Export rules split the market for AI compute in two

Washington now restricts advanced AI chips and chipmaking tools bound for China. Access to compute becomes a question of geography as much as price.
Selling a top AI chip to China now takes a US licence. Two rules issued on 7 October by the Commerce Department's Bureau of Industry and Security restrict China's access to advanced computing chips, supercomputers and the tools to make leading-edge chips; from 21 October, exports of the most capable chips and of computers that contain them need a licence, including some chips made abroad with US technology.
Compute for training large models now has a passport. Access depends on where a company is based, not only on what it can pay, and frontier training concentrates in the US and allied clouds. Any investment that needs a steady supply of advanced chips carries policy risk that one rule can change.
The equipment rules cut deeper. Tools for Chinese fabs that make logic chips at 16 or 14 nanometres and below, DRAM at 18 nanometres half-pitch and below, or NAND with 128 layers or more need a licence, with a presumption of denial for Chinese-owned plants. US persons need a licence to support such fabs, and 31 entities joined the list of parties whose end use cannot be verified. China loses the means to make leading-edge chips, not only to buy them.
Each side of the line pays. These rules build on the licence requirement Nvidia disclosed in August for its A100 and H100 data centre chips, which put about $400 million of potential quarterly sales to China at risk, and US chipmakers lose part of a large market for data centre accelerators. Chinese AI companies lose the most capable training hardware, including the H100 generation that resets training cost, which raises their cost per training run and pushes them toward domestic chips or designs that stay under the thresholds.
Outside China, the picture reverses. Access to top GPUs through cloud providers stays open, and the field of well-funded competitors at the high end narrows. Founders should map their exposure, because customers, data centres, suppliers and team members in China all shape what licences a company may need as the rules evolve, and hardware startups and makers of chip design tools face licence reviews on sales they took for granted. Efficiency pays on both sides of the line, so software that trains and serves models with fewer chips becomes a stronger business.
Controls leak when allies stay out. Europe and Japan are home to other leading equipment makers, and unless their governments adopt matching controls, the split holds for chips but softens for the machines that make them. Chipmakers can also design data centre products that fall just under the new thresholds for Chinese customers. What settles it is how fast Chinese chip designers and clouds close the gap, and what Chinese AI companies pay for compute in the meantime.
Advanced compute is now a strategic resource governed by national policy. Fanaura favours teams whose compute, customers and talent sit where the rules are stable, as in the US, EU and Latin American markets it invests in, and teams that get more out of each chip than their competitors. Compute now has an address as well as a price.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building software that trains or serves models with less hardware? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.