Nvidia's GB300 FP4 Still Trails Google's TPUv7 on Cost Efficiency
Key Summary
Nvidia's GB300 FP4 submission remains 1.64 times worse in performance per dollar than Google's TPUv7 on an agentic inference benchmark, despite improvements after feedback and engineering work.
Precedent & Market Context
Nvidia has been the dominant player in AI hardware for years, but Google's TPUv7 is now challenging its lead on cost efficiency.
Main Event
According to SemiAnalysis, Nvidia's GB300 FP4 submission remains 1.64 times worse in performance per dollar than Google's TPUv7 on an agentic inference benchmark.
Analysis & Market Impact
The gap persists despite Nvidia's efforts to improve its results after receiving feedback and putting engineering work into the submission.
The format detail used in the benchmark makes the result more striking, with Nvidia's entry using FP4 and Google's TPUv7 running on FP8.
Implications & Outlook
The disparity in cost efficiency has significant implications for the AI hardware market, with cost-conscious enterprises having a concrete reason to look beyond Nvidia for agentic inference workloads.
The next round of results will show whether this is a temporary software gap that Nvidia can tune away or a structural cost advantage for Google's architecture on agentic inference.