Join War on the Rocks and gain access to content trusted by policymakers, military leaders, and strategic thinkers worldwide.
China’s plan to become a world leader in AI by 2030 is a fixture of practically every Congressional briefing and expert commentary on Beijing’s AI ambitions. The plan’s logic — introduced in 2017 — was simple and alarming: Beijing would direct capital, mobilize its firms, recruit talent, and execute with the strategic patience of a state-led innovation ecosystem. Nearly a decade later, that frame has only hardened. Beijing’s recently issued 15th Five-Year Plan directs Party organs to take “extraordinary measures” to strengthen technological self-reliance and launch a new “AI+” initiative to integrate AI across the nation’s strategic sectors. Beijing has the legal architecture to compel its firms to do its bidding, so Washington has largely concluded that Beijing’s AI sprint reflects deliberate industrial policy, and built America’s response around that assumption.
That conclusion, however, mistakes the frame for the picture. China’s AI rise is being driven as much by market forces as by state direction.
The more you look inside China’s AI ecosystem, the more it looks like the most brutal AI capitalist knife fight across the world. The domestic competition among firms is so fierce that Chinese commentators have a word for it: involution, or neijuan.
Involution means fierce price wars between battling AI labs that destroy margins across the entire industry, founders cannibalizing each other’s researchers and customers, and provincial governments bankrolling rival AI champions in zero-sum competition with neighboring provinces. Western observers may be unaware that, when it comes to China’s AI policy, the state and market often drive in different directions. Beijing is left struggling to impose coherence on a market of 1.4 billion people, 5,100 AI firms, and a messy array of colliding incentives that no five-year plan fully anticipated.
Involution in China’s AI market takes three primary forms, each shaped in part by Beijing’s own policies, and none of which Beijing can fully control. Price wars that destroy margins across the entire industry are the first and most obvious form. In May 2024, ByteDance slashed the price of its Doubao model by 99 percent, making it nearly free for enterprise users. Alibaba, Baidu, and Tencent followed within days, gutting mid-tier firms and thinning the field. Two years later, the price wars are continuing to inject volatility into the competition, with DeepSeek announcing it would cut its latest V4-Pro prices by 75 percent permanently, which will likely set off a fresh round of retaliatory cuts from competitors with no price floor in sight.
The second form of involution is talent cannibalization. In China, the entire frontier AI ecosystem operates within a two-hour flight, the labor market is enormous, and the competitive pressure to strip rivals of their best people is existential. A competitive talent market is not strictly a negative byproduct of involution, but firms poaching top researchers from rivals rather than developing new talent from scratch is. AI firms from Bytedance to Tencent have recruited rival researchers, offering millions and launching elite recruitment drives to stay ahead.
Involution’s third engine is provincial competition. Local governments across China are bankrolling rival AI champions in competitive races to claim credit for producing the next DeepSeek, sometimes backing companies on other provinces’ turf entirely. Thirty-plus provincial governments are doing exactly that, each with its own capital, its own preferred firms, and its own definition of winning, often in direct tension with Beijing’s national priorities. Anhui province’s own AI development plan explicitly cites regional competition, not national security imperatives or Beijing’s strategic directives, as the driver of its AI investments.
Involution is not unique to China’s AI sector, but is happening at a faster speed. In batteries, solar, and electric vehicles, Beijing helped build conditions that produced catastrophic overcapacity, and has spent years trying to unwind with limited success.
However, in the case of AI, Beijing is far more adept at harvesting the results of involution, even as it struggles to govern the process. The Party has selectively pulled winners toward the state while trying to suppress the dynamic’s most destructive tendencies. The anti-involution campaign launched in 2025 called on industries from electric vehicles to AI to compete on quality rather than price. It is the clearest public admission that the competitive dynamic Beijing had once encouraged was now undermining necessary innovation. A state that is simultaneously trying to harvest and suppress a market dynamic is not executing a strategy so much as improvising one.
Beijing’s relationship with its tech juggernauts is not strictly command-and-control. It is something more opportunistic, and in some ways more adaptive.
In the early 2020s, China’s tech giants grew powerful enough to openly defy the party-state. Jack Ma publicly criticized China’s regulators at a financial conference. A month later, Beijing cancelled Ant Group’s $37 billion IPO and fined Alibaba $2.75 billion. The crackdown that followed sent the message that the knife fight could produce winners, but Beijing would decide what happened to them.
