The Story Washington Keeps Getting Wrong
Every Congressional briefing on China’s technological ambitions tends to open with the same foundational document Beijing’s 2017 plan to become the world leader in artificial intelligence by 2030. The plan is ambitious, clearly articulated, and backed by the full institutional weight of the Chinese Communist Party. Analysts cite it, policymakers quote it, and strategic thinkers in Washington have built entire response frameworks around it. The story it tells is clean, orderly, and deeply alarming: a patient state mobilizing vast resources, directing compliant corporations, and marching with disciplined purpose toward technological supremacy. That story, compelling as it is, describes only one dimension of what is actually happening inside China’s AI ecosystem. The version being written on the ground inside the research labs of Hangzhou, the compute farms of Shenzhen, and the pitch rooms of Shanghai looks less like a coordinated military advance and more like a street fight with no referee. Chinese firms are not marching in formation toward a state-determined goal. They are tearing into each other with a ferocity that even Beijing finds difficult to manage, let alone fully direct.
The phenomenon has a name in Chinese: neijuan, commonly translated as “involution.” The term describes a condition of intensifying competition in a closed system, where the collective energy expended keeps rising while the collective gains keep shrinking. It entered popular usage as a complaint about academic and professional culture in China brilliant young people running faster and faster on a treadmill that does not actually move them forward. China’s AI sector has absorbed that dynamic and amplified it to an industrial scale, producing an ecosystem defined by price wars that destroy margins, talent battles that cannibalize the very firms they are meant to strengthen, and redundant provincial investments that duplicate effort while generating the illusion of coordinated progress. Understanding this dynamic matters enormously not just for academic analysis, but for practical policymaking in Washington. The tools that the United States has deployed to slow China’s AI progress were largely designed around the assumption that Beijing is the primary engine of China’s advancement. When a market dynamic, rather than a ministry, is driving the pace and direction of development, those tools operate on incorrect assumptions about what they are trying to constrain. The narrative of a monolithic, top-down Chinese AI strategy is not merely incomplete. It is operationally misleading for any government attempting to formulate a competitive response.
How Beijing Harvests What It Cannot Control
The State’s Selective Co-option of Market Winners
Beijing’s relationship with its AI sector is not one of command and control. It is something more operationally interesting and more strategically significant a pattern of selective co-option, in which the state allows markets to run, observes who wins, and then absorbs the winners into a structure that serves state objectives. This is not a strategy that any ministry designed in advance. It is a behavioral pattern that has emerged from the intersection of an authoritarian state’s instinct for control and a market’s capacity to generate capability faster than bureaucracies can direct it.
The Jack Ma episode of 2020 and 2021 established the terms of engagement with unusual clarity. Ma publicly criticized Chinese financial regulators at a major industry conference. Within weeks, Beijing canceled the Ant Group initial public offering at the time one of the largest in history and subsequently imposed a fine on Alibaba that sent an unambiguous signal across the entire technology sector. The message was not that private firms could not grow large or generate enormous value. The message was that private firms grew at the state’s sufferance, and that public defiance of state authority would be met with consequences calibrated to remind the market of that fact. When the AI boom arrived and competitive dynamics began generating genuine frontier capability at speed DeepSeek’s R1, Alibaba’s Qwen series, Moonshot AI’s Kimi — Beijing adjusted its posture. Regulatory requirements for generative AI that Chinese authorities had proposed in 2023 were softened after industry pressure, a concession that reflected Beijing’s recognition that overly restrictive regulation would slow capability development that the state wanted to harvest. Provincial investment flowed freely. Founders competed, cannibalized each other’s talent, and optimized without the kind of heavy regulatory oversight that Beijing had imposed on earlier technology cycles.
DeepSeek’s trajectory illustrates the co-option dynamic with particular precision. The company began as a side project of a hedge fund, deliberately avoided outside capital and state funding, and structured itself to maintain the independence that global ambition requires. Founder Liang Wenfeng understood that a company too visibly connected to the Chinese state would face credibility problems in international markets where data security concerns are already acute. Yet by early 2026, DeepSeek was in advanced talks to raise capital led by China’s state-backed National AI Fund, and its latest model had been optimized to run on Huawei’s Ascend chips — a Chinese-made alternative to Nvidia’s hardware that the state has strong incentives to promote as export controls deepen.
