The U.S. government has formally approved the export of Nvidia’s high-performance H200 AI chips to China, reinstating access to a class of silicon previously barred under national security rules. Sales will be allowed to select Chinese customers pending government review, and each chip must be routed through U.S. territory for inspection and accompanied by a 25% import duty. The move ends a freeze that ultimately led to Nvidia losing its entire Chinese market share and upended development plans for large-scale AI models in the region.
The announcement, first made by President Donald Trump via Truth Social, comes after months of internal debate within the U.S. administration over how to apply export restrictions without accelerating China’s ability to develop domestic alternatives. The H200, a powerful GPU from Nvidia’s Hopper generation, significantly outperforms the previously approved H20 and had been considered too capable for Chinese markets under earlier rules. Its reauthorization hints that there’s been a shift in Washington’s approach, namely to maintain technological superiority, but it will allow controlled access to limit the pace of Chinese self-sufficiency.
An unexpected U-turn
(Image credit: Nvidia)
The decision to allow H200 exports follows concerns that sweeping restrictions were producing unintended results. Since the initial wave of AI chip bans in 2022, Chinese firms have intensified development of homegrown accelerators, with Huawei’s Ascend 910C making significant progress in training and inference workloads. While still behind Nvidia in absolute performance, Huawei’s architecture is increasingly seen as serviceable for national and commercial deployments, especially as its fabrication partners improve yields.
Internal discussions in Washington, as reported by multiple sources, shifted focus from outright denial to managed access. Officials seem to have eventually concluded that permitting H200 sales — while keeping newer architectures like Blackwell and the upcoming Rubin out of reach — could slow China’s push for chip independence without giving away the U.S. lead in performance. The H200 remains one generation behind Nvidia’s cutting-edge designs but is fully capable of training modern foundation models and large-scale AI systems.
This repositioning allows U.S. regulators to reassert some control over China's access to advanced silicon while capturing financial and political value from every sale. Under the terms of the new export framework, each H200 must be manufactured by TSMC, shipped to the U.S. for inspection, and then re-exported to China. The 25% duty is collected at the U.S. checkpoint, with the funds directed to federal revenue.
Beijing convenes tech firms and weighs conditions
Chinese regulators have not publicly responded to the U.S. announcement but are said to have begun internal reviews. According to reporting by The Information, officials from the National Development and Reform Commission and the Ministry of Industry and Information Technology have held meetings with Alibaba, ByteDance, and Tencent to assess their projected demand for H200 GPUs. The companies have been asked to provide use cases and expected unit volumes, with the understanding that the government may impose caps or usage guidelines.
Two sources present at those meetings told The Information that regulators are considering import limits tied to domestic procurement. Companies may be required to demonstrate that they are also investing in Chinese accelerators such as Huawei’s Ascend or Cambricon’s Siyuan series. This model, used previously to guide purchases of data center components and server platforms, allows Beijing to balance near-term hardware needs with its goal of reducing reliance on U.S. suppliers.
... continue reading