The AI Race Is Now a Mutual Siege

Last month the Pentagon designated Anthropic — an American company — a "Supply-Chain Risk to National Security." The reason: CEO Dario Amodei refused to let the company's Claude model enable mass surveillance and autonomous lethal warfare. Defense Secretary Pete Hegseth made the administration's position plain.

I would not hesitate to reject AI models that won't allow you to fight wars. — Pete Hegseth

The blacklisting cleared the way for Musk's Grok to receive classified military clearance instead — the Pentagon was shopping for a more compliant AI, and it found one ^[1]. The same administration frames the entire AI competition with China in moral terms. In April, OpenAI submitted formal testimony to Congress casting the contest as one between political systems.

OpenAI believes the best defense is offense: the best way to ward off a fast-oncoming PRC making headway around the world for autocratic AI is continued investment in American AI leadership and global adoption of responsibly developed, democratic AI. — OpenAI

The State Department circulated a global diplomatic cable accusing Chinese firms of industrial-scale model theft, and the House passed sanctions legislation ^[2]. A nonprofit linked to a tech-backed political network began paying social media influencers to warn Americans about Chinese AI as a national security threat ^[3]. The message was consistent: the United States stands for democratic AI; China builds tools for autocracy. Then look at what China is saying. In March, the Defense Ministry warned the United States against "Terminator"-like dystopian military AI and called for human primacy in weapon systems ^[4].

A dystopia depicted in the American film The Terminator could one day come true. — Jiang Bin

Two months later, Beijing certified nine domestic AI chips — from Huawei, Alibaba, Biren, and others — under a national security and reliability framework, creating a government procurement guide for replacing foreign hardware and mandating that domestic chipmakers source 50 percent of equipment from Chinese suppliers ^[5][6]. The same government that lectures Washington on responsible military AI is building a state-directed chip industry explicitly walled off from foreign supply. Same sermon, opposite pulpit. The symmetry is not accidental. It is the product of a causal chain that Huawei's rotating chairman stated without equivocation.

If the US hadn't forced our country, our companies, and our industry, we wouldn't have done something like this. — Xu Zhijun

US export controls, designed to keep advanced chips out of China, triggered the very self-reliance they were meant to prevent ^[7]. What has emerged is a reciprocal containment regime — a mutual siege in which each side now restricts the resources the other needs most, and both have extended those restrictions far beyond chips. The US blocks what enters China. The House Foreign Affairs Committee advanced the MATCH Act, which would require Japan and the Netherlands to align chip-equipment restrictions with US rules within 150 days or face unilateral enforcement ^[6]. The administration launched Pax Silica, a 35-nation coalition to secure the AI supply chain, complete with a $50 million AI-powered cargo verification platform ^[8]. And it suspended advanced AI model access — Anthropic's Mythos and Fable — even for trusted partner India, prompting Indian officials to warn about "abrupt cutoffs" ^[8]. China blocks what leaves. In May, Beijing imposed sweeping AI export controls: it blocked Meta's $2 billion acquisition of the AI firm Manus, barred AI founders from departing the country, required government sign-off before Chinese AI firms accept American capital, and granted itself legal authority to unwind completed overseas transactions ^[9]. It restricted ByteDance, Moonshot AI, and other top firms from taking US investment without approval — a direct response to the Meta-Manus deal ^[10]. And it labeled the practice of moving AI company headquarters to Singapore to evade Beijing's review as "Singapore-washing," a direct mirror of US accusations about Chinese firms laundering trade routes ^[9]. The architecture is a mirror image. The US restricts what enters China; China restricts what leaves. Both have now extended those restrictions to capital, talent, model access, and third countries. Neither is merely defending a technological lead — each is trying to deny the other a complete stack. And the race is shifting beneath both of them. Stanford's 2026 AI Index found the US-China model performance gap has narrowed to near-parity, with Anthropic holding only a 2.7 percent lead as of March ^[11].

The US-China AI model performance gap has effectively closed. — Stanford University

Chinese AI models have outpaced US models in global weekly token usage for five consecutive weeks — 12.96 trillion tokens against 3.03 trillion — and all six of the world's most-used AI models are now Chinese, driven by pricing at roughly one-tenth of US costs ^[12].

Users apply clear criteria: first, the model must be intelligent, user-friendly and responsive, and capable of solving complex problems; second, pricing must be reasonable and commercially sustainable. — Minimax

China's AI chip self-sufficiency rose from 10 percent in 2021 to 41 percent in 2026, with Morgan Stanley projecting 86 percent by 2030 ^[7]. DeepSeek's V4 model was rewritten to run on Huawei Ascend 950 chips and is priced up to 35 times cheaper than OpenAI's GPT-5.5 ^[13]. Alibaba has already delivered 560,000 of its Zhenwu-series AI chips to more than 400 customers across 20 industries ^[14]. The contest is no longer about who has the best frontier model or the fastest chip. It is about who can build a complete self-reliant stack fastest — and both sides' restrictions are accelerating the very independence they were designed to prevent. The outcome is genuinely uncertain, and the honest limits of the evidence make that clear. Zhou Hongyi, founder of China's 360 Security Technology, openly acknowledges a 20 to 30 percent base capability gap for Chinese AI models due to US hardware restrictions ^[15].

Objectively speaking, domestic models still have a 20%-30% gap in base capability. — Zhou Hongyi

China's LineShine supercomputer reclaimed the world's fastest ranking using a CPU-only architecture with domestic Huawei processors, but it ranked only fourth in AI-specific benchmarks and burns 42 percent more power than the US El Capitan system ^[16].

China can adapt to develop its own version of technology as good as – or maybe even better than – existing technology, despite U.S. export controls. — Jack Dongarra

And 40 percent of US AI data center projects may miss their 2026 completion dates, delayed by labor shortages, equipment bottlenecks from tariffs on Chinese transformers, and utility infrastructure that cannot keep pace ^[17]. The uncertainty does not come from one side winning. It comes from both sides running the same siege against each other, each claiming the other is the threat to responsible AI, each shopping for the most compliant weapons system it can find. The Pentagon blacklists an American AI company for insufficient military compliance while the Defense Ministry in Beijing certifies domestic chips under national security law. Two powers, each preaching restraint, each racing to arm.