Nvidia just announced its DGX GH200 at Computex. It’s got 256 of its new “Grace Hopper Superchips” for an exaflop of AI performance — and contains 150 miles of optical fiber and over 2,000 fans. Google, Meta, and Microsoft will be “evaluating” it, though not necessarily purchasing kits: they tend to build their own compute clouds, even if they contain loads of Nvidia GPUs.

Nvidia says it’s 2.2x faster than a last-gen DGX H100 cluster at GPT3 training, as one example.