a user terminated for violating the terms gets no refund, while a subscription terminated for any other reason earns a pro- rata refund for its remaining period. My brother’s ban notice alleged a usage policy violation, which means no refund was owed. Yet the company also told him his payment had been refunded.

The contradictory messages are hard to reconcile, and both sit awkwardly beside his bank statement, where no refund has appeared.

Here is the structure underneath his story. AI services now operate in most countries around the world, but each company answers to courts and regulators in only one or two. Anthropic’s terms designate California law and San Francisco courts for disputes and assert that users waive claims under other jurisdictions. Lawyers debate how far such clauses hold up in any given country and if mandatory consumer protection rules can be contracted away.

For an ordinary student, though, the debate is academic. Whatever rights Bangladesh consumer law preserves on paper, enforcing them against a foreign company would cost far more than the claim is worth. Bangladesh does have a consumer protection authority, the Directorate of National Consumer Rights Protection, with a hotline and a complaint process. But a complaint against a foreign provider is hard to investigate and harder to remedy

Why call this an AI governance problem instead of normal internet commerce? Frozen accounts and unreachable support exist in gaming and streaming too, and the consumer-law gap predates AI. What AI adds is the depth of the dependence. These accounts are becoming personalized working infrastructure. They hold accumulated conversations, custom instructions and workflows that users build their studying and livelihoods around.

Enforcement decisions themselves are increasingly automated and opaque. And a handful of foreign firms are gaining traction in developing markets faster than redress institutions can keep up. Researchers have documented the pattern in studying how Global South stakeholders are sidelined in the institutions that write AI’s rules, and Global South governments have begun organizing against it, most recently through a New Delhi forum on international AI accountability.

Yet that conversation centers on sovereignty, data and national capacity. A framework that secures data sovereignty for states while leaving the individual customer without recourse will have solved the dignified half of the problem and skipped the daily half.

The daily half is fixable, and Europe has shown how. The European Union’s Digital Services Act requires platforms to give users a statement of reasons when accounts are restricted, to run internal complaint systems and to provide access to out-of-court dispute settlement. Whether standalone AI chatbots fall fully within those platform rules is still contested among lawyers.

The EU service proves that reason-giving, appeals and independent dispute resolution can be imposed on major digital services at scale. Anthropic itself now publishes transparency reports under the EU act, covering enforcement and complaint-handling for the Claude. ai features within the law’s scope in Europe. This procedural machinery already runs for covered services there.

A workable global baseline is not hard to sketch. Any restriction on a paying account should come with meaningful stated reasons, allowing narrow exceptions where disclosure would help bad actors evade security systems. A paying user who lodges a non-frivolous challenge to a permanent termination should get a meaningful human review and a decision within a reasonable fixed period, say 14 days -- a mark my brother’s unanswered appeal has long passed.

Unused paid balances should come back automatically, in a way the customer can verify, unless funds are lawfully withheld for documented fraud, sanctions or payment disputes. And the companies should be required to join an independent cross-border, dispute-resolution scheme, open to users in every market where they sell subscriptions and modeled on the ombudsman systems that already govern international banking and telecommunications. Each of these exists somewhere in law or practice today. What does not yet exist is a commitment to offer them everywhere the companies take money.

The venues for building wider commitment are taking shape. Last week, China introduced the World AI Cooperation Organization (WAICO), signed by 29 countries, that states AI governance should be inclusive and that developing countries deserve an equal say in writing the rules. Earlier this month, the UN convened the inaugural Global Dialogue on AI Governance in Geneva, which closed with a pointed message: what matters now is not new declarations but concrete actions.

Universal consumer redress is exactly the kind of step that call describes. It does not require the US and China to resolve their competition over technology chips, models and talent. Great powers have long competed while agreeing on baseline rules for shared systems, from aviation to international payments. American companies, for their part, have every commercial reason to want the durable trust of users in the emerging markets where their next billion customers live.

My brother will be fine. He found another tool, finished his assignment and moved on, the way students do. But I keep returning to that phone call, and how quickly he stopped expecting fairness and started shopping for alternatives. His shrug points to the larger risk. As dependence on a few foreign AI services deepens across the developing world, users are developing a level of skepticism that could undermine a stable foundation for the AI ​​century.

Helping a single student in Bangladesh address an injustice may not be earth-shaking in the global scheme of things, but rules to protect consumers are important first steps in addressing larger problems.

I never expected a hospital stay to become my crash course in China’s medical innovation ecosystem.

Usually, when people talk about innovation, they imagine laboratories filled with scientists, venture capital meetings or entrepreneurs presenting ambitious ideas in polished stages. I discovered it from a hospital bed, wearing a patient wristband and waiting for my next examination.

After spending nearly two weeks in hospital recently because of a health issue, I experienced something that is easy to overlook from the outside: China’s medical innovation is no longer just a story about patents, investment numbers or government plans. It is increasingly becoming a story about patients.

The equipment around my bed told part of that story. The monitoring and assistance devices used in my room came from a major Chinese medical equipment company. In the past, similar equipment in the same hospital was often supplied by global companies. The change was subtle. There was no dramatic announcement, no ribbon-cutting ceremony. It was simply there, helping doctors treat patients.

The medicines I received also reflected a changing landscape. One foreign drug I was prescribed had already moved to relying on raw materials produced in China. Another had just received its first global approval in China.

Out of curiosity, I asked my doctors about how these innovations actually enter hospitals. I learned that the hospital participates in numerous collaborative research projects with both multinational companies and Chinese innovation-driven companies. These partnerships were not theoretical. They involved real clinical needs, real doctors and real patients.

_{The photo of how my hospital bed looiks like was processed by AI to remove patient name and hopital and equipment logo.}

One detail surprised me. According to doctors involved in these collaborations, Chinese companies often respond faster when hospitals propose improvements or new requirements. A technical adjustment that might take a few months to complete locally could take longer with multinational companies because requests need to travel through global headquarters. Some global companies are now redesigning their internal processes to shorten that cycle.

