NVIDIA H800

The NVIDIA H800 is a datacenter GPU whose key distinction from the flagship H100 is its halved NVLink interconnect bandwidth — 400G bidirectional versus the H100's 900G, or about 160G measured unidirectionally per 真格基金's analysis . This bandwidth constraint became a defining feature of its use case: DeepSeek built its training infrastructure around H800s, and the engineering team's extensive "cuda optimization and orchestration between computation and communication" was described by 十字路口Crossing as a direct response to this limitation . The same 真格基金 piece notes that if DeepSeek had access to ample H100 supply, "there would be no need for these tricks" . The H800 was available for Chinese purchase after the October 2022 U.S. export controls but before the January 2025 second round of restrictions added it to the controlled list, after which H20 became the compliant alternative .