SemiAnalysis benchmarks TPUv7 on agentic inference, claims tokens-per-dollar edge

SemiAnalysis published what it describes as the first long-context, multi-turn agentic inference results for Google's TPUv7 (Ironwood) on the InferenceX/AgentX benchmark, covering four models served through vLLM and SGLang by ecosystem contributors including Inferact and RadixArk. On GLM-5.2 (FP8), the authors report TPUv7 as the cheapest FP8 option at 1.2x to 2.7x the tokens per dollar of B200, B300 and MI355X between 60 and 300 tok/s/user, while noting NVIDIA NVFP4 still wins with an acknowledged accuracy/quality loss; Kimi K3 (MXFP4) is reported at 1.15x to 2.5x the tokens per dollar of GB200, B200, B300 and MI355X. Qwen3.5 (FP8) with llm-d disaggregated serving is said to beat GB300 FP8 at every interactivity level, whereas Muse Spark Open (NVFP4 weights) is only described as competitive. These are SemiAnalysis-authored, partly community-submitted measurements rather than independent audits, and the piece cites Google's partnerships with Meta's PyTorch team, vLLM, SGLang, llm-d and Ray as evidence it is contesting NVIDIA's software moat.