A new analysis published in the War on the Rocks “Cogs of War” column by researcher David Lin offers one of the most rigorous portraits yet of how China’s domestic AI industry actually functions, and why Western policymakers may be fundamentally misreading it. The picture that emerges is not of a monolithic state-directed machine, but of something far more chaotic, more competitive, and in some respects more dangerous to American technological leadership than the consensus view suggests.
A Market of 5,100 Firms and No Mercy
China’s AI ecosystem currently comprises roughly 5,100 firms, a figure that by itself signals the scale of the competitive pressure within its borders. This is not a market where incumbents rest comfortably on government subsidies. It is, as Lin’s title suggests, a knife fight.
The clearest illustration came in May 2024, when ByteDance slashed the price of its Doubao model by 99%. Not discounted. Not restructured. Ninety-nine percent. Alibaba, Baidu, and Tencent followed within days, triggering a race to zero that has since become the defining feature of China’s AI economy. The episode forced every domestic player to re-examine its cost structure, its hardware dependencies, and its capacity to survive on razor-thin margins, a pressure cooker that, paradoxically, accelerates capability development even as it destroys profitability. As EastFrontier has reported, ByteDance’s profit fell more than 70% in 2025 as it bet its margins on AI, while Alibaba’s core profit plunged 84% under similar pressures.
The latest trigger for another round of cuts is DeepSeek’s announcement of a permanent 75% price reduction on its V4-Pro model, which Lin argues will likely set off fresh retaliatory pricing across the industry. Given the pattern established in 2024, there is little reason to doubt that prediction.
Provincial Governments as Fuel on the Fire
What makes this competition structurally different from market dynamics in the United States or Europe is the role of regional government. More than 30 provincial governments are actively backing rival AI champions, creating a patchwork of competing state-supported ecosystems that operate with local, not national, logic. Lin highlights Anhui province’s AI development plan as a particularly revealing document: it explicitly cites regional competition, not national security or military modernization, as the primary driver of its AI investment strategy.
This is a crucial distinction. The dominant Western framing of China’s AI buildup as a top-down, centrally coordinated military-industrial project misses the extent to which provincial rivalry, commercial vanity, and local GDP targets are driving much of it. Beijing sets the broad direction; local governments compete ferociously within it, often working at cross-purposes. The result is redundant investment, overcapacity, and the relentless price compression that has come to define the sector.
Beijing has acknowledged the problem. The anti-involution campaign launched in 2025 was an explicit attempt to redirect competition from price-cutting to quality improvement, an acknowledgment from the center that the knife fight had gone too far. Whether that campaign will succeed against the structural incentives of 30-plus provinces backing their own champions remains genuinely uncertain.
Open-Weight Models and the Global Footprint
Perhaps the most strategically significant data point in Lin’s analysis concerns global model downloads. Chinese open-weight models now account for 17.1% of global AI model downloads, compared to 15.9% for American models, the first time China has surpassed the United States on this metric. The primary vehicles for this lead are Alibaba’s Qwen family and DeepSeek, both of which have achieved extraordinary global reach. Alibaba’s Qwen family has surpassed one billion downloads, capturing over 50% of the global open-source AI market.
The downstream implications are significant. Lin reports that approximately 80% of American AI startups are now using Chinese base models, primarily Qwen or DeepSeek, as the foundation for their own products and services. Critically, he notes this adoption is not primarily driven by superior capability. It is driven by cost and accessibility. Chinese models are cheaper to run and, in the open-weight ecosystem, freely available for fine-tuning and deployment. American policymakers who have focused on preventing China from acquiring advanced chips have largely failed to reckon with the scenario in which Chinese software becomes load-bearing infrastructure for the American AI economy.
The Manus Moment and Beijing’s New Regulatory Muscle
Lin’s analysis also highlights a geopolitical inflection point that received surprisingly little coverage at the time. In December 2025, Meta acquired Manus, the AI agent startup that had generated enormous buzz in the Chinese tech community — for $2 billion. In April 2026, Beijing ordered the deal unwound. This was the first time China had used its foreign investment security review process to reverse a completed acquisition, a significant escalation in Beijing’s willingness to use regulatory tools as instruments of industrial policy in the AI sector.
The episode sits alongside the more famous case of Jack Ma publicly criticizing regulators, after which Beijing cancelled Ant Group’s $37 billion IPO and subsequently fined Alibaba $2.75 billion. Both cases illustrate the degree to which Beijing retains the ability to intervene decisively when it judges that commercial outcomes conflict with strategic priorities, even after deals are done.
At the same time, major capital continues to flow into the sector’s leading players. DeepSeek is currently raising $7.35 billion in a round led by China’s state-backed National AI Fund, with its V4 model specifically optimized to run on Huawei Ascend chips — a clear signal of the state’s interest in building a domestic hardware-software stack that reduces exposure to American export controls. This dynamic, covered extensively at EastFrontier in the context of AI chip smuggling networks and federal enforcement actions, underscores how the chip war and the model race are becoming increasingly intertwined.
Rethinking the Policy Response
Lin’s policy conclusion is pointed: the United States should calibrate its entity listing decisions to actual state integration rather than country of origin. The current approach, which effectively treats all Chinese AI companies as extensions of the state, both overstates the degree of central coordination in China’s AI sector and understates the genuine strategic differences between firms with deep People’s Liberation Army ties and those competing primarily in commercial markets. Blunt instruments, Lin argues, risk driving Chinese developers further toward Huawei hardware and closed domestic ecosystems, the precise outcome that chip controls were designed to prevent.
The broader picture that emerges from this analysis demands more nuance than the “China is winning” or “China is bluffing” narratives that dominate Western discourse. As EastFrontier has reported, an American researcher who spent time inside China’s AI labs found a similar complexity on the ground. The knife fight is real, the casualties are real, and the companies that survive it will emerge with a cost discipline and competitive resilience that should concern anyone who assumes American AI dominance is self-sustaining.