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.