Comment by zozbot234

If your point is purely about supply and demand for datacenter HBM and LPDDR, you're probably right. Local model inference (using the existing memory stock) can make a dent in current use, but not in projected future uses that will plausibly involve much larger models.