Export Controls for AI Chips

Modern AI programs need special chips to perform their calculations, chips that can carry out many computational steps simultaneously. These chips are expensive and are manufactured by only a handful of companies, chief among them the US corporation Nvidia. Export controls are laws that ban the sale of such goods to certain countries or make it dependent on government approval. Since October 2022, the United States has repeatedly issued such rules targeting exactly these computing chips. The goal is to prevent China from obtaining hardware with which it could build militarily usable AI systems on a large scale. Affected are not only the chips themselves, but also the machines used to manufacture them.

The dispute over access to computing power

Anyone who wants to build a very powerful AI model doesn't need just one chip, but tens of thousands. They are networked together in large data centers and train a single model for weeks on end. Computing power is thus the most important raw material of AI development. Whoever controls access to it indirectly controls how quickly others can advance.

This is precisely why export controls have become a central issue in global politics. The United States views AI as a military technology, relevant for example to surveillance, drones, or cyberattacks. China, in turn, regards the restrictions as an attempt to slow its economic development. Both sides have since taken countermeasures, for instance regarding the export of rare metals.

For investors, these rules have concrete consequences. For years, China was a multi-billion-dollar market for Nvidia, one that has shrunk considerably due to the restrictions. That's why every new regulation out of Washington moves the stock prices of the chip industry.

Thresholds, licenses, and workarounds

The rules operate using technical thresholds. What's measured, for example, is how many computing operations a chip can perform per second and how quickly it exchanges data with neighboring chips. If a product exceeds the threshold, it may only be exported with a license. This license is granted or denied by a US agency, the Bureau of Industry and Security.

What matters is the broad reach of these laws. They also apply to companies outside the United States if those companies use American technology or software. Since practically every chip factory in the world depends on US-made components, the rules have an almost global effect. Experts call this principle extraterritorial reach.

In practice, this leads to a game of cat and mouse. Nvidia developed scaled-down special models such as the H20, which fell just below the thresholds. In response, authorities lowered the thresholds or began requiring licenses for these variants as well. On top of this comes smuggling through third countries and the renting of computing power in foreign data centers, both of which make enforcement even more difficult.

What shows up in the news

Export controls are most often mentioned in news about Nvidia and other chip companies. Typical headlines involve new blacklists, slashed revenue forecasts, or exemptions granted to individual countries. Companies like the Dutch firm ASML also appear regularly in such reports, since they supply the lithography machines used in chip manufacturing.

A second thread involves news from China itself. There, Huawei, SMIC, and other companies are working on their own AI chips in order to become independent. Whether the controls are actually effective is therefore a matter of debate: some analysts see China being held back, while others see the controls as precisely what is accelerating the domestic chip industry.

In everyday life, these rules are felt indirectly. Which AI services are cheap and quickly available depends on where enough computing power is located. One should not confuse export controls with tariffs: tariffs make goods more expensive, while export controls either ban a sale outright or tie it to a required permit.