Speed matters in medicine. A three-month improvement may sound small in a corporate meeting. For a patient waiting for better treatment, it can mean something very different.

This is where China’s medical innovation story becomes more complicated than the simple narrative of “China catching up with the West.”

China’s speed has not come from one single factor. Regulatory reforms have played an important role by improving review processes and encouraging faster approval of innovative drugs and devices. Companies have increased research investment because a large market can create commercial returns. Hospitals and doctors have become more willing to participate in innovation and testing.

The result is a more connected system. Regulators, companies, hospitals and clinicians are increasingly moving in the same direction: turning ideas into products that reach patients.

But this system also faces a difficult question: how do you balance the interests of everyone involved?

Innovation requires investment. Drug makers and medical equipment manufacturers need reasonable returns. Hospitals need resources. Doctors need incentives. Patients need affordable access. The interests do not always perfectly align.

The challenge for China, and for every country trying to build a stronger, more innovative healthcare system, is not simply creating more products. It is creating an environment where innovation can survive, scale up and ultimately improve people’s lives.

Shanghai offers an interesting example of this experiment.

The city has invested heavily in healthcare infrastructure and has one of China’s strongest medical ecosystems. But what impressed me during my hospital stay was not only government investment but also the interaction between public resources and market forces.

One small device used during my treatment was actually the result of hospital-led innovation. A doctor developed the idea, the hospital helped transform the invention and a company eventually produced it for wider use.

I later learned that the hospital regularly organizes “matchmaking” meetings, bringing doctors and companies together to identify new ideas that can be commercialized.

This may sound ordinary. In reality, this is how innovation often happens. Not in a dramatic “eureka” moment, but through hundreds of small improvements created by people who understand real problems.

At the same time, China’s medical strides should not be viewed through a narrow national lens. Innovation does not belong to one country.

Some multinational pharmaceutical companies are increasing investment in China, expanding production capacity and research capabilities. From conversations with people familiar with the industry, several global healthcare companies are considering deeper commitments to China in the coming years.

Their message is clear: China is becoming an increasingly important part as part of the global healthcare supply chain.

These companies are not only looking at China because of market size. They are also looking at China’s research capabilities, clinical resources and industrial ecosystem. The goal is to connect Chinese innovation with global development to bring new treatments to more patients.

At the same time, Chinese companies are also learning from the world. Some domestic companies are working with international research teams to develop specific technologies and components. The future may belong less to companies from a particular country and more to companies that can connect the best ideas from different places.

This is why the recent decline in global technology stocks may not be entirely bad news.

When technology stocks around the world fell, Chinese technology and healthcare-related companies were also affected. The decline was painful for investors, but market corrections have one useful function: they separate real innovation from attractive stories. When the tide rises, many companies look impressive. When the tide goes out, we see who is actually swimming.

Healthcare innovation is no different. A good idea is not enough. A company needs technology, clinical value, commercial discipline and the ability to deliver effective products to patients.

After my hospital experience, I have become more optimistic about China’s medical innovation. But my optimism is not based on the idea that China will replace everyone else.

The more interesting possibility is that China is becoming a major participant in a global innovation network where different strengths come together, with the patient as the ultimate judge.

During those two bed-ridden weeks, I did not become an expert on medicine. I still cannot explain every machine beside my bed, and I certainly do not plan to volunteer for another “field research project” anytime soon.

But I did learn one thing: innovation is not an abstract word.

Sometimes it is a new drug.

Sometimes it is a better medical device.

Sometimes it is a doctor’s idea turning into a product.

And sometimes, it is simply something quietly working beside a hospital bed while a patient is trying to get better.

I went to the hospital hoping to recover my health. II did not expect to leave with an idea to share with others.

My doctor may advise me to curtail my opinion-writing next time. I hope he also prescribes fewer hospital visits.

#ChinaShapingtheFuture

_{BYD displays its flash electric-vehicle charger, with plans to install more than 3,000 of them in Europe next year./Feiran Lyu}

A quiet consensus is forming in the 2026 global auto market: the next vehicle you drive, or the robotaxi you hail, is increasingly likely to run on a “brain” designed in China, even if the badge on the hood is German, Japanese, or American.

For a century, the auto industry’s profit pool sat with hardware suppliers — Bosch, Continental, ZF — who built empires machining standardized components and licensing closed operating systems.

That model is being displaced.

The new suppliers aren’t judged on tolerances; they’re judged on how fast they can retrain a model and how much data their stack can turn into usable driving decisions.

WAIC 2026, which closed in Shanghai on Monday, was effectively a trade show for that transition — several of the technologies on display are already running on foreign roads, not just foreign PowerPoints.

Why the value moved

The mechanism is straightforward: hardware margins compress once components are commoditized, but software compounds.

A traditional Tier 1 supplier sells a part once per vehicle generation, on a multi-year refresh cycle. A software stack updates over the-air, continuously, and the company that owns the data pipeline keeps capturing value long after the car leaves the factory. That’s the structural reason Chinese AV and cabin-software firms — several of them still unprofitable — are getting valued like infrastructure plays rather than parts suppliers.

_{WeRide is planning worldwide penetration of its autonomous driving technology in collaboration with Bosch./Feiran Lyu}

What’s independently checkable is the footprint: WeRide operates in 12 countries and 40-plus cities, holds self-driving licenses in eight countries, and runs a global fleet north of 3,000 vehicles, with driverless taxi operations in Riyadh and Singapore and expansion underway in Madrid and Zurich.

The company also disclosed last week that its ADAS software is in road testing and localization in Germany, France, and Japan — traditional strongholds of the industry it’s now supplying into.

Westwell, the Shanghai port-automation firm, showed its Q-Truck autonomous heavy duty EV alongside its operations platform. By the company’s own count, 70% of the world’s top 20 container ports had partnered with it as of 2025; it’s automating ports in Thailand, the UAE, the UK, Egypt, Malaysia, and Mexico.

