Nvidia H200 access limited

U.S. officials said on April 22 that Nvidia had sold no H200 artificial-intelligence chips to Chinese companies. (reuters.com) Commerce Secretary Howard Lutnick said “zero H200 chips” had gone to China as of April 22, according to Reuters. Nvidia’s H200 is one of the company’s higher-end data-center chips for training and running large artificial-intelligence models. (reuters.com) (nvidia.com) Nvidia says the H200 uses 141 gigabytes of HBM3e memory and 4.8 terabytes per second of memory bandwidth, which lets servers hold larger models and move data faster than older chips. Nvidia markets it for generative artificial intelligence and high-performance computing workloads. (nvidia.com 1) (nvidia.com 2) Washington has spent the past several years tightening export controls on advanced computing chips to China. The Commerce Department said in October 2022 that it was imposing new restrictions on advanced computing and semiconductor manufacturing items to the People’s Republic of China. (bis.doc.gov) Those rules reshaped Nvidia’s China business by separating which chips could ship there and which could not. The H200 now sits on the wrong side of that line, leaving Chinese cloud groups and model developers to rely on older Nvidia products, domestic alternatives, or both. (reuters.com) (bis.doc.gov) Nvidia has previously built China-specific products when U.S. rules cut off faster chips. Reuters reported in April 2025 that the company took a $5.5 billion charge tied to new U.S. export curbs on its H20 chip, another product designed for the Chinese market. (reuters.com) The split is not only about one model number. It means the same Nvidia software stack can sit on very different hardware pools in the United States, the Middle East, Europe, and China, with capacity and performance now shaped as much by export policy as by demand. (reuters.com) (nvidia.com) For now, the clearest fact is the one Lutnick gave Reuters on April 22: China has not received Nvidia’s H200. That leaves Nvidia, its customers, and regulators still negotiating where the next boundary in artificial-intelligence hardware will be drawn. (reuters.com)