Post

here is an insane thing I have learned: it appears that nobody has any real idea of how much actual, functional GPU capacity exists. Not "in progress," not "set to come online," I mean actual functional in-use capacity.

Considering the US stock market is based on selling millions of these things, you'd think it'd be much easier to say A) where they went and B) what they were doing. 6 million blackwell GPUs shipped - I can find evidence of maybe...50,000 installed? 75k max? bsky.app/profile/bria...

it's more that, despite "gigawatts of capacity" being discussed regularly, I am finding it difficult to find more than two examples of AI data centers over 100MW of capacity. Even Colossus 1 doesn't seem to be fully energized? bsky.app/profile/ianv...

Even the 125MW of "capacity" that Core Scientific has in Calvert KY is for bitcoin mining, and those rigs aren't anywhere near powerful enough for AI. There are SIX MILLION Blackwell GPUs sold, apparently - where did they go, and are they ...on? www.gem.wiki/Calvert_Data...

What I'm saying is that I believe the majority of generative AI is run on A100 and H100 GPUs, and I see no compelling evidence of large-scale or even minor-scale deployments of Blackwell GPUs bsky.app/profile/ombu...

are there enough PC enthusiasts and bitcoin nutters to explain the discrepancy?