The UK’s Port of Felixstowe just placed a third Q-Truck order, taking its fleet to 100 vehicles — reportedly one of Europe’s largest automated electric truck deployments. ModelBest, the Beijing cabin-AI startup, demoed SuperMate, its in-car agent already shipping in the Changan Mazda EZ-60 and Geely Galaxy M9.

Co-founder Lei Shengtao pitched it as solving the lag problem in voice assistants — simultaneous see/hear/speak rather than ask-and-wait — and claimed China’s cabin and AV software is running roughly two years ahead of overseas competitors.

Pony.ai (Nasdaq: PONY) is scaling into production with Toyota, targeting 1,000 robotaxis deployed this year, and co-developing fully autonomous vehicles with Stellantis in Europe — Chinese software running inside a European chassis, literally. Nullmax is pairing with Renesas to build cost-competitive driving stacks for global automakers.

The strategic logic — and the part the boosters skip

The pattern reads as a calculated response to tariff walls: rather than exporting finished vehicles into markets that tax them heavily, Chinese firms are exporting the architecture, the algorithms, and increasingly the chips — embedding themselves inside vehicles that never touch a Chinese port.

Because these systems are AI-native, they adapt to autobahn speed limits, European roundabouts, or Southeast Asian traffic density without a full rewrite. What this framing tends to leave out: Regulatory risk is rising, not falling.

Washington has already moved to restrict Chinese-made connected-vehicle software and hardware over data-security concerns; the EU is watching the same question. A strategy built on embedding Chinese code inside Western vehicles is a strategy exposed to the exact security review that’s currently expanding, not shrinking. “Two years ahead” is a sales claim. Useful as a signal of confidence, not as a verified technical gap.

Scale claims (fleet counts, port-partnership percentages) are company disclosed. They’re plausible and directionally consistent with public deployment news, but they haven’t been checked against counterparty confirmation here. 1,000 robotaxis “this year” is a target, not a delivered number — worth revisiting closer to year-end.

The bottom line

The direction of travel is real: value in the auto stack is migrating from the machinists to the model-builders, and Chinese firms got there first partly because their home market forced faster iteration under harsher price competition. Whether that lead survives contact with Western security regulators — rather than Western consumers — is the open question we will be tracking through the rest of 2026.

It is tempting to see two separate futures from the twin events -- China making AI tangible while America chasing artificial general intelligence. A better reading is that each side was confronting what the other already knows: machine intelligence must leave the laboratory, but deployment without deeper capability soon reaches a ceiling. The next stage will need both.

_{Humanoids turned from big toys to workmen this year./Dong Jun}

Xi’s presence at the Shanghai event reinforced the top priority the government places on AI as a national strategy, but the origins of the science are international. The quest for artificial intelligence began in 1956 at a Dartmouth University workshop in the small US state of New Hampshire. That’s where the term “artificial intelligence” was first coined, leading to seven decades of work by researchers around the world.

Xi sought to build on the international future of AI, pledging that China will spearhead access to lesser developed economies through the creation of the new, Shanghai-based World Artificial Intelligence Cooperation Organization, with 29 signatories and a global body dedicated to governance and AI collaboration.

That vision was reinforced two weeks ago when the Global AI Capacity Development Network was launched in Geneva, involving 23 centers from 19 countries and organizations focused on education, open science and policy capability.

_{The United Nations-promoted Global Network of Centers for Exchange and Cooperation on AI Capacity Building was set up in Geneva in early July.}

On the exhibition floor at the Shanghai conference, China’s bet on the globalization of technology was clear. World models, embodied intelligence and physical AI drew crowds because they connect software to China’s factories and households. The catch-phrase huandao chaoche (换道超车–), or “changing lanes to overtake,” can sound like a search for a shortcut. In practice, the new lane is built from hard-won strengths in batteries, electric vehicles, motors, sensors, factories and supply-chain engineering. Companies such as Unitree and AgiBot are trying to turn those assets into machines that can navigate warehouses, production lines and homes.

Robots are the visible end of a deeper stack. Their performance depends on language and multimodal models, agents, simulation, data and computing infrastructure – the less photogenic work being done by DeepSeek, Moonshot AI, Z.ai, Alibaba’s Qwen teams, ByteDance, Huawei and many specialist suppliers. China’s distinctive opportunity is not robots instead of models. It is the integration of models into the physical world, followed by the diffusion of that capability beyond a small group of technology companies.

Moonshot released its Kimi K3 model on the eve of the Shanghai conference, grabbing world attention with its 2.8 trillion parameters, 896 experts with 16 activated for each token, native vision and a one-million-token context window. Moonshot says architectural changes produced 2.5 times the scaling efficiency of Kimi K2.

More revealingly, K3 is not priced as a bargain substitute. Its output costs US$15 per million tokens. Moonshot is asking customers to pay for completed work, not merely inexpensive inference.

That is where “taste” enters the Chinese model race. Moonshot demonstrated K3 building a compact graphics processing units compiler, conducting a 48-hour simulated chip-design run, recreating an astrophysics workflow and editing a launch video from 56 clips.

But such demonstrations are not independent scientific validation. Moonshot candidly admits that K3 still trails Claude Fable 5 and OpenAI’s GPT-5.6 Sol overall, and has a “noticeable gap” in user experience. But that admission helps define the next contest. Taste means knowing when a chart is unreadable, when an agent should ask rather than improvise and when a technically correct result is a work in progress.

This is the “second half” of the race described by Tencent chief AI scientist Yao Shunyu.

China has the users to test that proposition: QuestMobile counted 499 million monthly users of AI-native apps in May, up 85 percent in a year, while the China Internet Network Information Center counted 602 million generative-AI users by December 2025.

Those users are already testing AI in commerce, entertainment, education, automobiles and office work. Scale does not automatically produce intelligence, but it reveals which forms of intelligence people can actually use.

