Comment by Hamuko

Would it? First of all, all current AI users still need memory in their devices. You can't use ChatGPT with pen and paper. There would always be a consumer demand for memory, even without local inference. Secondly, it'd drastically decrease the demand for HBM, which has shit yields when compared to conventional memory. It's also questionable as to how much of the produced HBM is actually in use and not sitting in a warehouse.