TPUs and ASICs run in data centers too. Your argument only holds true if there's some satisfied limit to demand for inference. If not, data centers will continue to spring up to host more and more agents. Even if agents were running on hardware and software as efficient as the human brain, its conceivable we want trillions of them running at any given time which would require data center scale.
Everything has some satisfied limit to demand, often depending on the price. If you assume there will never be any satisfied limit to demand for inference at any price you can justify any investment.