The U.S. government has imposed new export-license requirements on some of the world’s most powerful artificial-intelligence processors, restricting shipments of Nvidia and Advanced Micro Devices chips to China and escalating Washington’s effort to prevent advanced computing technology from reaching military users.
Nvidia disclosed that the government informed it on August 26 that any future export to China, including Hong Kong, of its A100 and forthcoming H100 integrated circuits would require a license. The company said in its quarterly filing that the requirement also covers A100X products, DGX systems incorporating the chips and future processors meeting specified performance and chip-to-chip interconnect thresholds. Nvidia said officials cited the risk that covered products could be used in or diverted to a military end use or military end user.
The controls target the engines behind large-scale AI
The affected products are not ordinary graphics cards. Nvidia’s A100 is a data-center accelerator designed for machine-learning training, inference, analytics and scientific computing. When the company launched the architecture in 2020, it described the A100 Tensor Core GPU as capable of accelerating AI training and inference by as much as 20 times compared with its previous generation, with systems designed to scale across cloud and supercomputing installations.
The H100 is the next generation. Nvidia announced the Hopper architecture in March, saying the H100 contains 80 billion transistors and is designed for large language models, recommendation systems, genomics and other computationally intensive workloads. The processor is not yet broadly available, which means the export restriction reaches a product before its commercial rollout is complete.
That timing creates an unusual engineering problem. Nvidia said the license rule could affect its ability to complete H100 development on schedule or support existing A100 customers and could force some operations to move out of China. The company’s initial Form 8-K disclosure described the restriction as effective immediately and said it would seek licenses where necessary.
Nvidia says $400 million in China sales are exposed
The financial impact could be material even before the H100 reaches volume shipments. Nvidia said it had approximately $400 million in potential third-quarter sales to China that could be affected if customers do not buy alternative products or if the government declines to issue licenses. The figure reflects the importance of Chinese demand for data-center accelerators used by cloud providers, internet companies, research organizations and other large computing customers.
Washington has not publicly released a detailed rule explaining all of the performance thresholds behind the new license requirement. In contemporaneous reporting, the Guardian, citing Reuters, reported September 1 that the Commerce Department said it was reviewing China-related policies and taking additional steps to keep advanced technologies from uses inconsistent with U.S. national-security and foreign-policy interests.
The controls also apply beyond Nvidia. AMD told Reuters that it had received new licensing requirements affecting its MI250 accelerator, although the company did not believe the MI100 would be covered. AMD launched the broader Instinct MI200 series in November 2021 for high-performance computing and AI, including the MI250X used in exascale-class systems.
The distinction among products indicates that Washington is trying to draw a technical line around the highest levels of computing performance rather than restricting every data-center processor. The government’s concern is that accelerators powerful enough to train large AI systems can also support military modeling, intelligence analysis, image recognition and other dual-use applications.
A temporary authorization reduces disruption to H100 development
After the initial disclosure, the government granted Nvidia limited authorizations designed to reduce immediate operational disruption. In a September 1 Form 8-K, Nvidia said it may continue exports and transfers required to develop the H100, support U.S. customers of the A100 through March 1, 2023, and conduct A100 and H100 order fulfillment and logistics through its Hong Kong facility through September 1, 2023.
Those authorizations do not remove the underlying restriction on sales to covered Chinese customers. They instead preserve specific development and logistics functions while the broader licensing requirement remains in place. For Nvidia, that distinction is important because sophisticated chips depend on globally distributed engineering, manufacturing, testing and distribution systems even when the final customer is located in one country.
The H100 is especially significant because Nvidia has positioned it as the foundation for a new generation of AI infrastructure. The company’s March DGX H100 announcement said eight H100 GPUs in a single system could deliver 32 petaflops of AI performance at new FP8 precision, with larger clusters aimed at massive language models and other workloads that require thousands of processors operating together.
The policy targets capability rather than a specific application
The export decision reflects a broader problem facing technology regulators: the same hardware can support commercial research and military applications. A high-end accelerator can train a consumer recommendation model, analyze medical imaging, simulate physical systems or process satellite imagery. The chip itself does not reveal the purpose for which it will ultimately be used.
Licensing therefore gives the government a mechanism to review particular exports rather than relying solely on a categorical prohibition. Companies can seek authorization for transactions, and regulators can consider customers, end uses, destinations and other factors. But the process also introduces uncertainty for manufacturers and customers because a planned sale may depend on a government decision that is not under the company’s control.
The new requirements may have effects beyond near-term revenue. Chinese technology companies could accelerate efforts to develop domestic alternatives, while U.S. manufacturers may need to redesign products or supply chains to serve markets that fall below controlled thresholds. Cloud providers operating across borders will also need to understand whether access to advanced accelerators through remote computing services raises additional policy questions.
A significant escalation in the technology competition with China
Washington has used export controls for years to restrict specific Chinese companies, telecommunications equipment and military-linked end users. The new action is different in scale because it reaches a class of advanced computing hardware based on technical capability and applies to products from multiple major U.S. chip designers.
That makes artificial-intelligence computing itself a more explicit national-security concern. Nvidia’s A100 and H100 and AMD’s MI250 are valuable precisely because they can perform enormous numbers of calculations in parallel, allowing researchers and companies to build increasingly capable models and simulations. The same performance advantage that creates commercial value is what makes the processors strategically sensitive.
For Nvidia, the immediate challenge is operational: protect H100 development, support existing customers, apply for licenses and determine how much of the approximately $400 million in at-risk quarterly sales can be preserved. For AMD, the MI250 restriction creates a similar need to identify affected customers and transactions.
For the broader technology industry, the significance is larger. Advanced AI hardware has moved from being treated mainly as a commercial product to being treated as a strategic capability subject to national-security controls. The exact boundaries of that policy are still emerging, but this week’s licensing requirements establish that the performance of a processor can now determine not only what systems it can power, but where it can legally be sold.
