No, while what you are saying sounds intuitive on one hand, OpenAI is building $100b supercomputer for training new models. As long as people want to train bett

No, while what you are saying sounds intuitive on one hand, OpenAI is building $100b supercomputer for training new models. As long as people want to train better models than a year before(which seems to be a good assumption for new few years), training will keep on giving good revenue to Nvidia. Inference requires 1/3rd the FLOPs than training per token and current models are already trained on entirety of internet. You can generate 3 times the internet for 1 training run.