AI chip sales to China: Six senators call for a ban

Six Democratic senators have asked the Trump administration to block the export of NVIDIA H20 and AMD MI308 GPUs to China in exchange for a 15% cut of the revenue from their sale. Meanwhile, DeepSeek has reportedly decided to use NVIDIA's AI chips to train its R2 model.

National security and violation of the law

The Trump administration has reauthorized the sale of NVIDIA's H20 and AMD's MI308 GPUs to China. The US president later confirmed that the government will receive 15% of the revenue (and may do the same for exports of Blackwell GPUs). Some national security experts have highlighted the strategic error, calling for the sale to be halted.

After the one at the end of July , sent to Howard Lutnick (Secretary of Commerce), six Democratic senators sent a second letter to Donald Trump , asking to once again ban the export of NVIDIA and AMD chips to China.

The senators highlight national security risks because the Chinese Communist Party could use GPUs to improve communications, surveillance, and military systems. The letter also emphasizes that the 15% tax is not permitted under export law and could be unconstitutional.

Senators want to receive information from the Trump administration by August 22. An NVIDIA spokesperson said:

The H20 GPU wouldn't have improved anyone's military capabilities, but it would have helped America attract the support of developers around the world and win the AI race. Banning H20 cost American taxpayers billions of dollars, for no benefit.

AMD hasn't released a statement yet. Instead, this is the White House spokesperson's comment:

Democrats and irrelevant “experts,” who were completely lost when Joe Biden’s administration let H20 chips and other advanced technologies flow freely to China, are now pretending to care about our national and economic security.

The aforementioned experts predict that Chinese companies will leverage US AI chips for their operations. According to the Financial Times , DeepSeek postponed the launch of its new R2 model due to problems during the training phase with Huawei's Ascend chips. The latter could only be used for inference, while training would be performed with NVIDIA GPUs.