More efficient AI is possible in principle but when the next breakthrough arrives there's no guarantee that it can be implemented on current hardware architectures. Something fundamentally different may be needed, as different from current GPU/TPU chips as they are from regular CPU/FPU chips.