Big tech is delighted to rent American chips to Chinese companies banned from buying them

One of the defining geopolitical struggles of the last few years has been America’s attempts to limit Chinese access to advanced artificial intelligence chips through export bans. China is America’s biggest rival, the world is in an AI arms race, and America doesn’t want China to take the lead.

For private companies that produce AI chips, however, China doesn’t represent a geopolitical enemy. It represents a customer base. Despite current export bans on chip sales to China, American companies have still found ways to sell “access” to Nvidia’s AI chips to Chinese companies. From The Information:

Both Google Cloud and Microsoft Azure are offering to rent Nvidia’s AI chips to Chinese companies, including AI startups, for use in data centers outside China. But apart from the biggest U.S. tech giants, there is a whole sector of smaller cloud providers specializing in offering access to Nvidia-powered servers around the world, and their services are available to Chinese customers. Some of these cloud providers are based in the U.S., but numerous others are based in Europe and Asia.

Last summer, for example, Google Cloud’s Asia Pacific team contacted a prominent Chinese startup that develops large-language models and offered to rent servers in Europe with Nvidia’s A100 and H100 chips, according to a person with direct knowledge of the talks. U.S. rules block the export of both kinds of chips to China. The approach didn’t lead to a deal.

Microsoft also offers its Nvidia-chip server rental services, including servers with A100 and H100 chips, to Chinese customers through data centers outside China, according to a Microsoft employee with knowledge of the services and a person directly involved in the sales.

In 2019 and 2020, the United States added Huawei and SMIC to a restricted entities list to limit their ability to design chips that rivaled those produced by western companies like Nvidia. This decision was manufacturing-based: the US government didn’t want China to leverage western technology to create its own powerful chips that could be used by its military.

However, since the generative AI boom kicked off with the launch of OpenAI’s ChatGPT in November 2022, the export ban became more complicated. Training a large language model requires a tremendous amount of computing power, and Nvidia’s chips are widely seen as the most powerful on the market. While US export bans still hinder Chinese manufacturers’ abilities to improve their own chips, Chinese tech companies looking to train their own LLMs don’t need physical possession of Nvidia chips. They just need access to servers with Nvidia chips, regardless of where those servers are located.

US Commerce Secretary Gina Raimondo has noted that the US needs to block this practice to prevent China from “training their frontier models,” but for now, the secondary cloud rental market is wide open.