When the AI boom arrived, and competition was generating strategically-significant capability at speed — from DeepSeek’s R1, to Alibaba’s Qwen, and Moonshot AI’s Kimi — Beijing granted the companies and their founders a longer leash. The central government imposed a lighter regulatory touch, provincial investment flowed freely, and founders were left to compete, cannibalize, and optimize. Regulations on generative AI proposed by Chinese regulators in 2023 were softened under industry pressure.
But this latitude for private industry did not last. At Chinese leader Xi Jinping’s February 2025 private sector symposium, DeepSeek’s Liang Wenfeng was seated prominently alongside Huawei’s Ren Zhengfei. DeepSeek was born as a hedge fund side project that initially avoided all outside capital and state funding to protect its global ambitions from Beijing. Yet it is now in advanced talks to raise $7.35 billion led by China’s state-backed National AI Fund, with its V4 model optimized to run on Huawei’s Ascend chips. DeepSeek has recently prioritized state-backed investors that can provide access to compute infrastructure over those that can only offer capital, a move driven more by compute scarcity rather than proactive government alignment. The company that emerged from market competition is being pulled toward the state, through capital, silicon, and proximity to power. Further, when Meta acquired Manus for $2 billion in December 2025, China’s regulators ordered the deal unwound in April 2026. This marked the first time Beijing had used foreign investment security review measures to unwind a completed acquisition. Every Chinese AI founder building for international markets now understands that any initial period of unfettered competition and growth will eventually give way to government direction.
China’s cutthroat domestic market has trained its AI firms to do whatever it takes to survive, and that conditioning does not stop at the border. Firms cut off from frontier compute by U.S. export controls have sought lawful procurement where possible, but built sophisticated smuggling networks to procure restricted advanced AI chips through intermediaries in Southeast Asia and the Middle East. Firms like DeepSeek, Moonshot AI, and MiniMax have conducted distillation attacks on U.S. frontier models like Anthropic, extracting capability from a competitor the way they would poach a researcher from a rival lab. The White House called it a “deliberate, industrial-scale campaign,” but this is involution applied internationally: If you can’t build it, extract it from someone who did.
Additionally, domestic competitive pressure, compounded by external constraints on compute, pushed Chinese labs to aggressively develop open-weight models. Unlike closed models that charge for access, open-weight models make their underlying code publicly available, allowing developers anywhere to download, modify, and deploy them on local hardware for free. This allows the models to scale a global developer ecosystem rapidly while bypassing the compute constraints that export controls impose and spawning hundreds of thousands of derivatives. A recent study found that Chinese open-weight models accounted for 17.1 percent of global AI model downloads, surpassing the U.S. share of 15.9 percent for the first time. This happened not because of government mandates, but because developing open-weight models became one of the only ways to survive China’s brutal domestic competition. With open-weight models, Chinese companies could still make money by capturing the global developer ecosystem. The firms that learned to optimize for survival at home are now shaping the foundation layer of global AI development.
The behavior of Chinese AI firms, illicit and overt alike, is state-enabled but market-conditioned. Market-conditioned behavior finds workarounds faster than state-directed programs because survival, not ideology, is the motivation. The policy implications of market-conditioned behavior and state-directed behavior are not the same, and Washington cannot continue to address them with the same tools.
Export controls, entity listing, and investment screening were all designed on the premise that Beijing is the engine of China’s progress in developing advanced AI models. If involution is driving China’s AI progress as much as Beijing’s industrial policy, then policy centered exclusively on counteracting Beijing’s directives will keep missing the progress that Chinese firms can make without being forced to do so by a ministry or five-year plan.
As President Donald Trump and Xi signal a shared desire to avoid further escalations in bilateral tensions, Washington should use the relative calm to recalibrate.
On export controls, categorical controls on frontier chips remain the right instrument. Compute is China’s genuine weakness, and they are harder to route around than targeted ones. But sustaining this advantage is difficult, and export controls are a depreciating asset. Beijing is now actively blocking its own companies from purchasing advanced U.S. chips they could otherwise legally acquire, steering domestic demand toward Huawei as a matter of industrial policy. Extending that window of advantage does not mean redesigning controls, but closing transshipment loopholes, coordinating with allies on end-use verification, and building a credible enforcement architecture.
