Sort of. At the gate counts we're talking, it doesn't take long until you can't even fit the chip on a whole wafer of an older node. IIRC, an Nvidia GA100 doesn't fit on a 300mm 65nm wafer (to say nothing of being able to clock it at a viable speed even if you were to cut it down). Going off chip destroys perf for a lot of designs too.

It's not just a price component, but a performance and viability of new designs one, at a time where the DoD internally is pushing ML hard as a major pillar of investment in weapons technology.