For large machines like tanks, yes, the power the AI would consume would be an insignificant fraction of the total power. However, I wasn't thinking large. Far cheaper than tanks would be swarms of tiny autonomous drones, each carrying flechettes loaded with neurotoxin. If you released a hundred thousand of them, they would totally destroy any human infantry. Now imagine releasing them inside a city: urban warfare would be almost easy.
Of course, such drones would need AIs more advanced than anything any country currently has, to operate in complex, real-world environments -- and to distinguish friend from foe. The AI would consume a significant fraction of such a drone's total power.
For example, consider a drone the size of a
, which for many people is the state of the art today for small drones, or close to it. It's nowhere near autonomous, but let's ignore this. The drone's battery capacity is 3.8 Ah, it runs at 11.4 volts, and it flies for 27 minutes max. Thus it consumes 96 watts while flying. Now even a monster CPU like an Intel Core i9 10900K would be barely enough to run the simplest, most primitive AI -- and yet this monster will consume 250 watts TDP, several times the Mavic's total power. This is not "insignificant".
Thus even EUV lithography may not be enough, though breakthroughs in AI could happen to reduce the necessary computation load. If these breakthroughs don't happen, my hopes are on the neuromorphic chips, which are at least partly analog and therefore consume less power than fully digital devices. Analog circuits are less precise, but AI's neural nets are rather fuzzy anyway.