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

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.