SemiAnalysis publishes TPUv7 results for multi-turn AI agent inference
SemiAnalysis released AgentX results for TPUv7 across four models, measuring the cost and speed of long-context, multi-turn AI agent inference. Its report finds more tokens per dollar in some configurations, but results depend on the model, numerical precision, software and latency metric—not a universal lead over GPUs.
SemiAnalysis published InferenceX/AgentX results for TPUv7 on November 6, 2026 (Singapore time), covering Muse Spark Open, GLM-5.2, Qwen3.5 and Kimi K3. The tests focus on AI agent workloads such as long-context, multi-turn interactions rather than only single-turn question answering.
The report says that for GLM-5.2 at FP8 numerical precision and 60–300 tokens per second per user, TPUv7 delivers 1.2–2.7 times the tokens per dollar of B200, B300 and MI355X. Tokens are the basic units models process and generate when working with text.
The results do not mean TPUv7 is cheaper or faster in every scenario. The report also notes that GPUs using lower-precision FP4 configurations still offer better cost performance across many speed ranges, potentially at some loss of accuracy or quality. The TPUv7 configuration for Muse Spark Open also still needs improvements in total end-to-end request latency. Conclusions vary with the model, software configuration and latency metric.
These are SemiAnalysis’s reported benchmark findings, not independently reproduced tests by this site.