The Export-Control Squeeze

The United States has spent four years trying to deny China the compute needed to train frontier AI. The effort has been serious, and its second-order effects have been the opposite of what was intended. A short timeline explains why.

October 2022. The Commerce Department’s first sweeping controls restricted exports of top training chips, the A100 and H100 class, to China, aiming to choke off advanced compute.^[1] Nvidia responded within weeks, launching the throttled, China-specific A800 and then the H800 to stay just under the thresholds.^[2]

October 2023. Washington tightened the test to performance density, and banned the A800 and H800 workarounds. Nvidia built down again, to the H20 and its siblings.^[2]

2025. In April, a new licensing requirement effectively halted H20 sales, forcing Nvidia to take a $5.5 billion inventory charge; by July, the policy reversed and sales resumed.^[2] Then came the unprecedented part: Nvidia and AMD agreed to remit 15% of their China AI-chip revenue to the US government in exchange for export licences.^[3]

The policy oscillated between a ban and a tax. Neither stopped the chips.

The controls leak, and accelerate

Enforcement is porous. The Financial Times found at least $1 billion of Nvidia chips were smuggled into China in the three months after the April 2025 curbs.^[4] And Beijing has begun pushing the other way: Chinese regulators raised security concerns about the H20 and discouraged local firms from buying it, steering demand toward Huawei and domestic silicon.^[5]

The deeper dynamic is reflexive. Each tightening does two things at once: it raises the near-term cost of frontier compute for Chinese labs, and it raises the long-term certainty that a domestic alternative will find a market. Stockpiling ahead of each rule, the reason China’s 2024 chip imports hit records, buys time; the guaranteed demand does the rest. A control designed to deny a capability instead subsidises the will to replace it. That is not an argument that the controls are ineffective; it is an argument that their second-order effect is precisely the domestic build-out they were meant to forestall.

What the numbers imply

Export policy has become the single biggest exogenous variable in China’s AI market, and its net effect has been to guarantee a domestic customer base for domestic chips. Each tightening cycle transfers demand from Nvidia to Huawei, Cambricon and their peers, and hardens Beijing’s resolve to build a parallel stack. We underwrite that reflexivity directly: the more durable the controls, the more durable the domestic-substitution thesis they were meant to prevent.