Q&A with Hao Chen, co-author of “China’s Diffusion-Forward AI Strategy: Chatbots, Robots, and Political Economic Possibilities”

Q: Your central argument is that the dominant U.S. framing of the “AI race” has obscured what is actually being contested. If diffusion rather than frontier capability is the decisive variable, what does that imply for the current U.S. toolkit, i.e. export controls on advanced chips, the AI Diffusion Rule, restrictions on Chinese model access in third markets? Are these instruments fighting the wrong war, or do they remain consequential under your framing? And is there a positive industrial agenda the U.S. could pursue that takes diffusion seriously without simply mimicking the Chinese model?

A: On one hand, the current U.S. policy toolkit has slowed the innovation pace of leading Chinese large language model (LLM) labs, but on the other hand, it may harm U.S. national interests in the long run. Regarding the war metaphor, I don’t think the U.S. toolkit and policies are fighting the wrong war, but I do think there is no single war between the U.S. and China. Model performance and AGI are not the end game – diffusion is, because it determines how broadly AI gets embedded into the economy and society. If we want to use the war metaphor, it is really multiple wars at once, and however important, the race on LLMs is just one of them.

Crucially, superiority in LLMs does not automatically translate to superiority in AI’s deployment in the physical world. As my recent paper co-authored with Meg Rithmire shows, diffusion at scale requires an entire ecosystem—including central guidance cascading through local governments, the investor-state mobilizing capital into application-layer firms, anchor manufacturers willing to absorb humanoid robots and industrial AI agents on the factory floor, and a domestic market deep enough to sustain rapid commercialization. That ecosystem does not emerge naturally from frontier model leadership.

This points to a serious side effect of export controls that may, in the long run, harm U.S. economic interests. Based on my other research, the biggest one is that U.S. export controls have created an acute “sense of urgency” in China — a clear signal to Chinese competitors that they have either lost the game or must build their own stack. They have unsurprisingly chosen the latter. Before the 2022 controls, for example, Chinese domestic AI accelerator suppliers accounted for less than 10% of the Chinese AI chip market; that share has climbed to nearly 50% in just three years. DeepSeek V4 is now optimized for Huawei Ascend and was not trained on NVIDIA hardware before release— a striking departure from earlier generations.

The forward-looking question is this: If DeepSeek and Huawei Ascend co-build an AI-and-applications ecosystem in China, and then export that ecosystem to third markets – particularly across the Global South, where price sensitivity and industrial-upgrading needs align well with what China offers – what happens to U.S. global influence? The ultimate question is: Can the U.S. still maintain global leadership in AI in the long run?

As for a positive U.S. industrial agenda that takes diffusion seriously without mimicking China: I don’t think the answer is to replicate China’s investor-state model. The institutional preconditions existing in China simply do not exist in the U.S. But the U.S. could take diffusion seriously on its own terms by investing in the deployment side rather than only the supply side: workforce retraining for AI-augmented work, procurement policies that pull AI into manufacturing and public services, support for embodied AI and industrial robotics startups, and targeted partnerships with allied manufacturing economies. The decisive contest may not be who reaches AGI first, but which political economy can more rapidly diffuse AI into productive activity. The U.S. still has time to take that contest seriously.

Q: You situate AI within a longer arc of investor-state activism: solar, EVs, batteries, semiconductors. Each of those sectors has produced global Chinese dominance alongside chronic overcapacity, deflation, and escalating trade conflict. What, if anything, is distinctive about embodied AI and industrial robotics that might break or repeat that pattern? And at what point does the investor-state model’s success, measured in production capacity and export volumes, become the constraint that forces a course correction, whether through fiscal stress, trade retaliation, or domestic political backlash?

A: What distinguishes humanoid industrial robotics and embodied AI more generally is that the sector has not yet proliferated, largely because of model constraints. From a technical standpoint, humanoid industrial robotics is harder than EVs or solar on both the hardware and software sides. On hardware, precision control remains difficult, and the “hand” of a humanoid robot still struggles to replicate the dexterity of a human hand. Even more fundamentally, humanoid industrial robotics are constrained by the underlying models. Today’s frontier models are LLMs, trained on text, but operating on a production line in a complex factory environment requires models that go well beyond language. A robot needs to measure distance precisely, anticipate the direction and speed of moving objects, and preempt potential hazards in real time. That requires models that speak to the physical world, not just to text. The UBTECH Robotics case we documented in the paper has made some progress but still needs much more improvement.

