> appearance of new techniques that make inference computationally very cheap So long as training (computation of the gradient and update) remains expensive, n

> appearance of new techniques that make inference computationally very cheap So long as training (computation of the gradient and update) remains expensive, nVidia - and others - will be fine. At the moment, between the global cloud providers and smaller players, there appears to be effectively "unlimited" demand for GPUs. This is likely to continue, as ML gets used for more and more things. Remember - the GPTs are the new hotness, and certainly hog their share of GPUs right now, but there are a lot of models just quietly working without making anywhere near the splash; all of these need GPUs too.