There is also a return to research fundamentals. In July, Z.ai co-founder Tang Jie announced a two-year “Touch High” plan that would shift the focus from short-term application revenue to long-horizon agents, self-training and safety governance. That is a meaningful departure from the internet-era habit of treating daily active users as the final measure of value.

China’s compute constraints remain real. Researchers interviewed during a spring tour by a group of AI writers and researchers repeatedly raised the issue of limited access to advanced computing, while Nathan Lambert from the Allen Institute estimated that research and development can consume roughly 80 percent of the computing involved in producing a frontier model.

Yet scarcity has also rewarded efficient architectures and open weights. DeepSeek, Qwen and Kimi have given universities, smaller companies and developers across the global south tools they can modify rather than merely rent. This is more than a defensive response to chip controls; it can become a durable contribution to global AI research.

_{Chinese companies are building up compute in various ways. / Dong Jun}

The next obligation, therefore, falls heavily on companies. Chinese laboratories seeking global trust should publish stronger evaluations, report serious failures, fund independent red teams and protect the rights of workers, creators and users. China already has extensive government rules. Its next phase requires corporate commitments that can be examined from outside the conference hall.

The AGI Summit in San Francisco expressed the Silicon Valley’s frontier conviction. Its program was more practical: agents, foundation models, robotics, coding, open source, five-minute product demonstrations and direct access to investors. The event’s slogan moved from “AI assists you” to “AI executes for you.”

American frontier labs still speak as if artificial general intelligence, and increasingly superintelligence, is an engineering destination. Their behavior is more concrete. Anthropic released Claude Fable 5 on June 9, emphasizing long-running work and coding. OpenAI released GPT-5.6 on July 9 across ChatGPT, Codex and its application processing interface. In San Francisco, Anthropic’s Aengus Lynch was billed to discuss Claude for enterprise and alignment; OpenAI’s Rohan Varma represented Codex. The frontier race is now simultaneously a research contest, a developer-tools war and a fight to become the interface through which companies reorganize work.

That battle is growing more intense. The Wired reported that contractors working on Meta’s Cannes project posed as minors and sent rival chatbots more than 45,000 prompts in one testing round, probing suicide, sex, eating disorders and drugs. Meta described it as safety benchmarking; Google said it had not authorized the testing.

Palantir Chief Executive Alex Karp attacked the frontier labs from another direction, saying companies were “paying for tokens that create no value” while risking the loss of proprietary knowledge. Microsoft then launched its US$2.5 billion Frontier Company with more than 6,000 specialists to embed inside customer organizations and integrate models without ceding the customer’s intelligence to one provider.

The competition is not simply Claude versus Codex. It is model labs versus cloud platforms, consultants and company-data incumbents versus who owns the learning loop.

Capital makes that competition unusually durable. Of the US$300 billion of global venture investment in the first quarter of 2026 reported by Crunchbase, AI companies captured US $242 billion. OpenAI’s US$122 billion round alone accounted for roughly two-fifths of the total. In a politically and economically unsettled America, investment in AI is still one of the few shared convictions.

Social permission is less assured. The “We Must Act Now” initiative, which attracted more than 200 initial signatories, including 16 Nobel laureates, calls for AI that complements rather than displaces people. Opposition to data centers, anxiety about child access to harmful online content and arguments over work displacements are now part of American politics. That resistance need not be treated as anti-technology. It is the process through which a society decides who suffers from disruption and who receives the gains.

_{The “We Must Act Now” initiative is backed by 16 Nobel laureates.}

China has become a convenient scapegoat in the process. When a lab wants tighter controls, China is accused of distilling US models. When investors want faster permits, China is accused of innovating too quickly. When reformers seek worker protections, Eric Schmidt, former Google chief executive who now heads Relativity Space, and his AI chief Selina Xu return from China reporting that some Chinese executives would rather slow deployment than trigger a Luddite reaction.

Schmidt has argued that while America builds the most powerful systems, China is using AI to build a more powerful economy.

The China-US competition is no cause for panic. It has an inherent complementary nature. America’s frontier culture needs stronger channels into industry and public life; China’s deployment culture benefits from continued penetration in those realms. Each country is stronger near the other’s missing pieces.

The most important China-US dialogue may begin with each side becoming curious about the other’s questions.

At the Shanghai conference, President Xi called for “mutual learning between civilizations” and for technology that preserves cultural differences. AI systems respond not only to calculations, but to grief, doubt, conflict, ambition and love. Every model, therefore, carries an image of human beings, whether their creators acknowledge it or not.

China is more open to this reflection than many Western observers assume. When Pope Leo XIV published the Magnifica Humanitas this year, Chinese technology outlets and social media produced lengthy readings connecting the encyclical to employment, education, healthcare and digital life. One summarized the pope’s statement memorably: “Every sentence mentions AI; every word is about humans.” This was recognition that a two-thousand-year-old institution was asking what benchmarks cannot: what must remain irreducibly human?

Claude’s “constitution” has received similar attention. The Beijing Academy of Artificial Intelligence described its explicit moral principles as a shift from reward and punishment toward education. Tsinghua’s Institute for AI International Governance asked the harder question: if AI has a constitution, who writes it, whose values does it represent and who may revise it? This is intellectual openness–not accepting an American company’s answers, but studying its experiment and continuing the argument.

Silicon Valley, in turn, is reopening its doors to the humanities. Anthropic’s Amanda Askell has helped shape Claude’s character and moral reasoning. In 2026, the company also recruited Harvey Lederman, a philosopher of AI mentality and Chinese neo-Confucianism. His award-winning work examines Wang Yangming’s “unity of knowledge and action,” the idea that genuine moral knowledge cannot be separated from conduct. The symbolism is rich -- an American frontier laboratory asking a scholar of Chinese philosophy to work on alignment and character. Ideas are crossing the Pacific even as chips face tighter borders.

