

The generation over generation computing power of large AI systems is pretty much doubling each generation as we go from Ada to Blackwell to Vera Rubin. Also software improvements have made old GPUs way more efficient. NVIDIA says the H100 us now 40% faster than it was new because of the software enhancements made. The losses they’re taking get smaller every day as the tech progresses.
Old GPU supplies are just about as constrained as the new ones too. You can get like 50 cents on the dollar for a 3-year old H200, and it’s not the hyperscalers buying them. Even DGX1 systems are hard to come by. Tons of companies are buying up the capacity that’s essentially worthless to the hyperscalers.
The hyperscalers aren’t getting rid of old capacity on the same cycle as they used to anymore either. 6 year old A100s are still all over Azure/AWS since the demand is still so high and stocks are so constrained. That’s a huge TCO win for the investment.

Best I can offer is a Canadian diesel electric hybrid.