China, chips and export controls

This thread follows the contest over who can build frontier AI, fought largely through the supply of the advanced chips that training requires. It opens with the US October 2022 export controls, which cut China off from the most advanced processors and the equipment to make them, and which Washington tightened again in 2023 and extended to 140 more entities in 2024. The controls’ premise — that compute is the choke point — was itself contested, argued at length in Situational Awareness.

Then the premise was tested. DeepSeek’s V3 and R1 matched leading US models at a fraction of the reported training cost, and markets read it as evidence the controls had spurred efficiency rather than prevented progress: Nvidia lost a record amount of market value in a single day. The policy response then oscillated. The Biden administration’s AI Diffusion Rule was rescinded months later; H20 sales to China were restricted, then allowed to resume under a revenue-sharing arrangement handing the government 15% of the proceeds.

By 2026 the Chinese labs had partly routed around the constraint — Zhipu trained GLM-5 entirely on Huawei Ascend chips — even as Anthropic accused several of them of distillation attacks on Western models. The thread’s recurring question is whether export controls slow a rival or teach it to need less of what is being withheld.