Earlier this week, the U.S. government lifted its ban on shipments of AMD's Instinct MI308 and Nvidia's H20 processors for AI to Chinese entities, which pleased the companies, their Chinese customers, and investors. However, John Moolenaar (R), the head of a House of Representatives panel on China, has sent a formal letter to Commerce Secretary Howard Lutnick criticizing the move and requesting a briefing to discuss the decision and future actions of the executive branch. There is an interesting part: he does not call for a ban on the said GPUs again, but to develop a new set of rules.
"I could not agree more, which is why I strongly supported the administration's ban on H20 sales [to Chinese entities] — we must not allow U.S. companies to sell these vital artificial intelligence (AI) assets to Chinese entities," the letter signed by John Moolenaar on behalf of the panel reads. "The H20 chip significantly outperforms anything Chinese chipmakers like Huawei can currently produce, particularly in high-bandwidth memory — a critical factor for AI inference workloads."
H20 too good for China, for now
Nvidia's H20 HGX GPU for AI applications was developed in response to the U.S. government's ban on shipments of more powerful H100 and H800 GPUs to Chinese entities in late 2023. Although the H100 can be 3.3 – 6.69 times faster than cut-down H20 HGX with AI data formats, H20 HGX is a massively successful product in China not only because it outperforms the vast majority of AI processors that can be produced in volume domestically, but also because Chinese companies prefer to use Nvidia's CUDA platform as it by far outclasses everything that Chinese companies have to offer.
The congressman cited growing signs that Chinese companies are using Nvidia's H20 processors in ways that may breach current export rules. Among other things, he mentioned Tencent's reported use of Nvidia's H20s to train its Hunyuan-Large model, which likely required over 200 PFLOPS of computing power. A cluster of this capability would meet the U.S. government's definition of a supercomputer, placing it under restricted usage guidelines.
Moolenaar also outlined the disparity between U.S. and Chinese production capabilities. While Huawei's partner SMIC is expected to produce only 200,000 of the company's Ascend 910B processors in 2025 (which is based on an assumption that SMIC produced 200,000 compute chiplets for the Ascend 910B last year and its production capabilities remained the same in 2025), American companies plan to deploy over 14 million AI processors (made by their manufacturing partners) this year. Given the disparity in production capabilities, approval of mass Nvidia H20 sales to Chinese entities could enable China to create powerful AI models and distribute them globally, possibly at no cost, as part of a strategy to capture a significant share of the AI market.
Ban H20 again? Not only
The congressman criticizes the idea of comparing H20s to top-tier U.S. chips, arguing this misses the point. In his view, the correct reference point should be the most powerful chip China can produce at scale. Because the H20 substantially exceeds that level, Moolenaar warns that it would significantly advance China’s AI sector if allowed to enter the market there.
Moolenaar recommends that export rules adopt a dynamic standard that stays just ahead of China's known domestic capabilities, rather than relying on fixed U.S. performance metrics.
While such a move would enable AMD, Nvidia, and other American companies to compete against China's AI processors that local companies can produce domestically, their shipments will not allow Chinese entities to develop something they could not using only domestic solutions. In addition, this could put a roadblock on the expansion of China's AI hardware outside of China, addressing concerns that companies like Huawei could set their own AI standards globally.
There is a catch, though. Export controls are intended to prevent China's People's Liberation Army from acquiring advanced AI technologies prematurely. However, PLA does not necessarily need to use mass-produced processors, but can rely on hardware that is not built in high volumes, yet delivers performance that is actually higher than that offered by American companies.
Call to a briefing
In his letter, Moolenaar requested a detailed briefing from the Commerce Department by August 8, 2025. He asks for clarity on licensing decisions, projected shipment volumes, risk assessments, enforcement mechanisms, and any potential policy updates related to AI technologies and export controls.
Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
