The US restricts advanced chip exports to China

Training a large model is a hardware problem before it is anything else. The systems released in 2022 were trained on thousands of specialized processors run for weeks, and a small number of firms outside China designed and manufactured the processors that could do it. Two months earlier, Congress had addressed one side of that dependency by funding domestic manufacture through the CHIPS and Science Act. The October rule addressed the other side, by restricting what could be sold abroad.

On 7 October 2022 the Bureau of Industry and Security, part of the Commerce Department, issued an interim final rule covering three things: advanced computing chips, the equipment used to manufacture semiconductors, and the technology and services associated with both. Sales of those items to customers in China now required a license, and the rule set out that licenses for many of them would be presumed denied. The provisions were phased in over two weeks, taking effect on 7, 12, and 21 October. The stated purpose was to limit China’s access to technology that could support military modernization, a category the department described as including advanced AI systems.

The controls were written around categories of hardware and capability rather than a list of named companies, which meant they applied to any firm shipping into the restricted categories, wherever it was based. Chipmakers, equipment vendors, and their customers redrew supply plans around the rule through the rest of the year, and the government revised and extended it in the years after its issue.

The rule turned computing hardware into an instrument of foreign policy. From October 2022 the question of who could train the largest models was no longer settled only by budgets and engineering, but also by export licenses.