This makes a new narrative possible. Liberal traditions speak about rights and agency; Chinese thought deepens ideas of relationship and harmonious coexistence. Christian social thought insists on dignity and the common good; other traditions bring their own visions of community, wisdom and the living world.

The task is not to blend them into a slogan, but to let each reveal what the others fail to see.

The robots in Shanghai offer a final reminder. Intelligence enters history through bodies: the hands of a worker, the attention of a teacher, the vulnerability of a patient, the curiosity of a child and the ecosystems sustaining them. Mankind’s flourishing cannot be reduced to tokens, automation rates or the arrival date of artificial general intelligence.

As the two most important players of the AI era, China and America can still build together if they imagine a future larger than victory–one in which AI helps different peoples live, create and coexist without surrendering the relationships that make AI worth having. The deepest openness is allowing another civilization’s question to change one’s own answer.

_{Kimi booth at #WAIC, July 18, 2026}

The next phase of AI competition may not be just about building the smartest models, but about building the largest ecosystems around them.

At the ongoing World Artificial Intelligence Conference in Shanghai, I stopped by Moonshot AI’s Kimi booth. It was surprisingly low-key for a company that had just achieved something few AI labs anywhere have managed: releasing one of the world’s largest open-weight frontier models.

This week, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model. It features a one-million-token context window, native multimodal capabilities and a new architecture called Kimi Delta Attention, designed to make inference at that scale more economically viable. Without architectural innovations like this, deploying models of this size would become significantly more expensive and challenging. Parameters are the learned values within a machine-learning model that determine how it maps inputs to outputs.

Moonshot, which is backed by Alibaba and Tencent, said Kimi K3 is the world’s first open-weight model to approach the 3-trillion-parameter mark and is designed for advanced reasoning, long-horizon coding and knowledge work.

Its release follows the unveiling of GLM-5.2, a 744-billion-parameter open-weight model from Zhipu, which Reuters noted “narrowly trailed leading closed-source models” from the top US firms on benchmark tests. China’s MiniMax is also developing its own 2.7-trillion-parameter model.

On Arena.ai’s Frontend Code Arena leaderboard, Kimi K3 currently ranks No. 1 among tested models, ahead of Claude Fable 5, GPT-5.6, GLM-5.2 and Claude Opus 4.8. While a single benchmark cannot determine overall frontier AI leadership, the result sends an important signal: an open-weight Chinese model is now competing at the top tier of coding benchmarks alongside leading proprietary systems.

Like every major release since Kimi K2 last year, K3 is open-weight, meaning its core model weights are available for developers and researchers to download, run and build upon.

That is where the bigger story begins.

The Strategy Behind China’s Open AI Push

Moonshot’s success reflects a broader shift across China’s AI industry.

Many of China’s leading foundation-model developers — including DeepSeek, Alibaba’s Qwen, Zhipu’s GLM and Moonshot’s Kimi — are embracing openness as a core strategy. While their licensing approaches differ, they share a common belief: frontier AI creates more value when developers, researchers, startups and companies can build on top of it.

Open weights are one expression of a broader philosophy that differs from the dominant approach among many U.S. frontier AI companies, where models from OpenAI, Anthropic and Google largely remain proprietary and are primarily accessed through APIs. Elon Musk’s xAI’s Grok early on transitioned out of open-weight.

Meta’s Llama was once the major U.S. exception, publishing open weights at significant scale. That changed in April, when Meta pivoted to Muse Spark, its first fully closed model, citing concerns including unauthorized use of Llama’s architecture by other labs. Even that American holdout for openness has moved toward the closed side. The difference isn’t simply about licensing. It’s about where each ecosystem believes long-term value will be created.

The Culture Behind the Strategy

Moonshot’s commitment to openness also reflects something about the company’s culture.

Two of its four founders, Tsinghua University graduates Yang Zhilin and Zhou Xinyu, played in a campus rock band together as students. The company’s Chinese name, Dark Side of the Moon (月之暗面), is a tribute to Yang’s favorite Pink Floyd album, and Moonshot reportedly launched on the album’s 50th anniversary. Meeting rooms at the company’s Beijing headquarters are named after legendary rock bands.

_{A screenshot from the archived website of Splay, the Tsinghua computer sciences major rock band where Zhou Xinyu (second from left) played guitar and Yang Zhilin (third from left) was the vocal.}

It is a small detail, but one that hints at a company culture comfortable with unconventional thinking — a trait that also appears in its technical choices.

While much of the frontier AI race has focused on building larger proprietary models, Moonshot has repeatedly chosen a different path: releasing increasingly capable open-weight models and betting that developers, rather than exclusivity, will become its competitive advantage.

In that sense, K3 does not feel like a sudden breakthrough. It feels like the next step in a strategy that has embraced unconventional bets from the beginning.

Why Openness Matters

Every major open-weight release strengthens the ecosystem.

Developers gain another frontier model they can deploy and customize. Companies can run models within their own infrastructure, keeping sensitive data inside their own networks while fine-tuning models for specialized applications. Researchers can inspect and experiment with architectures rather than treating frontier models as black boxes. Startups can build products without relying entirely on a single API provider.

Progress compounds across the ecosystem instead of remaining concentrated inside a small number of companies.

K3 matters not only because of its benchmark performance, but because it reinforces that ecosystem dynamic.

But How Do Open Models Make Money?

Companies exist to generate profit, so the obvious question is: how can open-weight models become a viable business?

The answer is that releasing model weights does not mean giving away the business.

There is also a pragmatic hardware calculation. With advanced US chips restricted, distributing model weights allows more companies and developers to share inference workloads across their own infrastructure. This can reduce pressure on individual labs’ GPU clusters while addressing some of the scalability challenges facing API-first deployment models.

Many customers will still prefer managed APIs rather than operating massive models themselves. They are willing to pay for deployment, customization, fine-tuning and long-term support.

