The Chip Controls Have Built the Thing They Were Meant to Prevent

The pattern is now visible across enough stories to stop calling it a trend and start calling it a system. Every US restriction designed to deny China access to advanced technology has produced a Chinese response that narrows the gap the restriction was meant to preserve. And in the last several months, that response has moved from the software layer to the silicon layer, while both governments have quietly built mirror-image regimes to control what flows out, not just what flows in. Start at the point where the causal chain is explicit. DeepSeek, the Chinese AI lab whose models have already matched frontier Western performance, is now designing its own inference chips. The company's founder, Liang Wenfeng, said directly that the move was prompted by US policy ^[1].

chip export controls were a challenge for the company. — Liang Wenfeng

That single admission connects the dots. DeepSeek had already optimized its V4 model to run on Huawei's Ascend processors rather than Nvidia GPUs ^[2]. Now it is recruiting chip engineers and talking to foundries, aiming to reduce dependence on both Nvidia and Huawei ^[1]. The substitution trajectory runs from software workarounds to custom silicon to full stack control, not from US supplier to Chinese supplier. And the silicon is arriving at scale. Huawei's Ascend 950 is now considered comparable to Nvidia's H200, with ByteDance, Tencent, and Alibaba all placing orders ^[3][4]. Alibaba's own Zhenwu M890 chip has already shipped 560,000 units to more than 400 customers across 20 industries, with Alibaba's CEO projecting AI-related products will be half of cloud revenue within a year ^[5]. On May 26, China certified nine domestic AI chips from seven manufacturers under its national security procurement framework, the first time AI chips were included as a standalone category, making domestic silicon mandatory for government agencies and state-owned enterprises ^[6]. The market consequence is stark. Nvidia's share of China's AI chip market has collapsed from roughly 95 percent to an estimated 8 percent ^[4]. Jensen Huang frames the loss as permanent.

Well, we were in China for 30 years, and before the export control banned Nvidia out of China we had about 95% market share, and so we were competing just fine. — Jensen Huang

95% → 8% Nvidia's China AI-chip market share, pre-controls to 2026 — The collapse is documented in Nvidia's own market reporting ^[4].

Here is where the logic of the controls inverts on itself. The original rationale was military diversion: keep advanced chips out of Chinese hands so they cannot power weapons systems. But Nvidia, the company the controls were meant to protect, now says the rationale is spent.

Just like it would be nonsensical for the American military to use Chinese technology, it makes no sense for the Chinese military to depend on American technology. — Nvidia

If the military-diversion argument is moot by the chipmaker's own assessment, the remaining justification is economic leverage, and that is eroding too. The Trump administration lifted the H200 export ban in January 2026 and authorized sales to China with a 25 percent fee, but no sales have occurred because China's government is blocking the purchases to keep investment focused on domestic chips ^[7]. The controller wanted to deny access; the controlled is now denying the access it was offered. This is where the pattern widens from chips to AI models. On June 12, the Commerce Department banned Anthropic's Fable 5 and Mythos models for all foreign nationals after a jailbreak showed the models could identify software vulnerabilities ^[8]. The ban was not a free-standing policy. It was triggered by Zhipu AI's release of GLM-5.2, a model trained entirely on 100,000 Huawei Ascend 910B processors, bypassing Nvidia at every stage. GLM-5.2 reached frontier-competitive coding performance at less than one-tenth of Anthropic's cost ^[9].

This changes things. — Guillermo Rauch

A Silicon Valley CEO saying he is shocked by a model trained entirely on Chinese hardware is the signal that the substitution has crossed a threshold. The US response, restricting model access for all foreign nationals, is itself an admission that the competitive gap has narrowed enough to require protection. Anthropic told its investors as much.

the window of opportunity to lock in that lead will not necessarily remain open for long. — Anthropic

China, meanwhile, is building the same instrument pointed the other direction. The State Council codified outbound restrictions effective July 1 that ban unauthorized cross-border talent transfers, bar senior AI professionals at Alibaba and DeepSeek from overseas travel without approval, require government sign-off before Chinese AI firms accept American capital, and authorize retaliatory probes against countries that discriminate against Chinese investors ^[10]. The NDRC blocked Meta's $2 billion acquisition of the AI startup Manus even after it relocated to Singapore, and placed exit bans on Manus executives ^[11][12]. The Ministry of Commerce met with Alibaba, ByteDance, and Z.ai to discuss restricting API access and blocking model-weight downloads ^[13]. Both sides publicly condemn the other's controls while quietly building reciprocal versions of the same thing ^[14]. The US side keeps tightening. The Bureau of Industry and Security closed the subsidiary loophole in June, clarifying that licensing requirements apply to all China-headquartered companies regardless of where a subsidiary sits, targeting Nvidia Blackwell GPU flows through countries like Malaysia ^[15]. The House Foreign Affairs Committee advanced more than 20 export-control bills, including one that would require Japan and the Netherlands to align chip-equipment restrictions with US rules within 150 days or face unilateral enforcement, and ban ASML from servicing existing equipment at SMIC and Huawei ^[16]. Anthropic separately identified "distillation attacks," Chinese firms using Claude outputs to train competing models, as a software-level workaround alongside chip smuggling and offshore datacenters ^[17]. Each of these moves is individually rational. Together they form a cycle. A restriction denies access. The denied party substitutes. The substitution narrows the gap. The narrower gap triggers a new restriction on a different layer. The cycle does not reverse, because each side has now institutionalized the posture. China made domestic-chip adoption mandatory procurement policy. The US made AI model weights export-controlled technology. China made senior AI talent travel a security question. The US made allied chip-equipment policy a domestic legislative matter. Huang, who has the strongest commercial incentive to read this clearly, puts it plainly: China's AI moves on with or without US chips, and the question is whether the world's largest AI market runs on American platforms ^[18]. The CUDA software ecosystem, the one durable lever the US had, is being replaced every time a Chinese lab optimizes for Ascend instead. The limits are real and worth naming. Huawei's domestic manufacturing still stands at 7nm, with 1.4nm-equivalence only projected by 2031, while TSMC plans actual 1.4nm by 2028 ^[19]. DeepSeek's custom chip still needs a foundry, and if that foundry is TSMC, the dependency merely moves from US chips to Taiwanese manufacturing ^[20]. China controls most of the world's rare earth processing but holds fewer than 20 percent of high-value patents for advanced rare earth materials ^[21]. These qualifications matter for how far the substitution has gone. They do not change the direction. What follows from all of this is a question neither government is asking aloud. The export-control regime was built on a premise of asymmetric leverage: China needs US chips, the US controls supply. That asymmetry is dissolving. China is building its own chips, its own models, its own outbound controls, its own investment-screening regime. The US is building the same set of instruments pointed the other way. Two self-sufficient, mutually excluding technology stacks, each governed by a control architecture that mirrors the other's, is not the decoupling the controls were designed to manage. It is the one they were designed to prevent.