Frontier AI risks are now increasingly urgent for both the United States and China. The good news is that both sides know they need to talk. And arguably, they want to talk. But there is a more basic problem before those talks can begin: the mismatch between and within each political system. It’s not just that the two sides speak different languages. Neither side has developed a common language of its own! Not just the Babel problem, but the Babel-squared problem, if you will.

I think this common language / coordination challenge within each system is equally hard, if not more so, than the coordination challenge between the U.S. and China.

As my colleagues and reported from their on-the-ground experience at the Shanghai WAIC, this is a fast-evolving, still nascent, and highly fluid field in China. But the thing is, the people who understand AI technology, the people who control the resources, and the people with the political authority to act are often not the same people.

Which makes it all the more important to understand where the expertise sits and how those experts are politically positioned within both systems, to make sense of how Beijing is likely to tackle frontier AI risks.

So the CCA Substack will do exactly that, i.e., an institutional map / who’s who in China’s frontier AI governance.

More details will be posted there! But here I’ll do a trailer version of the players.

Subscribe now

• The Chinese system has its traditional cybersecurity apparatus. Unsurprisingly, at the top of China’s party-state security architecture is Cai Qi, who oversees the political and ideological machinery around information security. Below Cai, authority over cybersecurity splits into two arms: one runs through the Cyberspace Administration of China (CAC), which manages disclosures and incident responses. This is perhaps the most technically sophisticated cohort in the Chinese civilian bureaucracy when it comes to addressing AI risks. The other runs through Wang Xiaohong and the Ministry of Public Security (MPS), with a dedicated research arm, the MPS Third Research Institute. This apparatus is responsible for setting up technical standards and charting rules for compliance, audits, and procurement requirements for government agencies and critical infrastructure operators.

• The traditional cybersecurity ecosystem also has a sprawling network of commercial vendors, including 360, Topsec, etc. They serve governments and SOEs effectively as government contractors. Incidentally, the commercial contractor economy has not been doing well financially in recent years, with governments tightening their belts and costs skyrocketing (compute is expensive!). And as expected, they don’t have either the financial resources, the incentives, or the technical capacity needed to address frontier risks.

• What’s more notable is an emerging ecosystem of state-affiliated research institutes and frontier AI labs. This network connects macroeconomic planning, top-level science-and-technology policymaking, elite universities, and state-backed research platforms, all under a broader policy system overseen by Executive Vice Premier Ding Xuexiang.

• Ding Xuexiang’s portfolio on science and technology, in addition to traditional macroeconomic planning, sits at the nexus of the Ministry of Science and Technology (MOST) and the National Development and Reform Commission (NDRC). This pipeline is revealing because it creates a more direct institutional channel and political infrastructure connecting state capital, basic research, and technical architectures, especially the national laboratories.

• Beneath the central level, regional research hubs are especially important to understand.

  • Again, the fuller version will be in the CCA Substack, but one important person to note is Huang Ru currently in Beijing.

    • Her career trajectory cuts across several major institutions, including those in Nanjing. As people familiar with the Chinese system know, Nanjing is famous for having a high-quality academic cluster, with Southeast University and the Nanjing University of Posts and Telecommunications as two leading examples, especially in microelectronics.

    • Then, at Peking University, Huang became a leading scholar in microelectronics and served as the founding director of its Institute of Artificial Intelligence. She returned to Nanjing in 2022 as president of Southeast University and subsequently joined the governing council of Purple Mountain Laboratories, a major state-backed research platform in Nanjing focused on cybersecurity. (For the Xi watchers out there: Xi Jinping visited the laboratory in July 2023, while Huang was still president of Southeast University.) Notably, in 2025, Purple Mountain’s Endogenous Security Research Center served as a blue-team technical adviser during a Hong Kong government cybersecurity exercise.

