On Monday, December 8, local time, U.S. President Donald Trump announced that artificial intelligence (AI) chip giant Nvidia will be allowed to ship its H200 AI chips to "authorized customers" in China and other regions, on the condition that the U.S. government receives a 25% cut of Nvidia’s H200 sales revenue in these regions.
In a post on Truth Social, Trump stated that China had given a "positive response" to the proposal. He wrote that the policy "will support U.S. jobs, strengthen U.S. manufacturing, and benefit American taxpayers." "The Department of Commerce is finalizing the details, and the same approach will apply to AMD and Intel," he added in the post, "as well as other great American companies."
Back in August this year, Nvidia agreed to hand over 15% of its sales revenue from H20 chips in China to the U.S. government in exchange for the approval to resume exports of its China-specific H20 chips.
However, almost simultaneously, Chinese state media warned of potential security risks associated with the H20 chip and noted that its performance was already outdated.
Subsequently, Nvidia encountered sales difficulties with the H20 in China. Nvidia CEO Jensen Huang also acknowledged that the company’s market share in China’s AI chip market had dropped to zero, and he predicted that it would remain at zero for the next quarter and the one after that.
In response, Huang has been pushing for the export of Nvidia’s H200 chips and even the more advanced Blackwell GPUs to China. He stated that he hopes to achieve a breakthrough between the U.S. and Chinese governments and reach cooperation with both to re-enter the huge Chinese market.
According to Huang’s forecast, "China’s AI chip market is approximately $50 billion in size and is likely to grow to $200 billion by the end of the decade."
"It’s really regrettable that U.S. companies cannot participate in this market. This is a very important source of revenue," Huang said. "With this revenue, we can increase investment and accelerate the pace of investment, so I hope we have the opportunity to return to the Chinese market. But for now, we can only assume zero revenue."
Nevertheless, while Trump opposes the export of Nvidia’s advanced Blackwell GPUs to China, he is considering approving the export of the H200, which features the same architecture as the H20 but offers stronger performance, to China.
Regarding Trump’s approval of H200 exports to China, a Nvidia spokesperson said in a statement: "We applaud President Trump’s decision to allow the U.S. chip industry to compete, supporting high-paying U.S. jobs and manufacturing."
The spokesperson further noted: "Providing H200 to commercial customers approved by the Department of Commerce is a thoughtful balance and a positive development for the United States."
According to available information, both the previously approved H20 chips for export to China and the H200 are based on the previous-generation Hopper architecture, but the H20 is a cut-down version of the H100. The H200 is equipped with 141GB of HBM3e memory and boasts a memory bandwidth of up to 48TB/s, representing a significant performance improvement over its predecessor, the H100. Estimates suggest that the H200’s performance will be twice that of the H20 chip.
However, due to existing export control restrictions on performance and bandwidth parameters, even if the U.S. government plans to approve H200 exports to China, Nvidia will need to make certain "cutbacks" to the H200, with further reductions in core count and bandwidth expected. But at the very least, its overall performance is projected to deliver a substantial upgrade over the original H20 (which only has 96GB of HBM3 memory).
For domestic AI technology firms in urgent need of high-performance AI chips, this may be welcome news, as better chips can help them iterate their AI large model capabilities more quickly. However, for domestic AI chip manufacturers, this will undoubtedly expose them to more intense competition.
(Reprinted from https://news.eccn.com/）