On investment screening and entity listing, the priority is precision. Blunt entity listing collapses the distinction between firms whose behavior is driven by market competition and those integrated into China’s military-civil fusion strategy. Not all Chinese AI firms have the same proximity to Beijing. Military-civil fusion touches some companies deeply and others barely at all. An investment screening and entity listing regime calibrated to actual state integration rather than country of origin would be more durable, more convincing internationally (which matters because export controls, bans, and entity lists are more effective with allied buy-in), and more likely to give market-driven firms pause before siding with the state. State integration could be measured not just by government ownership stakes or direct state investment, but by the firm’s reliance on government contracts as a share of total revenue, government-affiliations of key personnel at the firm — such as concurrent appointments at state research institutes, Party membership, or prior service in government or military roles — and whether the firm’s technology stack offers the state bespoke capabilities it cannot find elsewhere within the government apparatus.
This is not an argument for exempting market-driven firms from scrutiny, but for prioritizing state-integrated ones, and addressing market-conditioned international aggression through better-suited enforcement architecture. Applied without that precision, these tools risk doing Beijing’s consolidation work for it. None of these actions will prevent Beijing from co-opting its most strategically valuable firms. The goal instead is to delay, create friction, force tradeoffs, and compound the pressures that involution is already generating inside China’s AI ecosystem.
If that toolkit buys time, Washington needs to build a new competitive strategy to meet China at the open-weight model layer of global AI development. Involution has optimized Chinese AI for a different kind of winning. Roughly 80 percent of U.S. startups use Chinese base models as their foundation, not because Qwen or DeepSeek are necessarily superior, but because they are cheaper and more accessible. A model that is good enough to power enterprise applications and developer workflows at a fraction of frontier costs doesn’t need to top performance leaderboards, but be embedded in the ecosystem before superior alternatives arrive. Without actively promoting viable Western alternatives, Washington risks repeating mistakes made in the 5G competition, where trying to undercut Huawei without backing its competitors on price and financing ceded Global South markets to China by default.
The United States should build sustained public-private partnerships to develop and deploy Western-origin open-weight models, paired with accessible compute — lower-tier Western hardware and allied semiconductor alternatives that non-Western data centers can actually use. And working with industry partners and allies to establish open-weight model security standards, from training data provenance to architectural transparency and security auditing, would create market conditions that reward trustworthy alternatives without requiring direct restriction after they are widely deployed.
The one layer of this competition where Washington’s advantage is most durable, but also most at risk of being squandered, is basic research in American universities and national laboratories. Xi Jinping himself has acknowledged China’s significant gaps in original innovation and convened a rare symposium of top Party officials and researchers to better align research projects with national priorities. The optimization and speed that price wars, talent cannibalization, and provincial competition produce do not substitute for the foundational scientific breakthroughs that have historically originated in American universities and national laboratories. Sustained investment in basic research and development, through government institutions like the National Science Foundation, National Institute for Standards and Technology, and the Department of Energy National Laboratories, is not a generic competitiveness argument. It is the specific response to involution’s structural blind spot: a market dynamic that optimizes for deployment and efficiency, but systematically underinvests in the long-horizon fundamental research that defines what the next generation of AI looks like.
Washington has spent nearly a decade anchored to China’s 2017 AI development plan as its mental model of Beijing’s AI ambitions. That plan is real, but Beijing’s plans only describe intentions. What China’s AI decade has actually produced emerged from a messy market dynamic: thousands of competing firms undercutting each other to the point where Beijing itself has acknowledged the dynamic is undermining China’s capacity for original innovation. Involution has also built hungry Chinese AI firms, ready to fight dirty in the international markets, by smuggling chips, launching distillation campaigns, and claiming market share through optimization rather than frontier breakthroughs. The way to confront this complex web of market and state-induced dynamics from China is to recalibrate the tools Washington already has, to compete where Chinese firms are already winning by default, and to remember that Washington’s sharpest long-term advantage may be its own research and development base. Until Washington learns to read that dynamic as carefully as it reads Beijing’s plans, it will keep designing policy for a China that exists on paper — and missing the one that is actually being forged.
David Lin is a senior advisor for technology leadership at the Special Competitive Studies Project. He served for over a decade as a CIA intelligence analyst and U.S. Department of State foreign service officer covering Asia and technology policy.
All statements of fact, opinion, or analysis expressed are those of the author and do not reflect the official positions or views of the U.S. Government. Nothing in the contents should be construed as asserting or implying U.S. Government authentication of information or endorsement of the author’s views.
Image: Midjourney
