SemiAnalysis: TPUv7 Delivers 1.2 to 2.7 Times More Tokens Per Dollar Than NVIDIA/AMD on GLM-5.2 AgentX
SemiAnalysis released the first cross-model results for TPUv7 on the AgentX agent inference benchmark on November 5, 2026. At FP8 precision, TPUv7 on GLM-5.2 delivers 1.2 to 2.7 times more tokens per dollar than the B200, B300, and MI355X. On Qwen3.5, TPUv7 FP8 performance per dollar is 1.67 times that of the GB300 FP4. However, NVIDIA FP4 still leads on GLM-5.2 by a factor of 3.5 to 3.6, and TPUv7 currently lacks native FP4 support. Results were submitted separately by Google (Qwen3.5), RadixArk (Muse Spark Open), and Inferact (Kimi K3) and have not been independently reproduced.
SemiAnalysis released the first cross-model results for TPUv7 (Ironwood) on the InferenceX/AgentX agent inference benchmark on November 5, 2026. AgentX simulates real-world agent workloads characterized by multi-turn interactions, long contexts, and high prefix reuse, which differ significantly from single-turn chat scenarios. This release covers four models: Meta's Muse Spark Open (816B parameters), Zhipu AI's GLM-5.2/5.3 (744B parameter MoE), Alibaba's Qwen3.5 397B, and Moonshot AI's Kimi K3 (2.8T parameters).
At FP8 precision, the results for each model must be viewed separately. On GLM-5.2, within an interaction range of 60 to 300 tokens per second per user, TPUv7 delivers 1.2 to 2.7 times more tokens per dollar than the B200, B300, and AMD MI355X. On Qwen3.5 397B, SemiAnalysis reports that TPUv7 FP8 performance per dollar is 1.67 times that of the GB300 NVL72 FP4—this marks the first AgentX model result submitted by Google itself. On Kimi K3, results submitted by Inferact, founded by the vLLM founding team, show that TPUv7 delivers 1.15 to 2.5 times more tokens per dollar than the GB200, B200, B300, and MI355X. Muse Spark Open was submitted by RadixArk, maintained by SGLang, achieving competitive decode cost-effectiveness even on a GPU-optimized architecture theoretically unfriendly to TPUs.