The Divide

The Chip Curtain: Do Export Controls Slow China, or Just Cede the Market?

Two frontier-tech CEOs read the same Chinese chip data and reach opposite conclusions about whether Washington's controls are a chokepoint or an own goal.

7 min read2026-07-08

The question

Since 2022 the United States has restricted the export of the most advanced AI accelerators to China, and the policy has only tightened since. The question is not whether the controls have costs — everyone concedes they do — but whether they achieve their stated aim: keeping China a meaningful step behind the compute frontier. It is live again this summer. Nvidia disclosed in its fiscal-2026 annual report that it was "effectively foreclosed from competing in China's data center compute market," and its latest quarterly filing records that Washington began, in February 2026, granting licences to ship small volumes of the cut-down H200 part into China under a 25 per cent tariff — a partial thaw layered on top of the ban.⁷ The same month, a June executive order stood up a voluntary pre-release review of frontier models, extending the security logic from chips toward the models themselves.⁹

The disagreement is unusually clean because it has two named, credentialed principals who have made their cases in public. On one side stands Anthropic's Dario Amodei, who argues in a signed essay that well-enforced controls are the decisive lever over whether the AI era is led by democracies.¹ On the other stands Nvidia's Jensen Huang, who has called the policy, in as many words, "a failure."⁵ Both are reasoning from the same facts about Chinese progress. This desk takes no side; it lays out each at its strongest.

The security camp

Compute is the one input China cannot yet make at the frontier, so controlling it is the highest-leverage move America has.

The strongest version of this case is not that controls stop China outright, but that they hold open a lead that compounds. Amodei's argument is that because AI can help build better AI, "a temporary lead could be parlayed into a durable advantage," and that "well-enforced export controls are the only thing that can prevent China from getting millions of chips."¹ The evidence that the chokepoint is real is physical. A December 2025 analysis by the Council on Foreign Relations argues that SMIC, China's leading foundry, is "stuck at 7nm" because it cannot buy EUV lithography, and that high-bandwidth-memory controls left Huawei able to turn perhaps 1.5 million dies into only 200,000–300,000 finished accelerators — concluding that "Huawei is not a threat that justifies loosening controls; it is evidence that the controls are working."²

On this reading, the leakage the other side points to is an argument for enforcement, not surrender: the response to chips reaching China through cloud rentals is to close that door. That is precisely what the House Select Committee on the CCP moved to do in June 2026, its chairman warning that "China will buy what it can and steal the rest, which is why it is actively trying to get backdoor access to U.S. data centers and train its AI models via cloud computing."³

Dario Amodei, CEO, Anthropic — standing here because he runs a frontier lab that both depends on compute and has argued, under his own name, that the chip gap decides a unipolar versus bipolar AI world.¹

The industry camp

Controls do not stop China from building AI chips; they hand it the domestic market, the motive, and the state backing to do so faster.

The strongest version of this case is empirical, not ideological. Speaking at Computex in May 2025, Huang said flatly, "I think, all in all, the export control was a failure," and that it "gave them the spirit, the energy, and the government support to accelerate their development"; he put Nvidia's share of China's AI-chip market as having fallen from roughly 95 per cent to about 50 per cent over the period.⁵ The mechanism is substitution: a market Nvidia is barred from does not vanish, it transfers. In mid-2026 Nvidia warned Washington that "if Nvidia is kept out of the China market, Huawei production and sales will soar," and that "the ability of China's domestic ecosystem to yield and scale competitive AI systems is a fact."⁴

The cost is not hypothetical. Nvidia's own filings record a $4.5bn charge against stranded H20 inventory when the rules tightened, and describe the Chinese data-center compute market as one it is now "effectively foreclosed" from — a forfeited position that a domestic champion is being paid, in effect, to fill.⁶

Jensen Huang, CEO, Nvidia — standing here because he runs the company the controls bind most directly and competes head-on with the Chinese supplier they are meant to hobble.⁵

The crux

Strip away the rhetoric and the two sides are not arguing about values; they are arguing about a forecast. Both agree China is racing to build domestic accelerators and that it is, for now, constrained. They disagree about which force wins the next two years: the physical bottleneck, or the substitution it triggers. The bottleneck is well documented — SMIC at 7nm without EUV, a memory shortage that strands most of Huawei's dies, and the CFR's striking claim that "Huawei's next-generation chip in 2026 will actually be less powerful than its best chip today."² The substitution is equally documented: Huawei plans to roughly double output to about 600,000 Ascend 910C chips in 2026, has laid out a three-generation roadmap, and has committed some 15 billion yuan a year to its software ecosystem.¹⁰

The leakage that sits between the two cases is what makes the forecast hard. A CSIS analysis notes that DeepSeek trained a leading model on export-compliant H800s for roughly $5.6m in final-run compute, that Huawei stockpiled more than two million TSMC-made dies before restrictions bit, and that smuggling networks move H100s in $100m-plus lots — so controls have "slowed but not stopped" China, with their largest effect still prospective.⁸ If the bottleneck holds, controls buy exactly the compounding lead Amodei describes. If substitution and leakage win, they buy Huang's own goal: a forfeited market and a subsidised rival.

What would change your mind

This is a forecast that reality will grade, and the grading has dates. Watch Huawei's actual 2026 output against its ~600,000-unit target: hit near it, and the substitution case strengthens; miss badly on SMIC yields or memory supply, and the bottleneck case does.¹⁰ The CFR's most falsifiable claim — that Huawei's 2026 flagship will be weaker than its 2025 one — is a clean test; if the next Ascend part benchmarks ahead of the current one on domestically produced silicon, the "controls are working" thesis loses its sharpest piece of evidence.²

Two more signals would move a fair reader. First, whether a Chinese lab trains a genuinely frontier-scale model predominantly on domestic accelerators rather than on stockpiled or smuggled Nvidia parts; the DeepSeek precedent shows how much can be done with compliant and diverted chips, so the diagnostic is the hardware underneath the next leap, not the leap itself.⁸ Second, whether enforcement actually closes the cloud and smuggling routes the security camp is now targeting; if rented and re-exported compute keeps flowing at scale despite the new measures, the controls are porous enough that Huang's substitution dynamic runs regardless of the paper rules.³

The question

The security camp

The industry camp

The crux

What would change your mind

Sources