    • More recently, BAAI, which incubated Z.ai, developer of the GLM model series, is getting more attention. Slightly less well known is that, in August 2026, BAAI published an evaluation benchmark, SPIKE-Bench, with a scholar from Peking University’s Institute for Artificial Intelligence.

  • Shanghai is a separate top hub for frontier risk research anchored by the Shanghai AI Laboratory (SAIL). SAIL has political backing through the Shanghai municipal government under Party Secretary Chen Jining, a Politburo member, who has high ambitions and potential for something bigger. (SAIL deserves a separate Substack, so I will keep it brief here.)

• Hint: as you can imagine, there is a fierce behind-the-scenes turf war going on between Beijing and Shanghai! The stars in both Beijing and Shanghai eyeing the next Party Congress have every incentive to show results on AI, which is now a top political priority…

• This is a key reason driving these two most vibrant innovation ecosystems. Here bureaucratic KPIs actually act in similar ways to the profit incentives that drive private entrepreneurs to innovate (and to involute!). Even more reason to understand how Chinese bureaucrats’ KPIs and incentives work. But that’s another Substack for another day…

The TL;DR version:

The story begins with an almost forgotten collaboration, one powered by the U.S. motor giant Ford, that began in the 1990s. The partnership never produced a commercial breakthrough. Ford eventually abandoned methanol, while China moved on to other technologies. Nearly three decades later, methanol vehicles have reappeared in China, where Geely now operates one of the world’s largest fleets of methanol-powered commercial vehicles.

To go back to the beginning, in 1995, Ford began working with China’s State Science and Technology Commission, the predecessor of today’s Ministry of Science and Technology, on a flexible-fuel vehicle program. The broader initiative eventually involved the U.S. Energy Department, the U.S. National Science Foundation, the Massachusetts Institute of Technology, Tsinghua University, Shanxi province, and several industrial partners.

But the two countries ultimately followed very different paths! In the United States, the technology largely disappeared. In China, it survived—because provincial governments and entrepreneurial firms saw their own reasons to keep it alive. (More details in the full article…) Geely emerged as the most important corporate backer, while provincial governments helped a lot in creating a market.

As it turned out, China’s approach was also more technologically inclusive than is often assumed. The Chinese government certainly prioritized battery EVs. Yet it rarely relied on a single technological pathway. Rather than forcing an immediate technological winner, policymakers allowed multiple pathways—including battery EVs, fuel cells, plug-in hybrids, and methanol—to develop simultaneously under different regional conditions. This is the critical industrial policy lesson we discussed in the article, which I think people haven’t yet paid enough attention to.

This brings us to July this year when Ford and Geely announced plans to establish a Europe-focused joint venture at Ford’s Valencia plant in Spain. Quite a striking reversal of roles! The new Ford-Geely venture suggests that the direction of technological advantage may have changed: In the 1990s, China deepened its understanding of methanol vehicles by building on Ford’s decades of research into alternative fuels. Today, Ford stands to benefit from Geely’s sustained investment in electrification and advanced powertrain technologies over the past two decades…

Full article here for interested readers. And hats off to Fengming, who first flagged this history to me—and who, during fieldwork at one of Geely’s principal methanol truck facilities in Sichuan, saw how far the technology had advanced in both engineering and production scale!

https://foreignpolicy.com/2026/08/03/ford-china-methanol-energy-electric-vehicles/

Finally, a few additional reflections on industrial policy then and now that that ended up on the cutting-room floor, which I’m including here for Substack readers. Still preliminary, but this is something I’ve been asking myself: given that replicating China’s industrial policy is both infeasible and to be honest, perhaps not the right way to think about these challenges today, is there a smarter, modern way to achieve what industrial policy did for China—especially when starting from a second-mover disadvantage? Here’s how we put it in the earlier draft:

This reversal also points to a broader lesson for industrial policy. The United States and Europe may well be unable to reproduce the industrial ecosystems that China has built through decades of policy investment. Carefully structured technology partnerships can offer another path, allowing U.S. and European firms to draw on capabilities created by Chinese industrial policy while fast-tracking technological learning at home.

