Controlling GPU exports to China, Now the H20

Blindsiding Nvidia to the tune of $5.5 billion and much more was not a good look

This week it became apparent that the new Trump administration intended to step up controls around already controlled GPUs…here is a quick look at how this has unfolded and the implications…

Intel, AMD, and Nvidia each released different details about the “is informed” letter they had received from the Bureau of Industry and Security (BIS) at the Commerce Department, likely signed off on by Undersecretary of Commerce Jeff Kessler. Nvidia mentioned the supercomputer/diversion justification, AMD did not mention supercomputers, and Intel revealed the performance parameters that are now operative for controlling GPUs, which jack up a lot of interconnectivity and memory bandwidth numbers. "Is informed” letters are not public, but because of the material impact on company’s financial situation, they are required to reveal what the impact may be once they become aware of the letters.

“As a company that has been deeply involved in the Chinese market for 30 years, we have grown together with the Chinese market and achieved mutual success. China is not only one of the world's largest consumer markets, but its booming industrial ecology and leading software strength have become an important driving force for our continued innovation. Our successful experience in the Chinese market has driven us to continuously increase our R&D investment, and our in-depth cooperation with Chinese companies has also enabled us to grow into a more competitive international company. Therefore, we will continue to spare no effort to optimize our product system that meets regulatory requirements and unswervingly serve the Chinese market."—Jensen Huang

The fact that Nvidia was blindsided by the action is a very very bad look here for the US government. This is not some rogue company, Nvidia has twice, and a great cost, already re-engineered its GPUs for the China market to comply with US export controls. This is a company that the US government is also counting on to be the core player in building out US AI infrastructure over the next decade. But at the same time, someone thought a $5.5 billion write off and 10s of billions of revenue was a good idea for the US government to cause for Nvidia’s business model because of concern over diversion of H20s for “supercomputer use?” There is no evidence I have seen, and I follow the issue closely, that H20s are being used as accelerator in Chinese supercomputers. Chinese companies developing AI models now seem the H20 as a useful GPU to do inference, and that is why Baidu, Alibaba, Tencent, and Bytedance put in orders for around $16 billion or H20s. I.e., not for “supercomputer use.”

On the H20, the “is informed” letters form BIS had to be signed off by the Undersecretary, here Kessler, but it remains unclear how much visibility where was on this issue at higher levels, within the White House, and the inner circle. The optics of a sudden and unexpected $5.5 billion hit on Nvidia remain pretty bad, and not certain if anyone at the White House asked about what was going on here, particularly after the NPR story last week suggested that CEO Jensen Huang’s attendance at a 1 million a plate dinner at Mar-a-lago had somehow bought Nvidia a reprieve on the H20 —I was very suspicious of this story, as there were no details and not clear who was at the meeting. The WSJ reported today that NSC officials met with Huang and Nvidia several weeks ago to talk about the H20. It is not clear who was at the meeting, and it sounds like it occurred after the firing of Senior Director for Technology and National Security David Feith. With Feith out, who actually made the decision to move the H20, and what as communicated to Nvidia at this meeting? The entire thing is a really really bad look. Then CEO Huang headed to Beijing, held talks with VP He Lifeng, and DeepSeek CEO Liang Wenfeng! With Liang, Jensen apparently talked about how Nvidia could still make GPUs for China that comply with US export controls? And this leaks to the media? This all seems just a bit nutty from end to end.

The Commerce Department will soon release a new rule to cover the new licensing requirement for exporting the H20 to Chinese end users.¹ There will almost certainly be a “presumption of denial” for these licenses, given that the language used in the Nvidia K-8 suggested that the Commerce Department was concerned about the potential for diversion of the H20 to particular end users. While the justification given in the letter to Nvidia by Commerce related to the use of the H20 in supercomputers, in fact that real reason related to the use of the H20 for inference for running advanced AI applications.

Details from the K-8 filings

Nvidia: On April 9, 2025, the U.S. government, or USG, informed NVIDIA Corporation, or the Company, that the USG requires a license for export to China (including Hong Kong and Macau) and D:5 countries, or to companies headquartered or with an ultimate parent therein, of the Company’s H20 integrated circuits and any other circuits achieving the H20’s memory bandwidth, interconnect bandwidth, or combination thereof. The USG indicated that the license requirement addresses the risk that the covered products may be used in, or diverted to, a supercomputer in China. On April 14, 2025, the USG informed the Company that the license requirement will be in effect for the indefinite future.

