Google's first TPUv7 FP8 vLLM results reportedly top GB300 NVL72 FP4 on agentic AI performance per dollar

SemiAnalysis says Google's first TPUv7 FP8 vLLM results showed 1.67x better performance per dollar than GB300 NVL72 FP4 in its Agentic Inference Benchmark. It says a four-chip TPUv7 setup beat GB300 NVL72 FP4 on performance per total cost of ownership across most of the Pareto frontier, though GB300 FP4 still won at low interactivity. The benchmark covers long conversations, accumulated context, prefix reuse and bursts of sub-agents.