Products like Kimi can generate subscription revenue. Unlike earlier Chinese models that competed aggressively on price, Kimi K3 is priced much closers to OpenAI and Anthropic’s frontier offerings, signaling confidence that developers will pay for performance. DeepSeek’s flagship API is about 17X cheaper than Kimi K3 for output tokens.

Cloud providers such as Alibaba, Tencent and Huawei can benefit from increased demand for computing infrastructure.

But the larger opportunity is the ecosystem itself.

If thousands of developers build applications on your model, your architecture, tooling and developer community become the foundation others build upon.

The model becomes the starting point, not the final product.

Rather than monetizing exclusive access to intelligence, the strategy is to monetize the ecosystem that grows around it.

Openness is not a blank check. Commercial licensing terms still matter, and deploying models at this scale requires enormous computing resources. Success on a coding benchmark is evidence of strong capability, but not definitive proof of overall frontier leadership.

Why It Matters

The AI race is increasingly taking shape along two different paths.

Much of the US frontier AI industry remains focused on building the most capable proprietary models and monetizing access through closed APIs.

Many of China’s leading AI labs appear to be betting that openness can accelerate adoption, expand developer ecosystems and create more long-term value than keeping frontier models behind closed doors.

In a recent internal letter, Tang Jie, founder of Zhipu(Z.ai), argued that his company was “not pursuing short-term application monetization, but directly targeting the next high ground of artificial general intelligence (AGI).”

He added: “If the destination is AGI, then short-term interests or industry trends are merely scenery along the road to the end-game.”

The connection is notable: Tang mentored Moonshot founder Yang Zhilin during Yang’s time at Tsinghua University.

That does not mean Chinese AI labs are not competing with each other — they are. But they are also contributing to an ecosystem where each major open release raises the capabilities available to everyone building on top of it.

For everyday users, Kimi K3 means a more capable AI assistant.

For the industry, it signals something larger.

The first phase of AI competition was about building the smartest models.

The next phase may be about building the largest ecosystems around them.

Kimi K3 suggests China’s open approach is not simply an alternative to the closed-model strategy. It represents a different theory of competition:

Do not just build the best model. Build the ecosystem everyone else wants to build on.

Twenty-nine countries signed a treaty in Shanghai yesterday. Not one of them got a single GPU out of it.

The establishment of the World AI Cooperation Organization (WAICO) marks an ambitious effort to broaden participation in the AI era, particularly for countries that have yet to develop significant AI capacity of their own. Members join as founding participants rather than observers—a notable shift for economies that have had relatively limited influence in AI governance frameworks led by the Group of Seven and the Paris-based Organisation for Economic Co-operation and Development.

But a seat at the table was never the hard part of this problem.

A permanent institution can give countries a voice in AI governance. It cannot generate a single token of inference. Access ultimately depends on physical infrastructure: power, cooling, land, fiber-optic networks, and data centers. Without sufficient computing capacity, AI access remains an aspiration rather than a capability.

If WAICO is the institutional framework, China’s expanding compute infrastructure could become the physical foundation that gives it practical meaning.

Over the past several years, China has pursued an infrastructure strategy designed to expand AI computing capacity while reducing many of the constraints that are increasingly slowing infrastructure development elsewhere.

Under the Eastern Data, Western Computing (东数西算) initiative, much of China’s new computing capacity is being built in sparsely populated western provinces, where land, renewable energy resources, and transmission capacity are more readily available than in densely populated coastal cities.

Electricity pricing is also managed differently. Rather than relying on utility rate proceedings that determine how infrastructure costs are allocated among customers, China’s electricity prices are administratively managed, and local governments have in numerous cases subsidized data center electricity costs to attract AI investment.

China is also experimenting with new engineering approaches to reduce infrastructure bottlenecks.

Hainan’s underwater data center at Lingshui expanded this June with a new module capable of processing roughly 7,000 DeepSeek queries per second. In May, Shanghai’s Lingang underwater data center began commercial operations using offshore wind power and seawater cooling to reduce freshwater consumption, cooling requirements, and pressure on land-based infrastructure.

Whether underwater data centers become mainstream remains uncertain. But they illustrate a broader point: China is attempting to engineer around infrastructure constraints rather than simply regulate them.

The contrast with parts of the United States is becoming increasingly apparent.

This week, New York became the first US state to impose a statewide moratorium on new hyperscale data centers while it develops a regulatory framework for future projects. The move reflects growing concern over electricity costs, grid upgrades, water consumption, and community impacts. It also follows broader warnings from Harvard Law School’s Electricity Law Initiative, whose researchers argued in a March 2025 paper that existing utility rate structures and confidential contracts can force residential ratepayers to shoulder part of the cost of powering large data centers. Across the United States, hundreds of local moratoria and state legislative proposals now illustrate how AI infrastructure has become a political as well as an engineering challenge.

The immediate questions being asked are different.

Neither approach is inherently right or wrong. One prioritizes protecting consumers from infrastructure costs. The other prioritizes expanding infrastructure while attempting to engineer around resource constraints.

This matters because AI is increasingly becoming a deployment business rather than simply a model-development business. The industry is gradually shifting from rewarding the most capable models to rewarding those that can deliver inference at scale—and at low cost.

History suggests that countries export technology only after they have first built abundant domestic capacity. AI is unlikely to be different.

That does not mean WAICO can guarantee AI access for its founding members. Domestic computing abundance is a precondition for broader access—not its fulfillment. Nothing in the founding agreement obligates members to share computing capacity, subsidize inference, or provide preferential access to AI infrastructure. Commercial incentives, export controls, connectivity, and geopolitics will all shape how much of China’s expanding compute base ultimately reaches other countries.

That is why the real significance of WAICO lies less in the agreement itself than in the infrastructure that may eventually support it.

If China succeeds in building abundant, low-cost AI infrastructure, it will eventually have something meaningful to offer beyond governance principles: compute.

