SemiAnalysis — Rubin NVL72 benches at 7x the performance per watt of GB300, not the 3x claimed

X post

SemiAnalysis — Rubin NVL72 benches at 7x the performance per watt of GB300, not the 3x claimed

SemiAnalysis says its own testing of Rubin NVL72 against GB300 NVL72 shows roughly 7x better performance per watt, against the 3x Nvidia claimed for the part at GTC. It is a measurement rather than a vendor slide, and if it holds the power drawn per unit of AI compute falls faster than the grid-constraint arithmetic currently assumes.

Open on X SemiAnalysis (@SemiAnalysis_)
SemiAnalysis✓ @SemiAnalysis_
Sep 16

JENSEN IS A LIAR, HE IS UNDERPROMISING & OVERDELIVERING 🚀🚀🚀 At GTC 2026, Jensen only claimed 3x better perf per watt on Rubin NVL72 versus GB300 NVL72, but our tests SHOW 7x BETTER PERF PER WATpic.x.com/lGFm7qOYUbYUb

Why it's worth your time

SemiAnalysis says its own testing of Rubin NVL72 against GB300 NVL72 shows roughly 7x better performance per watt, against the 3x Nvidia claimed for the part at GTC. It is a measurement rather than a vendor slide, and if it holds the power drawn per unit of AI compute falls faster than the grid-constraint arithmetic currently assumes.

7 events
DateTypeEvent
Aug 28 ArticleThe Layers That Feed AI Compute All Trade Below Their Averages. Compute Doesn't. Aug 11 tweetSemiAnalysis — TileRT compiles the whole model into one persistent GPU kernel, claiming 1.9x decode interactivity on the same Blackwell chips Aug 8 tweetIPO Newsroom — NVIDIA is putting up to $3B into the Blackstone-backed firm that owns the land under Stargate Jul 28 ArticleFifteen of twenty broke together Jul 26 redditu/rdh2dmd — Samsung, SK Group seal $950B in AI deals with Nvidia and Broadcom Jul 22 pickAlphabet Q2: $44.9B quarterly capex, $180-190B FY guide, 2027 'significantly increases' Jun 22 ArticleWeek in Review — The Rotation Down-Stack

Tweets, videos, filings, articles, and source pages tied to this read.

1
X@SemiAnalysis_ tweet