Crucial question is how such partnerships are designed. Political stakeholders demand partnerships tied to local production, workforce development, domestic suppliers, and shared research, rather than straightforward import-and-assembly arrangements, so that the gradual absorption of technological capabilities can strengthen the host country’s industrial competitiveness. This is easier said than done, and the political sensitivity is high. But at least in principle, it represents a form of industrial policy through negotiated integration rather than an attempt to reinvent the entire wheel from scratch.

The immediate benefits of the Ford-Geely partnership in Spain are readily apparent: higher utilization of Ford’s existing manufacturing assets, lower costs and reduced execution risks for Geely compared with building a greenfield plant in Europe, and positive economic spillovers for Spain, Europe’s second-largest automobile producer. The partnership’s greater significance, however, lies in its long-term strategic potential. At its heart is a new multi-energy family crossover scheduled for launch in 2028, which Ford describes a being “designed by Ford and jointly developed with Geely”.

More importantly, the commitment of co-development creates opportunities for genuine technological exchange. Thanks to China’s more inclusive technological approach, Geely has a full array of drivetrain technologies available, from advanced internal combustion engines, hybrid and plug-in hybrid systems, battery electric vehicle, or even alternative fuels like methanol. The partnership allows Geely’s diversified portfolio of technology to eventually empower an American automaker’s product development and manufacturing, with technological know-hows transferring through day-to-day collaboration of engineers, technicians, and manufacturing workers. In the meantime, China’s cost-effective electrical/electronic (E/E) architectures, a core strength of Chinese EVs, can also be used in American-brand vehicles after rigorous test, verification, and validation from Ford’s engineers.

Critics are correct to point out that such partnerships inevitably carry risks involving technological dependence, and concerns about national security are not trivial. Clear guardrails and carefully negotiated terms are mandatory for any such arrangement. Yet the historical episode illustrates the enormous potential of technological collaboration between the United States and China when complementary strengths and interests align. The new Ford-Geely venture suggests that the direction of technological advantage may have changed: in the 1990s, China deepened its understanding of methanol vehicles by building on Ford’s decades of research into alternative fuels. Today, Ford stands to benefit from Geely’s sustained investment in electrification and advanced powertrain technologies over the past two decades. But the underlying logic of collaboration remains relevant: each side can achieve outcomes that would be much harder, slower, and more expensive to pursue alone.

As the title suggests, the clearest takeaway from my CCA colleagues and friends , , Ran Guo, and ’s on-the-ground observations at this year’s Shanghai WAIC is that China is turning export constraints into an organizing principle for its AI ecosystem. There is no substitute for seeing these shifts unfold in real time. The full issue will run on our Substack in due course, but here are a few reflections.

Export controls are imposing real costs. Chinese alternatives remain behind, and Nvidia chips are still preferred for many frontier training tasks. Chinese labs are diverting enormous engineering effort toward making domestic hardware work across comparable use cases.

What is striking, however, is the scale of the system-level rewiring now taking place—not just across the technology stack, but in the institutions surrounding it, from capital markets to business models. These are no longer adaptations in the passive sense. They are becoming increasingly institutionalized, coordinated, and forward-looking.

Innovating around the chip

A lot of the most interesting developments involve innovating around the chip, not necessarily in the chip, though China is working hard on the latter too.

Paul, for example, wrote about the shift toward full vertical-stack integration across hardware and software, a trend that has deepened over the past few years. Huawei’s Atlas 950 strategy is designed to compensate for the lower performance of individual accelerators by connecting thousands of them through advanced interconnects, unified memory, and system-level engineering.

Chinese model developers are also actively optimizing models for domestic hardware clusters. What emerges is a process of mutual adaptation: hardware optimized for domestic models, domestic models optimized for domestic hardware, and coordination extending across company boundaries.