If these technical bottlenecks are resolved over the next five to ten years, China’s advantages will lie in its deep supply chains and aggressive price competition. As we have shown in our paper, the fierce domestic competition that emerged in EVs and solar is already taking shape in humanoid robotics: UBTECH, Unitree, Leju, AgiBot, and Galaxy General are all racing to scale, with several already crossing the 1,000-unit production threshold and securing contracts with anchor manufacturers like BYD, Geely, NIO, and CATL. If this trajectory continues, industrial robotics could very plausibly repeat the EV and solar pattern: rapid commercialization, falling prices, overcapacity, and an export surge that triggers trade tensions abroad.

On the other hand, whether that pattern actually materializes depends heavily on the degree of mobilization from the center. China’s political-economic system has a strong mobilization feature embedded in it. Once a technology mission becomes a national political campaign, it mobilizes the whole government and societal resources on a scale that can lead to reckless investment and overcapacity. Course correction, when it comes, typically comes from the center as well.

So, the more honest answer to your second question is this: China’s investor-state model has built-in mechanisms for both acceleration and correction, but neither operates through the price signals or political channels that economists and policymakers in market democracies are accustomed to reading. The constraint that forces a course correction is unlikely to be fiscal stress alone, or trade retaliation alone, or even domestic backlash in the conventional sense. It is more likely to be a political signal from the top, and the timing, severity and credibility of that signal are difficult to predict from the outside.

Q: You note that the AI+ policy architecture, and your conversations with local officials, are largely silent on labor displacement, even as urban youth unemployment runs at 13–20% and “purple-collar” mismatches deepen. How do you read that silence? Is it confidence that demographic decline will absorb displacement, strategic neglect, an artifact of how the bureaucracy compartmentalizes industrial and labor policy, or something else? And what would it take politically or economically to force the issue onto the agenda before it manifests as instability?

A: It is a complicated issue, and my reading rests on two points.

First, in my view, there is a deep cultural and ideological backdrop in China that frames technology as inherently progressive. This mindset can be traced back to the May Fourth Movement of 1919, with its embrace of “Mr. Science” and “Mr. Democracy” as twin pillars of national modernization. That instinct continues to shape how the Chinese state and Chinese society read technological change today. Officials are predisposed to focus on the upside – for example, the new industries AI will create, the productivity gains it will unlock, the new categories of jobs that will emerge – rather than the displacement it will cause along the way.

Second, and more institutionally specific, the local political system in China structurally underweights labor displacement concerns. The fact is that China’s AI development is now heavily concentrated in a handful of developed cities, such as Beijing, Shanghai, Hangzhou, Shenzhen, Guangzhou, Suzhou, Wuxi and Hefei, which together account for the overwhelming majority of AI firms and capital deployment. Another fact is that these leading cities rely heavily on migrant workers in their manufacturing and service sectors, but those workers typically do not hold local hukou (household registration, or local residency-ship). This matters enormously. In the U.S., political representation tracks residency, which in turn links to voting. In China, residency rights and public services flow from hukou registration, not from where one actually lives and works. Local officials in destination cities have, in theory, no formal political responsibility for the migrant workforce that staffs their factories and warehouses. So, when AI displaces those workers, for example, the political feedback loop that would force the issue onto the local agenda is structurally weak.

I don’t think demographic decline will absorb the displacement, either. The labor force in China is shrinking, but AI is more likely to disrupt white-collar work – administrative, analytical, customer service, and junior professional roles – than to replace carpenters, electricians, or building maintenance workers. Youth unemployment among college graduates is already a problem, and AI development is unlikely to resolve it on its own. If anything, the diffusion-forward strategy we describe in the paper, which embeds AI agents and automation across white-collar service sectors as well as manufacturing, may sharpen the mismatch.

