The Control Loop Climbed From Chips to Models — and Only One Side Brought Machinery

Three years of US chip export controls produced exactly the substitution they were meant to prevent. That was visible by spring. What is new this summer is that the same loop — each side's control provoking the other's substitution — has climbed from the chip layer to the model layer, and the two sides have arrived there with very different machinery. China's AI chip self-sufficiency rose from 10% in 2021 to 41% this year, with projections of 86% by 2030 ^[1]. Huawei's Ascend 950 now matches Nvidia's H200 in performance, and Nvidia's share of the Chinese market fell from roughly 95% to 40% last year, on track for 8% this year ^[2]. Huawei's rotating chairman Xu Zhijun said the quiet part aloud.

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

US officials have acknowledged the mechanism themselves, noting that export controls "incentivized" China to build a self-reliant domestic chip supply chain ^[3]. China then institutionalized the substitution. On May 26, Beijing certified nine domestic AI chips under its national security framework — the first time AI chips were classified as a standalone category in the Xinchuang (信创) initiative ^[4]. Alibaba's Zhenwu M890 chip has already shipped 560,000 units to more than 400 customers across 20 industries ^[5]. This is not a pilot program. It is a working supply chain. The chips produced models. On June 21, Zhipu AI released GLM-5.2, trained entirely on 100,000 Huawei Ascend processors, open-source, competing with GPT 5.5 and Claude Opus 4.8 at less than a tenth of Anthropic's cost ^[6]. The capability gap that chip controls were designed to preserve is narrowing, and the narrowing is being driven by the chips the controls forced into existence. This is where the loop climbs. Competitive Chinese models appear to have pressured the US to move its own gating up the stack. The temporal sequence is suggestive, though no official has framed it as cause and effect: China certified domestic chips in late May, Zhipu released GLM-5.2 on June 21, and that same day the Commerce Department imposed an export ban on Anthropic's Mythos and Fable 5 models, requiring the company to suspend access for all foreign nationals after designating them national security risks ^[7]. Days later, the Trump administration established a voluntary pre-release review system for frontier models, with OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI all committing ^[8]. China then mirrored the model-gating, and then some. On July 7, Beijing began restricting overseas access to advanced AI models under national security law, limiting API access and blocking model-weight downloads, explicitly citing Trump's denial of foreign nationals access to Anthropic models as precedent ^[9]. The State Council's regulations, effective July 1, provide a legal framework to unwind completed overseas AI transactions, ban unauthorized cross-border talent transfers, and require official sign-off before firms like ByteDance, Moonshot AI, and StepFun accept American capital ^[10]. On July 8, China layered on additional restrictions, designating model leaks and theft as national security offenses, while selectively allowing limited Nvidia H200 imports for top firms including DeepSeek ^[11]. Both sides are now gating models. But the resemblance ends there. China's model-gating is systematic. The State Council regulations extend a years-long institutional architecture — Xinchuang certification, national security chip designations, talent travel restrictions, investment review — into the model layer. Each piece connects to the next. The US model-gating, by contrast, is improvisational and self-contradictory. Trump initially canceled a planned AI executive order to avoid slowing the US lead ^[12].

We’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead. — Donald Trump

Then his Commerce Department imposed the Anthropic ban. Then Trump said he no longer views Anthropic as a national security threat ^[7]. The Pentagon simultaneously designated Anthropic a supply chain risk and terminated its $200 million contract after CEO Dario Amodei raised concerns about autonomous weapons, entangling model-gating with a military procurement dispute rather than a coherent export-control framework ^[12]. The US has built a model-level gate, but it is a gate that swings open and shut depending on who is in the room. One leg of the loop is also missing. If Chinese substitution at the chip layer was driven by US chip controls, you might expect US substitution at the chip layer to be driven by Chinese model controls. It is not. OpenAI and Broadcom have unveiled the Jalapeño custom inference chip, and Anthropic is exploring custom 2nm chips with Samsung ^[13][14]. These moves are about cost and vendor independence from Nvidia, not about Chinese model restrictions. Broadcom's Hock Tan framed the logic plainly.

any company serious about leading in AI should not depend on third-party GPUs for such a critical part of its stack. — Hock E. Tan

DeepSeek is developing its own inference chips too, driven by both US export controls and Chinese government pressure for indigenous alternatives ^[15]. The two sides are converging on the same hardware strategy — owning inference silicon — but for different reasons. China's is a control response. America's is a margin response. The asymmetry tells you what is likely to hold. China's model-gating is built into law, regulation, and procurement requirements, with enforcement mechanisms that extend to talent mobility and capital flows. The US model-gating is, so far, a series of ad hoc executive actions that have contradicted each other within weeks. If the model layer becomes the primary terrain of AI competition — and the chip layer is increasingly a solved problem on the Chinese side — the question is whether Washington can build institutional machinery to match what Beijing has already codified, or whether it will keep improvising while China systematizes. The deeper open question is whether model-level gating can work at all when the models are open-source. Alibaba's Qwen already captures more than half of global open-source downloads with 600 million users, and Singapore has shifted to Qwen for sovereign AI needs ^[16]. If a model's weights are already public and globally distributed, designating it a national security asset after the fact may be an unenforceable control — the one layer of the stack where the gate has no latch.