Whether that compute ultimately reaches WAICO’s members—or simply fuels China’s own AI deployment race—will determine whether July 16, 2026 is remembered as the beginning of a new model for AI cooperation, or simply the day another international organization was established.

However, instead of walking away from such platforms, I asked a different question: what if I could keep YouTube and evict the algorithm? This year, I did something that would have been unthinkable for me two years ago–I built my own tools to do exactly that. Not just one tool, but a small stack of them.

I am not a web developer. My coding background is basic Python and R, not JavaScript, so the extension architecture that Chrome and Firefox expect was foreign to me. But I described what I wanted to an AI system in plain language, the way I would explain a problem to a colleague, and within an afternoon I had a working tool. Then I built another.

My browser now runs a Facebook extension that hides Reels and cuts down on suggested content from accounts I don’t follow. A YouTube extension now restricts my homepage to channels I choose and strips out unwanted shorts entirely. I also built a small calculator because I am studying for an exam and wanted to practice on something that looks like what I will actually use on test day. None of this required a computer science degree. It required a clear idea, patience, and a conversation.

_{The extensions I built, alongside the ones I did not: Facebook Clean Feed, a study calculator, and a YouTube homepage filter, sitting next to Chrome’s own tools.}

This practice now has a name: vibe coding–a term coined by AI researcher Andrej Karpathy and named 2025 Word of the Year by Collins English Dictionary. You describe what you want in natural language, and an AI system writes the code. What used to sit between an idea and a working product, teams of developers, weeks of planning and a budget now collapses into a conversation. Before, my relationship with these platforms was one-dimensional. They decided what I saw, and I could complain or quit, but I couldn’t edit the terms. Vibe coding gave me a third option. I rewrote the terms myself.

_{My YouTube homepage now: vocabulary lessons, a journalism report launch and a lecture on the brain – nothing I didn’t choose.}

Before I understood any of this, I had already spent plenty on subscriptions I could not fully justify, and just as often avoided apps I actually wanted because the price tag felt wrong for what they did. MacroFactor was one of the ones I paid for anyway, before I knew a tool like it was something I could build myself. I have not bothered to cancel it. But a separate workout app that asked for US$30 a year just to let me create more than a couple of routines, I refused, and built my own instead.

Instead, I asked an AI system to help me build my own training log. It is just an HTML file, connected to the internet through Netlify. That’s it! It is not as polished as the app I refused to pay for, but it does exactly what I need. The difference between the two examples is timing: one subscription outlived my ignorance; the other never got the chance.

None of this was effortless, and I do not want to overstate how smooth it was. Getting an AI system to build something correctly usually takes many rounds of back and forth, not one clean prompt that produces a finished tool. The skill is not in writing code. It is reading carefully what the AI proposes, catching what it misunderstood and telling it precisely what to fix next. People who struggle with vibe coding are usually struggling with that conversation, not with programming itself. The tools themselves are not expensive to run. None of mine needed a frontier model; a mid-tier system handled all of them. The only recurring cost is the AI subscription itself, about US$20 a month.

It would be easy to stop here, on a happy note, and many articles on vibe coding do just that. The individual is empowered. Power shifts from the platform to the person. Anyone can build now. I believed a version of that while I was building my tools. But I keep returning to that phrase “anyone can build now” because when I look closely at what building actually required of me, the word “anyone” starts to fragment.

Consider the US$20 a month. That’s about US$240 a year, and it has to be paid in hard currency to a foreign company. Across South Asia, where I am from, the annual cost that supposedly empowers anyone can amount to a month or more of a typical person’s salary. For me it is a manageable line in my budget. For a great many others, the entry ticket alone is the obstacle before a single prompt is written.

Then consider language. The whole premise rests on the ability to describe what you want in plain language, but the models do not understand every language equally. One of the AI systems that started this boom was trained on text that was about 93 English, leaving every other language to share the remaining 7 percent.

The performance gap follows that imbalance closely. So, the better you can write and understand English, the better your outcomes. Researchers describe this simply as a “rich-get-richer effect.” My prompts work well partly because I write them in fluent English. Someone describing the identical tool in a lower-resource language would get a weaker result for the same request.

So the tool sold as the great equalizer demands exactly the two things: hard currency and English fluency. It does dissolve the hierarchy between those who can code and those who cannot. I am, awkwardly, proof of both halves at once. I am from South Asia, the region where the democratization story is supposedly aimed, but I built these tools from a well-resourced university in Shanghai, paying in dollars and thinking in English.

None of this means the door is closed to everyone but the privileged. Here in China, the picture already looks more crowded than the English-first story suggests.

Bao Linyi, a classmate at Shanghai Jiao Tong University, built a full content management system during an internship, front end and back end, using a generative AI model to handle product documentation and generate demos–the kind of build a small engineering team would once have owned. Alongside foreign tools, home-grown Chinese models are maturing fast and lowering the cost further, which changes the calculation for hundreds of millions of people. The barrier is not the concept of vibe coding; it is whether a capable model exists within your language and budget limitations.

But, she added, “The shift places a clear demand on the user’s own judgment and capacity to express what they want. The tool is only as good as the problem you can articulate, and the result is only as sound as your ability to tell whether it is right.”

That last point stayed with me. Even for those who can pay and who can prompt, the tool hands back exactly the quality of thinking input. It does not supply judgment. My own tools are low stakes; if my workout log miscounts a session, nothing breaks. But a tool built in the same casual way to decide something real about someone’s health, finances or social life deserves the care a trained engineer can give it. So the divide this technology opens isn’t only about who can afford the subscription or who writes fluent English; rather it is also about who has been trained to question a result rather than accept it.

I open YouTube now and my homepage holds what I actually chose. That much is real, and I did build it myself, one conversation with an AI at a time. But I would rather be honest about the size of the process than sell it. A tool that lets me reshape my own digital life is a genuine gift. A tool that lets anyone do the same is a different and larger ask.

