NVIDIA published a case for treating performance per watt as a core metric for AI infrastructure efficiency. The company argues that an AI factory's revenue and profitability depend on how many useful tokens it can generate within a fixed power budget.
The argument reflects the physical limits now shaping AI deployment. As data centers face power constraints and rising demand, efficiency is becoming a competitive factor alongside raw model speed and hardware availability.