Senators lobby for SAFE Chips Act, which would curb leading-edge AI chip exports to China — proposed bill would restrict AMD and Nvidia to H20/MI308-class accelerator sales until 2028

Nvidia Hopper H100 GPU and DGX systems
(Image credit: Nvidia)

A new SAFE Chips Act proposed by a bipartisan group aims to lock current export rules into law, preventing AMD and Nvidia from selling any AI accelerators based on the latest architectures to China for the next 30 months and forcing companies to compete with H20 and MI308 against China's latest AI processors.

A bipartisan group of U.S. senators has introduced new legislation designed to lock in the existing export control rules on AI and HPC processors from companies like AMD and Nvidia to China and other "adversary" nations for 30 months. If the Secure and Feasible Exports of Chips Act of 2025 (SAFE Chips Act) passes, then the current GPUs tailored for China — AMD's MI308 and Nvidia's H20 — will be the best processors these companies can supply to China and other foes for the next 30 months.

  • Total processing performance ≥ 4,800
  • TPP ≥ 2,400 with performance density ≥ 1.6
  • TPP ≥ 1,600 with performance density ≥ 3.2
  • Total DRAM bandwidth ≥ 4,100 GB/s
  • Interconnect bandwidth ≥ 1,100 GB/s
  • Combined DRAM + interconnect bandwidth ≥ 5,000 GB/s

Both AMD and Nvidia already have processors — AMD's Instinct MI308 and Nvidia's H20 — that comply with the aforementioned thresholds, so the companies can continue to ship them to China (including Hong Kong and Macau), Iran, North Korea, or Russia even without a U.S. export license. Assuming, of course, China lifts its self-inflicted ban on Nvidia hardware.

However, both H20 and MI308 were originally introduced in 2023, and by now, Chinese companies have already introduced AI accelerators that offer comparable or higher performance. For example, even Huawei's existing Ascend 910C beats H20 in FP16/BF16 (780 TFLOPS vs. 148 TFLOPS), with the next generation Ascend 950PR/950DT NPUs, the company will offer 1 FP8 PFLOPS (vs. H20's 296 TFLOPS) and 2 FP4 PFLOPS of performance for training and inference, offering new levels of performance, efficiency, and scalability that will open doors to ZetaFLOPS-scale AI systems in the coming years. Therefore, to be competitive in China, Nvidia and its peers from America must offer something better to stay relevant and continue pushing their standards for compute and AI around the world.

So, for months, Nvidia has been courting the U.S. government and legislators to let it sell cut-down versions of its Blackwell processors or even full-fat Hopper H200 processors (which beat both Ascend 910C and upcoming Ascend 950PR/950DT in FP16/BF16 and FP8) to China, according to media reports. However, if the bill passes as law, then the best things that Nvidia will be able to sell to its Chinese customers in the next 30 months (i.e., till mid-2028) will be its HGX H20, L20 PCIe, and L2 PCIe GPUs that are outdated today and will look antique several years down the road.

Nvidia argues that if it cannot sell reasonably competitive, but performance-capped GPUs to Chinese entities, local hardware developers like Huawei, Biren Technologies, or Moore Threads will quickly dominate the market, which will permanently displace U.S. technology in the People's Republic and will challenge it elsewhere as Chinese companies tend to expand. Nvidia also claims that allowing exports of controlled, downgraded accelerators slows China more effectively than a total ban, because it maintains dependence on U.S. hardware and standards rather than encouraging China to accelerate its own ecosystem. Finally, Nvidia warns that losing China revenue weakens America's overall leadership in high-performance AI computing, as it affects revenue and R&D.

30 months after enactment, the U.S. Department of Commerce will be able to adjust these technical thresholds originally set in 2023, but only with the majority approval of the End-User Review Committee. Furthermore, any planned modifications, along with a detailed assessment of how such changes could alter the capabilities of Chinese AI developers or influence military and cyber applications, must be disclosed to Congress at least 30 days in advance.

The new bill was proposed by Pete Ricketts (R), Chris Coons (D), Tom Cotton (R), Jeanne Shaheen (D), Dave McCormick (R), and Andy Kim (D).

"The rest of the 21st Century will be determined by who wins the AI race, and whether this technology is built on American values of free thought and free markets or the values of the Chinese Communist Party," said Senator Coons. "As China races to close our lead in AI, we cannot give them the technological keys to our future through advanced semiconductor chips. This bipartisan bill will protect America’s advantage in computing power so that the world’s most next-generation AI models are built at home by American companies, and the world’s infrastructure is built on the American tech stack."

Google Preferred Source

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.