What we urgently need is more rigorous research. We do not yet have credible estimates of net job creation versus net job loss from AI diffusion in China or anywhere else. Universities and think tanks should be investing much more heavily in this. And to be clear, this is not a uniquely Chinese problem. Instead, it is a global problem. The U.S. has not had a serious public conversation about it, either. Part of the silence, I think, is that this is a politically sensitive issue for every government in every country, regardless of regime type. But the longer we delay confronting it, the more likely it becomes that displacement manifests as instability rather than as a problem we have actually prepared for.

Q: A successful diffusion-forward strategy means a continued export surge into a global trading system already absorbing a $1.4 trillion Chinese surplus. For Southeast Asian economies, particularly Indonesia, Vietnam, and Malaysia, which are simultaneously trying to industrialize and accommodate Chinese investment, what does the rapid embedding of AI into Chinese manufacturing imply for their development trajectories? Is there a productive form of engagement with Chinese AI-enabled production beyond defensive trade measures, especially for economies with manufacturing deficits and aging populations of their own?

A: I think it is too early to tell with any precision, and the next two to three years of domestic Chinese deployment will tell us a great deal about what eventually gets exported.

The starting point is the policy architecture itself. For example, the State Council’s August 2025 Implementation Opinion on the “AI+” Action and the December 2025 MIIT eight-ministry document on “AI+ Manufacturing” set extraordinarily granular targets for embedding AI across Chinese industry – production-process-level specifications for sectors ranging from shipbuilding to steel to textiles. This will play out first in the Chinese domestic market. What gets refined, scaled, and commercialized at home over the next few years is what will eventually be exported.

For Southeast Asian economies, the development trajectory will be shaped by AI more broadly, not necessarily by China specifically. These countries face a real choice about which technology stack to build their industrial upgrading around: American models and chips, Chinese models and chips, or some hybrid configuration. But the reality is that AI is developing very rapidly, and the pace is being set by the U.S. and China. Smaller economies will need to make their own strategic calculations, but they will inevitably be shaped by the broader trajectory of the technology.

I do think there are productive forms of engagement that go beyond defensive trade measures. The most promising, in my view, is to attract Chinese embodied AI and physical AI companies to invest directly in Southeast Asia – to build factories, transfer know-how, and upgrade domestic industrial capacity. This is essentially what China itself did over the past four decades by absorbing foreign direct investment from Japan, Korea, the U.S., and Europe. Southeast Asian governments could pursue something similar. Vietnam’s experience absorbing Chinese electronics manufacturing investment during the U.S.-China trade-war reshuffling offers a partial template.

Q: Your argument rests on observable institutional advantages like cascading center-local mandates, investor-state capital, dense supplier ecosystems around anchor firms, but also on bets that have not fully played out. Over the next two to three years, what evidence would lead you to revise or qualify the thesis? Conversely, what signals (like falling local guidance fund commitments, consolidation in humanoid robotics before scale is achieved, a frontier-model breakthrough that genuinely reshapes the value chain) would suggest the AGI-frontier path is in fact the dominant one after all?

A: Let me start by pushing back gently on the framing. I don’t see AGI and diffusion as contradictory. Frontier-model breakthroughs are enormously important, and arguably the most important upstream variable in this whole story. The point we make in our paper is not that frontier capability doesn’t matter, but that capability does not translate into diffusion automatically. Achieving AGI, or substantial breakthroughs in foundation models, does not by itself produce successful deployment across the physical economy. As the China case shows, diffusion at scale requires a well-designed institutional ecosystem and that ecosystem must be built deliberately, particularly when the goal is to compress the timeline from research to commercial deployment.

So, in the next few years, I would expect frontier-model progress and diffusion progress to reinforce each other. A genuine physical-world model, for instance, would dramatically empower the humanoid robotics sector by solving precisely the perception, prediction and real-time control problems that text-trained LLMs are not good at. But having the model is not the same as having the robots. Tesla offers a useful illustration: it has arguably the world’s most advanced autonomous driving system, but turning that capability into mass-produced, commercially competitive EVs required years of factory-level effort, supplier development and operational learning. None of that happened automatically.