Data Center Dynamics: Nvidia creates pared back H100 GPU for export to China, called H800

Secondary trade-press corroboration of the one H800 figure that is consistently reported: 'Now, Reuters reports that Nvidia has done the same with the H100, this time cutting its interconnect performance by about half - leaving it with around 300GBps.' Read that carefully: 'around 300 GB/s' is the interconnect figure the article attributes to the H100-class part, and since a Hopper SXM is normally quoted at 900 GB/s, 'about half' and 'around 300 GB/s' are two loose approximations of one qualitative claim rather than two precise values - so this source does NOT by itself pin a number. Retained as a SECOND source alongside acc-fill-reuters-h800-export-china because a single wire story is not enough to fix a figure on an undocumented part. It explicitly records that 'the company declined to disclose the exact differences between the H800 and H100', and gives the A800 sibling for contrast: 'The company began selling the A800 last November, cutting the GPU's 600GBps interconnect down to 400GBps.' Accessed 2026-10-03. KIND NOTE: the schema's kind enum has no 'news' member, so this trade-press article is filed as 'blog'. It is secondary reporting that explicitly attributes its central claim to Reuters, so a reader should treat the 300 GB/s and 'about half' claims as reported speech, not as DCD's own measurement. 2026-10-11: Re-read article. Confirms DCD reports Reuters statement that Nvidia halved interconnect on H100 down to around 300 GBps for H800. Confidence remains 0.6 as secondary reporting of reported speech.