The transition from independent market actor to state-adjacent institution was not driven purely by government pressure. DeepSeek prioritized investors who could provide compute infrastructure access over those offering only financial capital a decision driven by the hardware scarcity that U.S. export controls have created. The state’s gravitational pull operated through economic necessity as much as political coercion, which makes the dynamic harder to identify and harder to interrupt. The Meta-Manus episode of 2026 made the political dimension explicit in a way that left little room for ambiguous interpretation. When Meta completed its acquisition of the Chinese AI startup Manus for approximately two billion dollars in late 2025, China’s regulators ordered the deal unwound in April 2026 the first time Beijing had exercised foreign investment security review authority to reverse a completed transaction. Every Chinese AI founder building products for international markets now operates with the understanding that an initial period of market freedom will eventually give way to the state’s assertion of strategic interest.
How Involution Crossed the Border
The Open-Weight Strategy Born From Necessity
Perhaps the most consequential geopolitical output of China’s domestic AI knife fight is one that no Beijing planner specifically designed: the dominance of Chinese open-weight models in the global developer ecosystem. Open-weight models are AI systems whose underlying parameters the numerical values that encode the model’s learned behavior are published and made available for anyone to download, modify, and deploy on their own hardware. Unlike closed commercial models, which generate revenue through API access fees and restrict users to interacting with the model as a service, open-weight models allow developers anywhere in the world to build applications on top of them without paying ongoing costs to the original developer.
Chinese AI firms did not gravitate toward open-weight development because a government directive told them to. They gravitated toward it because the domestic price wars had destroyed the commercial logic of charging for API access, export controls had cut off access to the frontier compute that closed-model economics requires. Open-weight distribution offered a path to building a global developer base that could generate indirect returns through cloud infrastructure deals, hardware sales, and enterprise customization contracts even when direct model revenue was nonexistent. The scale at which this shift has operated is striking. Alibaba’s Qwen model family ranked as the most downloaded model series on the Hugging Face platform in both 2025 and 2026, surpassing Meta’s Llama models in cumulative downloads. On Hugging Face the primary repository through which the global developer community accesses and shares AI models DeepSeek is now the most followed organization, and Qwen ranks fourth. The papers drawing the most attention from the global research community on that platform come predominantly from Chinese institutions: ByteDance, DeepSeek, Tencent, and the Qwen team at Alibaba.
This outcome did not emerge from soft power strategy or deliberate international influence operations. It emerged from the logic of surviving in a domestic market where price competition had eliminated most conventional revenue models. Firms that learned to compete on a cost basis by necessity discovered, somewhat unexpectedly, that the same approach positioned them to capture a disproportionate share of the global developer ecosystem. The open-weight model that was designed to survive China’s domestic knife fight turned out to be the format best suited to capture the attention of the global developer community that values cost, accessibility, and the ability to run models locally. The strategic implications are profound and not yet fully appreciated in Washington. A developer who builds an application on a Chinese open-weight model as their foundation layer does not necessarily make a political choice or a security choice. They make an economic choice: the Chinese model is cheaper, often competitive on capability for most commercial use cases, and freely available. The switching costs that accumulate as developer communities build tooling, workflows, and institutional knowledge around a particular model architecture are real and durable, which means that the foundation layer of global AI development is being defined by competition dynamics that originated in a Chinese price war.
When Domestic Aggression Becomes International Aggression
The behavioral conditioning that involution produces does not stop at China’s borders. Chinese firms that have spent years operating in a market where conventional rules of competitive conduct do not apply have developed an institutional tolerance for aggressive tactics that they now deploy in international markets. The export control architecture that the United States built to restrict China’s access to frontier compute became, in the hands of firms hardened by involution, something to route around rather than comply with.
Sophisticated procurement networks operating through intermediary entities in Southeast Asia and the Middle East emerged as supply chains for restricted Nvidia hardware that Chinese AI labs could not legally acquire through direct channels. These networks operated not because Beijing mandated them but because firms facing existential competitive pressure at home had already developed the organizational muscle for operating in environments where the rules are adversarial. The same competence that a firm develops to survive a domestic market where competitors will destroy your margins, poach your researchers, and replicate your architecture the moment you publish it, translates directly to the competence needed to navigate an international regulatory environment designed to constrain you.
Perhaps the most technically sophisticated form of international aggression is what the United States government has described as distillation attacks a practice in which Chinese AI labs use frontier Western models to generate training data for their own systems, effectively extracting capability from competitors without paying for it. Anthropic publicly documented efforts by Chinese firms to conduct industrial-scale distillation campaigns against its systems. The White House, in a 2026 national security memorandum, characterized the practice as a deliberate, coordinated effort to circumvent the investment and research that produced the frontier models being exploited. The analogy to domestic talent poaching is direct: if you cannot build the capability yourself, extract it from someone who did. The tactics are different; the underlying logic is identical.