Alongside this coordination—which helps the system organize around constraints and buy time for catch-up—Chinese firms are gradually moving inference, fine-tuning, reinforcement learning, overflow capacity, and secondary training onto domestic systems.

As others have reported, access to, e.g., Nvidia chips often comes with a phase-down agreement between firms and Beijing, such as replacing 20% with domestic chips within one year, 50% by the second year, and so on.

Compute scarcity becomes an institutional model

Beyond building larger clusters, porting selected workloads onto domestic chips, and improving algorithmic efficiency—as our colleagues observed in real time; read the Substack when it is out for the full picture!—it is critical to see how export-control-induced compute scarcity is becoming a business and institutional model, not just a technological one.

Business strategies and domestic capital-market structures are reorganizing around it—and around the constraints of China’s domestic capital markets themselves.

Firms are training flagship models offshore when possible and generating revenue from inference workloads abroad. Domestic capital is also reorganizing around China’s AI ambitions, including through state investment and domestic listings that finance AI infrastructure.

China’s capital markets, corporate structures, research priorities, and relationship between government and industry are therefore all increasingly being reoriented around this constraint.

I think it is safe to conclude that the Chinese AI ecosystem is being designed to survive—and continue innovating—without access to frontier U.S. chips.

Openness as strategy

As one of our colleagues pointed out, China’s open-source strategy, both historically and today, also emerged in response to restrictions. The push gained real momentum after the 2018 ZTE shock. It was spearheaded by companies such as Huawei and sought to combine domestic control over core technologies with continued access to global innovation networks, while giving China a greater role, voice, and presence in global technology ecosystems.

Openness has now become an important strategy beyond the domestic market, positioning China as a major exporter of affordable, reliable AI capabilities abroad. As others have noted, Xi’s WAIC speech touting China’s open approach is very much the culmination of this longer evolution.

The initial domestic strategy now serves a global objective as well: lowering the cost of adoption and expanding access to AI at home, leveling the playing field for up-and-coming firms and developers, and making China’s offerings especially attractive across the Global South.

A distributed safety architecture

Nor is this transformation limited to markets and institutions. China’s approach to governance and safety is also evolving quickly, and in ways distinct from the U.S. model. That, I think, will make bilateral coordination and dialogue on AI safety both more urgent and more difficult.

Within China’s own AI safety system, coordination remains a challenge. The landscape is still fluid, and my colleagues see Chinese firms approaching safety in very different ways. Some treat it primarily as a problem of compliance, information security, and content control to be managed.

But it is increasingly outdated to say that “China has no AI safety institutions” or no real urgency around safety concerns. At a higher level, China is developing a distributed system.

Our colleagues offer detailed analyses of the role Shanghai AI Laboratory is playing in building technical evaluation capacity, including through OpenCompass, which is creating shared evaluation infrastructure. Institutes at Tsinghua and in Beijing are focusing more on defining risk frameworks and developing international standards. My colleagues see an emerging division of labor between Shanghai and Beijing: Shanghai is building much of the technical machinery for model evaluation, while Beijing-based institutions are defining which risks matter and how they should be governed.

CAICT, TC260, laboratories, and industry actors all perform overlapping functions.

Technical capacity is advancing quickly on multiple fronts and becoming increasingly substantive. But obvious questions remain around independence and authority. As my colleague flagged, Shanghai AI Laboratory is simultaneously a model developer, a government-supported institution, a commercial actor, and a potential evaluator. Whether it has the political authority to perform that role is also crucial. There is, for example, little evidence that it can compel private model developers to provide pre-deployment access to their models. Nor is it yet clear how its evaluations shape regulation.

The chip war is working—but what is the goal?

So I think we need to be honest about two things being true at once. The chip war is working. But that does not mean America is winning.

Perhaps Washington needs a more clearly articulated answer to a basic question: What is the goal of its AI strategy?

When a constraint becomes the organizing principle of an entire system, we need to think beyond its first-order effects. The second- and third-order consequences may prove far more consequential than initially imagined.