TPUv7 FP8 efficiency 1.85x that of GB300… falls behind on FP4

SemiAnalysis’s InferenceX benchmark, published October 19, 2026, found Google’s TPUv7 and Nvidia’s GB300 split on cost efficiency for AI agent inference depending on the comparison basis, using Qwen3.5-397B at 200 tokens per second per user. At matched FP8 precision, TPUv7 delivered 1.85x GB300’s tokens per dollar on external-customer TCO and 2.17x on Google’s internal cost basis. The advantage reverses against GB300’s FP4 results, where TPUv7 FP8 reached only 0.54x the relative throughput on external-customer costs. The results matter because they show precision choice and cost assumptions, not raw hardware, can flip vendor economics for inference workloads. Watch for further FP4-versus-FP8 comparisons and customer TCO disclosures as both platforms scale.