They are, but training models is hard and inference (actually using them) is (relatively) cheap. If you make a a GPT-3 size model you don’t always need the full H100 with 80+ gb to run it when things like quantization show that you can get 99% of its performance at >1/4 the size.

Thus NVIDIA selling this at 3k as an ‘AI’ card, even though it wont be as fast. If they need top speed for inference though, yea, H100 is still the way they would go.

If it crashes hard I look forward to all the cheap server hardware that will be in the secondhand market in a few years. One I’m particularly excited about is the 4000 sff, single slot, 75w, 20GB, and ~3070 performance.

I have them off, but if something comes in that I should respond to I do send some sort of response back (errands, plans, etc.) I just don’t like feeling I get that I should respond now to something just because I saw it and now the person on the other end knows that I saw it, and are now wondering why I haven’t sent something back yet.