Nvidia Says Watts, Not FLOPs, Decide the AI Race
Nvidia's blog just named performance per watt the ultimate AI infrastructure metric. Engineers now optimize three levers: performance per watt, tokens per watt, and cost per token.
Transcript
Nvidia just picked the one number that will decide who wins the AI race.
Nvidia says the ultimate AI metric is not raw power, it is performance per watt.
Nvidia's engineers now optimize three levers: performance per watt, tokens per watt, and cost per token.
Hopper landed in 2022. Blackwell followed in 2025, putting seventy two GPUs into one rack. Vera Rubin ships in late 2026.
Vera Rubin cuts the cost per token by ten times versus Blackwell.
The winner of the AI race will be whoever squeezes the most useful tokens from every single watt of power.
Inboxsmith is the AI receptionist that never misses a business call. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Vera Rubin is about ten times more performant in performance per watt compared to Blackwell, and cost per token is about ten times lower.(Deconstructing Nvidia's Vera Rubin, The Successor To Blackwell That's 10x More Efficient (CNBC))
- 72 Grace Blackwell chips connected over NVLink deliver 25 times more energy efficiency than the previous Hopper generation.(Nvidia exec talks AI, Blackwell chips, and energy efficiency (Yahoo Finance))
- A single NVL72 Blackwell rack pulls up to 130 kW under maximum load and delivers 1.4 exaflops per rack.(NVIDIA Blackwell: The Architecture of AI Superchips (Micro Learning))
- Framing and trigger: Nvidia's blog post naming performance per watt as the ultimate AI infrastructure metric.(Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency (Nvidia Blog))