Nvidia is evaluating whether to increase production of its H200 data center GPUs after demand from Chinese customers is expected to exceed current supply, according to people familiar with the matter, as reported by Reuters. The discussions follow a policy shift this week in which the U.S. government said Nvidia would be allowed to export the H200 to China under a framework that includes a 25% fee on sales.
Reuters reports that Nvidia has already briefed Chinese clients that it is assessing options to add capacity for the H200, its most powerful accelerator from the Hopper generation, which is currently limited. This comes despite the company’s broader focus on ramping production of its newer Blackwell GPUs and preparing for the next-generation Rubin platform, both of which are competing for the same advanced manufacturing resources at TSMC.
However, it’ll be difficult for Chinese companies to ignore the H200’s appeal. It is currently the most capable AI accelerator that they can legally obtain from Nvidia, delivering substantially higher performance than the China-specific H20. According to Nori Chiou, investment director at White Oak Capital Partners, speaking to Reuters, the H200’s “compute performance is approximately 2-3 times that of the most advanced domestically produced accelerators,” a gap that will naturally continue to drive interest despite political and regulatory risks.
Nvidia will no doubt have to think carefully about adding H200 capacity. Advanced wafer capacity at TSMC is already heavily contested by customers ranging from Nvidia itself to large cloud providers developing in-house silicon. Despite Washington’s apparent willingness to ease export restrictions, another U-turn is not beyond the realms of possibility, and China could quite easily slam the doors shut in spite of what the U.S. wants.
Follow Tom’s Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
