It's not that hard to see the old GPUs being used e.g. for inference on cheaper models, or sub-agents, or mid-scale research runs. I bet Karpathy's $100 / $1000 nanochat models will be <$10 / <$100 to train by 2031