Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency
NVIDIA Blog · Shruti Koparkar · 2026-07-14
NVIDIA argues that performance per watt is the defining metric for AI infrastructure efficiency, citing up to 25x improvement with its GB300 NVL72 Blackwell platform over the prior Hopper generation.
Appears in
Extraction
Topics: ai-infrastructureenergy-efficiencygpu-performancenvidia-blackwellinference-optimization
Claims
- Performance per watt is the ultimate AI infrastructure metric because power is the inescapable constraint limiting token throughput and profitability.
- NVIDIA GB300 NVL72 delivers up to 25x performance per watt compared to the Hopper generation on DeepSeek V4 Pro.
- Software optimizations alone improved DeepSeek V4 performance per watt by up to 5x in a single month.
- NVIDIA DSX MaxLPS enables operators to run up to 40% more GPUs within the same power budget.
- Leading AI labs including Anthropic and OpenAI use NVIDIA Blackwell NVL72 systems for production inference.
Key quotes
Power is AI infrastructure's inescapable constraint. How many tokens an AI factory can generate within a fixed power budget determines its revenue and profitability.
In AI factories, power lost to cooling and rack-level inefficiencies can mean only about 60% of the electricity pulled from the grid turns into useful AI work.
On DeepSeek V4, performance per watt improved by up to 5x in a single month.