Intel: Intel’s K-8 provided details of the actual performance parameters that BIS was altering to capture the H20: a total DRAM bandwidth of 1400 GB/s or more; I/O bandwidth of 1100 GB/s or more; or a total of both of 1700 GB/s or more.

AMD: On April 15, 2025, Advanced Micro Devices, Inc. (the “Company”) completed its initial assessment of a new license requirement implemented by the United States government for the export of certain semiconductor products to China (including Hong Kong and Macau) and D:5 countries, or to companies headquartered or with an ultimate parent in such countries (the “Export Control”). The Export Control applies to the Company’s MI308 products. The Company expects to apply for licenses but there is no assurance that licenses will be granted. The Company expects that the Export Control may result in charges of up to approximately $800 million in inventory, purchase commitments and related reserves.

Implications, and there are many

There are few alternatives to Nvidia GPUs in China, and leading Chinese AI companies, including Bytedance, Baidu, Tencent, and Alibaba, all already have a lot of H20s in their GPU inventory, and were planning to purchase collectively up to $16 billion more H20s to expand their inference capacity for corporate clients and consumers. There is no ready substitute for the H20 for inference in China at scale. The move will force more Chinese AI developers to consider using Huawei Ascend 910B/C based systems for inference as they add capacity beyond existing stocks of H20s and other Nvidia GPUs.

The move by the Commerce Department will mean that Huawei’s Ascend 910C GPU will now become the hardware of choice for AI model developers and for deploying inference capacity. Many of the major AI developers still have large stockpiles of H20s and so they will also be used for inference, as companies deal with a heterogeneous hardware base and Huawei and domestic foundry leader SMIC ramp up production to meet the growing domestic demand for AI hardware.

Huawei faces significant constraints producing its Ascend series GPUs at domestic foundry leader SMIC. SMIC is under heavy export controls targeting its capacity to manufacture at the advanced nodes, 7 nm and below, that are required for advanced GPUs. Huawei was able to acquire several million GPU dies for its Ascend design from global foundry leader TSMC by using a cutout before this was discovered and stopped late last year. So Huawei will continue to package these dies with high bandwidth memory (HBM) likely from Samsung and this will provide somewhat of a buffer that can satisfy some level of domestic demand while SMIC attempts to improve its yields, and pushes its existing foreign tools to levels of performance that no company has done before. Ironically, it appears that the H20E version Nvidia sold to Chinees customers used Samsung HBM3-e memory.

US controls on GPUs and semiconductor manufacturing equipment have primarily damaged US companies, including Nvidia and all US toolmakers, while having marginal impact on the ability of Chinese companies to develop frontier AI models, the justification given for applying the controls. The latest justification given for the H20 may be the most weak, of the many justifications included in Commerce Department documents since October 2022, as there is no evidence that any Chinese firms are using H20 GPUs that turn out to be optimal for inference for “supercomputers”, the justification given to Nvidia by the Commerce Department.

Costs piling up

In general, the unprecedented level of unilateral, extraterritorial export controls that started in October 2022, without any cost benefit analysis, have been hugely disruptive of the global semiconductor industry, damaged the competitiveness of leading US technology companies, undermined long-term R&D efforts, and encouraged the Chinese semiconductor industry to become more innovative, while designing out US technology. I have called this a “small gain, high cost” policy, instead of the stated “small yard high fence” approach as articulated by senior officials of the Biden administration.

Chinese firms have also found many workarounds for gaining access to advanced GPUs, including through diversion, and sustained access to GPUs at data centers outside China. This effort will continue and deepen across the board, and the US Commerce Department lacks the ability to enforce complex measures such as the AI Diffusion Framework, released late in the Biden administration, as an attempt to control where AI data centers are built around the world.

Finally, perhaps the biggest failure of the export controls, including GPUs, culminating in the H20 and the AI Diffusion Framework, is that their launch has almost certainly undermined any attempt to reach a level of collaboration between the US and China on the vital issue of AI safety, an issue that is increasingly salient as models become more capable. The risks are high, and the only way to forge a global agreement about how to regulate and control the most advanced models will be a process that sees both the US and China, whose companies are global leaders in AI development, sit down and agree on the terms of a global AI Safety Framework. The export controls are working in a manner diametrically opposed to moving towards collaboration on this vital issue, and with some predicting the emergence of artificial general intelligence within the next 2-3 years, this issue is even more critical.

1. Readers may remember the August 2022 K-8 from Nvidia that revealed Commerce licensing requirements on the A100, H00 and other GPUs. The rule governing performance for these GPUs came in early October of 2022.