This pattern of behavior is harder to address with conventional trade and investment policy instruments, because the motivating force is market survival rather than state direction. Export controls designed to constrain Beijing’s directed programs encounter a different challenge when the actors being constrained are firms operating under existential competitive pressure that would motivate the same behavior regardless of what Beijing instructed. The compliance infrastructure that deters state-directed programs diplomatic consequences, strategic relationships, reputational considerations does not necessarily deter a firm whose domestic market has conditioned it to treat circumvention as a standard operating procedure.
What Washington Needs to Recalibrate
Export Controls, Precision, and the Compute Chokehold
The United States has built its primary response to China’s AI ambitions around a set of tools export controls on advanced chips, entity listing of Chinese technology firms, investment screening that were each designed on the premise that Beijing’s directed industrial policy is the central mechanism of China’s advancement. Categorical controls on frontier semiconductor exports remain the right instrument for the core strategic problem: compute is the genuine constraint on frontier AI development, and restricting China’s access to the hardware needed to train the most capable models addresses a real vulnerability in China’s position.
Sustaining that advantage, though, requires active maintenance rather than passive policy entrenchment. Transshipment networks operating through Southeast Asian and Middle Eastern intermediaries have already demonstrated the capacity to route restricted hardware to Chinese end users at scale. Beijing has taken the unusual step of actively discouraging its own firms from purchasing advanced American chips they could otherwise legally acquire, steering domestic demand toward Huawei’s Ascend hardware as a matter of industrial policy. The compute chokehold is real, but it requires consistent enforcement attention and allied coordination on end-use verification to remain effective. A control regime that cannot be enforced reliably provides primarily psychological reassurance to policymakers rather than genuine strategic constraint.
Entity listing and investment screening present a different calibration problem. Blunt application of these tools treats all Chinese AI firms as equivalent threats, collapsing the distinction between firms deeply integrated into China’s military-civil fusion apparatus and firms whose behavior is driven primarily by market competition. The evidence from China’s AI sector suggests these are not the same population. A company whose primary investor is a state defense fund and whose technology stack provides bespoke capability to the People’s Liberation Army occupies a fundamentally different risk category than a startup whose aggressive international behavior is driven by the competitive desperation that domestic involution generates. Precision in entity listing calibrated to actual evidence of state integration rather than country of origin — would produce a more durable and internationally persuasive framework. It would be more likely to create real friction for the firms that pose genuine security risks, and less likely to inadvertently do Beijing’s consolidation work for it by eliminating market-driven competitors that the state has not yet been able to absorb. This is not an argument for reducing scrutiny of Chinese AI firms. It is an argument for applying scrutiny in a way that accurately reflects the actual structure of the risk being addressed.
The Open-Weight Competitive Gap That Is Already Opening
While Washington has focused its competitive attention on the frontier model layer the most capable, most compute-intensive systems produced by OpenAI, Anthropic, and Google DeepMind Chinese firms have been quietly capturing the foundation layer of global AI development through open-weight model distribution. The gap this creates is not hypothetical. A substantial share of American AI startups now use Chinese base models as their technical foundation, not because those models are necessarily superior on frontier benchmarks, but because they are dramatically cheaper and more accessible for the range of commercial use cases that most enterprise applications require.
The 5G competition provides a cautionary analogy. The United States spent years attempting to undercut Huawei’s position in global telecommunications infrastructure through restriction and entity listing, without simultaneously backing competitive Western alternatives at the price points and financing terms that emerging market customers could realistically adopt. The result was that the effort to contain Huawei succeeded in delaying its penetration of certain markets while ceding others to Chinese equipment by default not because Western alternatives were technically inferior, but because they were not available on terms that matched the competitive offer.
The open-weight AI layer is developing its own version of that dynamic. Chinese models are embedding themselves as the default technical foundation for developers in emerging markets, in academic institutions with limited compute budgets, and in commercial enterprises that need capable AI tools at costs that closed frontier models cannot match. Developers build tooling around these models. Research communities develop expertise in their architectures. Applications are built on top of their specific capabilities and limitations. Each of these investments represents a switching cost that makes future transition to Western-origin alternatives incrementally more difficult. Building a genuine competitive response requires more than placing restrictions on Chinese open-weight models after they are already widely deployed. It requires sustained public-private investment in Western-origin open-weight model development, paired with compute access that makes deployment practical for the developers and institutions currently defaulting to Chinese alternatives. The absence of a credible, accessible Western competitor at the open-weight layer is a market failure, and market failures at strategic technology layers are not corrected by restriction alone.