_{Neuracle's NEO implantable brain-computer interface system for hand motor function restoration}

More than a decade ago, Dr Wu Zehan was asking patients to raise their hands 20 times, recording their brain signals and copying the data into a computer for calculation.

The machine was learning slowly. Today, the response can happen within hundreds of milliseconds.

Faster decoding now allows brain signals to be translated almost in real time, enabling a robotic arm, a pneumatic glove or a cursor to respond to a patient’s intention. That speed is pushing China’s brain-computer interface industry into a tougher phase.

After years of futuristic promises, the next test is no longer only whether machines can read or write brain signals. Rather, it is whether the technologies can be approved, tested in hospitals, priced, reimbursed in medical plans and made accessible to patients.

Shanghai is trying to build that missing bridge.

At the Shanghai Brain-Computer Interface Future Industry Cluster, the point is not the exhibition hall itself. It is the attempt to bring start-ups closer to hospitals, testing agencies, incubators, investors and regulators they will need long before a device can reach patients.

The cluster has more than 50 registered companies doing research and development in flexible electrodes, brain-interface chips, visual reconstruction, ultrasound-based interfaces and rehabilitation devices. It also brings together the Sinan BCI Super Incubator, Huashan Hospital’s clinical translation platform, Fudan University’s technology transfer center and a neural data platform.

For early-stage companies in the industry, the interconnections matter more than the office space. A neural interface cannot be commercialized like an app. It has to survive animal studies, product testing, clinical trials, regulatory review, hospital procurement systems and, eventually, payment channels.

For companies, the question is whether Shanghai’s cluster can shorten the hardest stretch: turning a promising neural interface into something doctors will trust, regulators will approve, hospitals will buy and patients can afford.

Neuracle offers an early test case. The Shanghai-based company in March has received approval for the world’s first marketed Class III invasive brain-computer interface medical device. Its system is designed for quadriplegic patients with cervical spinal cord injuries, using decoded brain signals to help them regain hand-grasp function through a pneumatic glove.

But approval is only the first gate.

For Neuracle, that means turning a regulated implant into something hospitals can actually offer. The company is working with medical insurance authorities, disability federations and foundations to improve patient access, Wang said. It has also secured a billing code, allowing hospitals to charge for the service, though broader reimbursement channels still need to be negotiated.

The product has not yet been widely rolled out. Wang said the first hospital deployment has not started, partly because production still takes time. The device’s price is also expected to become clearer only after it appears on Shanghai’s hospital procurement platform.

That makes Neuracle’s milestone both symbolic and practical. It shows that invasive brain-computer interface technology can cross the regulatory line. It also shows what comes next: manufacturing, hospital listing, service pricing, reimbursement and the slow work of making an implantable technology affordable enough for real patients.

If Neuracle can boast a regulatory breakthrough, Shaonao Tech demonstrates a more immediate path into hospitals.

The company focuses on non-invasive applications for rehabilitation, using motor-imagery decoding and functional electrical stimulation to help stroke patients recover upper-limb functions. Its core product, a non-implant brain-signal device, received a Class II medical device registration certificate in Shanghai in late 2024.

Unlike systems that require surgical implants, Shaonao’s product already has visible hospital pricing.

Marketing director Chen Xuanbo said Beijing Tiantan Hospital was its first “customer,” with a tender price of 1.32 million yuan (US$194,000). Another recent winning bid was from Wuhan Union Hospital, at 1.04 million yuan.

The patient-side pricing is also taking shape. Hospitals in Beijing and the provinces of Hubei, Zhejiang and Jiangsu charge between 960-990 yuan for non-invasive brain-interface services. Shanghai is still under independent pricing, with different hospitals charging from several hundred yuan to more than 900 yuan.

That puts non-invasive brain-computer interface closer to routine clinical use over more costly implant systems. It also exposes a different commercial test: whether hospitals will buy the equipment, whether therapists can integrate it into rehabilitation work and whether enough patients will use it frequently enough to support the business model.

Mindtrix, another company in the industry, is advancing visual restoration, one of the most closely watched frontiers in the sector. Rather than reading neural activity, the system writes information directly into the brain. A camera captures the outside world, converts images into stimulation signals, and transmits them to electrodes implanted​ in the visual cortex, bypassing damaged eyes or optic nerves.

_{A brain-controlled wheelchair on display at the Shanghai Brain-Computer Interface Future Industry Cluster. Non-invasive BCI companies are expanding beyond rehabilitation into assistive mobility. — Wing Tan}

The immediate objective​ is not to replicate natural sight. Chen described it as “functional vision” - enough visual information for patients to discern​ rough shapes, avoid obstacles, navigate daily spaces and regain a basic level of social interaction.

He cautioned against overestimating the technology. Reconstructed images are closer to low-resolution black, white and grey patterns than natural vision. While a healthy human retina contains roughly 100 million photoreceptors, Mindtrix’s system targets approximately 1,000 stimulation channels.

Mindtrix is already testing the system with Huashan Hospital in Shanghai and other major health facilities in Beijing and Wuhan, Chen said. But a commercial product remains years away. The company expects to begin clinical trials around 2028 and hopes to obtain approval by 2031. A future procedure, including surgery and rehabilitation, could cost about 300,000 yuan, he estimated.

For all its futuristic appeal, visual reconstruction remains a medical-device story. Clinical evidence, safety, durability, reimbursement and patient selection will determine whether it moves beyond the demonstration phase.

For Huashan Hosital’s Dr Wu, the physician’s role is not to sell the promise of brain-computer interface, but to narrow the gap between what companies can build and what patients can realistically gain.

Patients need to know what risks they are taking, what benefits are likely, what may require a year or two of training, and what may never happen, he said.

That may be the clearest test for China’s brain-computer interface industry as it moves closer to the market. Not whether a signal can move faster, but whether the promise can be translated into evidence, cost and risk that patients can understand - and accept.