The US Is Running Three AI Strategies Against China. They're Undermining Each Other.

In February 2026, the Trump administration eased export restrictions on Nvidia's H200 chips to China. Beijing declined to approve the purchases. Chinese companies, the government signaled, would use domestic alternatives instead. ^[1][2] That moment is the structural fact at the center of the rest of US policy. The United States is now running three AI strategies against China simultaneously — and the first has already produced conditions that undermine the logic of the second and third. The three tracks are containment, competition, and persuasion. Containment is the oldest: export controls on advanced chips, restrictions on model access, and investigations of American firms that use Chinese models. Competition is the administration's preferred public posture — more than $700 billion in domestic AI infrastructure investment by Meta, Amazon, and Microsoft, plus the argument that the US should keep China on American platforms rather than push it toward domestic ones. Persuasion is the newest: a paid influencer campaign called Build American AI, and a stream of rhetoric from Congress framing the contest as a moral one. The problem is not that these tracks are different. It is that the first has already made the second and third commercially irrational — and the evidence is not a forecast. It is in the numbers. Containment was designed to deny China the chips required to build frontier AI. Jensen Huang warned what would happen instead.

if they can’t buy from us, they’ll build their own — and then we’ll have a competitor we wouldn’t have otherwise created. — Jensen Huang

That is exactly what occurred. China's AI chip self-sufficiency rose from 10% in 2021 to 41% in 2026, with Morgan Stanley projecting 86% by 2030. Zhipu AI trained its GLM-5.2 model entirely on 100,000 Huawei Ascend 910B processors — and Silicon Valley engineers who tested it called it a viable daily driver competing with GPT-5.5 and Claude Opus 4.8. DeepSeek excluded Nvidia and AMD from its V4 optimization cycle, favoring Huawei instead. The US-China Economic and Security Review Commission, a congressional body, acknowledged in March that export controls are "not well suited to addressing the physical loop" of China's deployment-driven strategy and that "open model proliferation creates alternative pathways to AI leadership."

Open model proliferation creates alternative pathways to AI leadership. — U.S.-China Economic and Security Review Commission

The self-sufficiency containment produced has now generated a cost structure the competition track cannot match. In May, DeepSeek permanently cut V4-Pro prices by 75%, to between 0.025 and 6 yuan per million tokens — making it 12 to 19 times cheaper than GPT-5.5 or Claude Opus 4.7 for equivalent tasks. The price cut was structural, not promotional: it was made possible by the Huawei Ascend 950 processors the model runs on, the same domestic chip ecosystem containment accelerated. Against this, the competition track is spending enormous sums to build capacity the market may not need at prices it cannot sustain. Meta committed $600 billion to US AI infrastructure. Amazon committed $200 billion. ^[3] In July, Meta launched Meta Compute — a service to sell excess AI capacity, a direct admission of oversupply. Mark Zuckerberg had earlier signaled the option was available "if we get to a point where we feel that we have overbuilt." ^[4]

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

The market has noticed. Meta's stock is down roughly 12% in 2026, Microsoft's roughly 20%, even as Meta's core advertising business grew 33% year over year and Microsoft's Azure grew 40%. ^[5] The sell-off is not a rejection of these companies' current earnings. It is a judgment that the infrastructure investment will not generate returns sufficient to justify its cost when Chinese alternatives are available at a fraction of the price. Meanwhile, the data center buildout is hitting physical limits: aging power grids cannot support the expansion without massive new investment in nuclear, natural gas, and renewable generation. ^[6] The competition track's own most prominent advocate, Jensen Huang, has made the case in precisely these terms: keep China on Nvidia's CUDA platform, because pushing them to Huawei's stack creates a competitor the US cannot later dislodge. ^[7]

China's AI moves on with or without U.S. chips. It has to compute to train and deploy advanced models. The question is not whether China will have AI, it already does. The question is whether one of the world's largest AI markets will run on American platforms. — Jensen Huang

But the administration is not following Huang's advice. It is running both tracks at once. In June alone, Commerce closed the subsidiary loophole that had allowed chip exports to continue through third countries, and ordered Anthropic to disable its frontier models for foreign nationals — even as Meta and Amazon were writing the checks for their infrastructure buildouts. The whipsaw is not a pivot. It is a contradiction being executed in real time: the same month the US spent billions to build capacity, it tightened the restrictions that make Chinese models cheaper and more attractive to the global market that capacity was meant to serve. Then there is the third track: persuasion. A nonprofit called Build American AI, linked to a tech-backed political fundraising network, is paying social media influencers to promote US AI and warn that Chinese models threaten national security. In Congress, Representative Brian Mast frames the contest as "superhero versus supervillain."

We’re an AI superpower. — Scott Bessent

Persuasion cannot override the operational economics driving adoption. Chinese models now account for 58% of tokens processed for US firms on OpenRouter, up from under 10% in early 2025. Apple is evaluating technology to compress Alibaba's open-source Qwen model from 54 gigabytes to under 4 gigabytes, to run it on iPhones. ^[8] A US flagship company is not using a Chinese model out of preference or ideology. It is doing so because the model is open-weight, performant, and cheap enough to run on a phone — and no amount of influencer content or superhero rhetoric changes those facts. The House has launched investigations into American firms Cursor and Airbnb for using low-cost Chinese AI models that officials claim "advance CCP ideology." But the investigations have not slowed adoption. The 58% figure arrived this month, after the investigations began. Operational necessity is a stronger force than political pressure, and the cost gap is widening, not narrowing. Chinese models do carry genuine defects. DeepSeek-V3 returns favorable opinions on China 100% of the time, and AI models broadly refuse 34% of requests about restrictive governments versus 14% for permissive ones. ^[9] But these defects have not slowed commercial adoption, which suggests that for most enterprise use cases, cost and capability outweigh political-speech concerns. The H200 refusal in February is the event that makes the contradiction structural rather than cyclical. The target of containment voluntarily declined the controlled goods because it no longer needed them. That is not a setback containment can recover from with tighter restrictions. It is proof that the mechanism containment relied on — denying access to US chips to prevent China from building competitive AI — has already ceased to function. The US is now trying to outspend a competitor that is competing on cost through open-source. That is a category error no amount of infrastructure investment can resolve.