Basic Research and the Blind Spot Involution Creates
The structural blind spot that involution generates in China’s AI ecosystem points directly to where the United States retains its most durable competitive advantage and most urgently risks squandering it. Xi Jinping himself has publicly acknowledged that China’s AI sector, despite its remarkable progress on deployment and efficiency, still has significant gaps in original innovation. The convening of a high-level symposium of Party officials and researchers to better align basic research priorities with national strategic goals reflects the leadership’s awareness that the market dynamics driving China’s AI efficiency gains are the same dynamics preventing the deeper scientific inquiry that the next generation of AI will require. American universities and national laboratories have historically been the primary source of the foundational scientific breakthroughs that define new AI paradigms. The transformer architecture that underlies essentially every large language model in use today originated from research conducted with access to academic freedom, long time horizons, and funding structures that do not require quarterly justification. The attention mechanisms, the training methodologies, the theoretical frameworks for understanding generalization these advances came from environments that were structurally insulated from the kind of competitive pressure that involution creates.
That advantage is not self-sustaining. It requires active, sustained investment in the institutions the National Science Foundation, the Department of Energy national laboratories, the National Institute of Standards and Technology that fund and protect the kind of research that cannot justify itself on a commercial timeline. Political pressure to cut basic research funding in favor of applied programs that demonstrate near-term returns is a perennial feature of American budget debates. In the context of the U.S.-China AI competition, yielding to that pressure would represent a strategic concession that no amount of chip export control can compensate for. The specific geometry of the competitive challenge makes the basic research investment argument more precise than it typically sounds. China’s AI sector has become genuinely formidable at the layers of the technology stack where market dynamics drive efficient optimization deployment infrastructure, model distillation, efficiency tuning, cost compression. Those are the layers where involution’s pressure produces capability. They are not the layers where the next fundamental paradigm shift in AI will originate. The paradigm shift will come from basic research. That is the layer where the United States has the structural advantage of an academic and institutional ecosystem that market pressure has not yet sufficiently eroded and that represents the most important competitive asset to protect.
The Frame That Keeps Failing
There is a particular kind of analytical error that becomes self-reinforcing over time. When policymakers organize their mental model around a single explanatory framework Beijing directs, companies comply, outputs reflect state intentions every data point that fits the framework gets amplified, and every data point that contradicts it gets absorbed as a modification rather than a challenge. The existence of the 2017 AI development plan is real. Beijing’s genuine efforts to steer technology development are real. The capital flows from state-backed institutions into Chinese AI firms are real and growing. None of this is fabricated. What the framework misses is the degree to which the most consequential outputs of China’s AI decade the efficiency innovations, the open-weight model ecosystem, the aggressive international market behavior emerged from dynamics that the state neither designed nor fully controls. DeepSeek did not become a global phenomenon because Beijing told a hedge fund to build a more efficient large language model. It became a global phenomenon because a market environment defined by brutal competitive pressure forced engineering choices that happened to produce remarkable results. The price wars that have made Chinese AI models competitive on cost at the global level did not emerge from a ministry directive. They emerged from the same competitive desperation that produces involution in every industry where many firms fight over a constrained market.
Understanding this does not reduce the competitive challenge that China’s AI sector presents. If anything, it makes the challenge more difficult, because market-conditioned behavior finds workarounds faster than state-directed programs. A state program operates within bureaucratic constraints, political approval processes, and institutional inertia that create predictable friction. A firm operating under existential competitive pressure operates with the full urgency of survival, and the full creativity that survival under pressure generates. The chip smuggling networks, the distillation campaigns, the open-weight distribution strategy these reflect the competitive instincts of firms that have been forged in a domestic knife fight, not the deliberate strategic calculations of a government agency executing a master plan. The most honest thing that can be said about China’s AI trajectory is that it has produced outcomes more consequential than Beijing planned for, through processes messier than Beijing intended, and with structural weaknesses that Beijing acknowledges but cannot easily correct. Washington’s response needs to be calibrated to that reality — precise where precision is possible, competitive where competition is required, and invested in the research foundations that the domestic market structure of an involution-driven ecosystem systematically undervalues.
