Google's TPUv7 Posts 1.67x Better Perf per Dollar Than GB300 NVL72 in SemiAnalysis InferenceX
scaling01 · x · 2026-10-15
Google has submitted its first TPUv7 FP8 results (running vLLM) to SemiAnalysis' Agentic Inference Benchmark (InferenceX), demonstrating 1.67x better performance per dollar than GB300 NVL72 FP4, 1.46x vs B300 FP4, and 1.92x vs GB200 NVL72 FP4 on the Qwen3.5 397B agentic workload. The chart shows TPUv7 chip TCO at $1.03/hr versus GB300's $2.31, with the biggest cost edge at high interactivity (P90 400-500 tok/s/user). SemiAnalysis notes Google is pushing to externalize TPU power beyond its own datacenters — a direct challenge to Nvidia's inference dominance.
More from Infra
- KAIST's KV4-RL: INT4 KV cache quantization speeds RLVR rollouts 1.8x without accuracy loss — kaist-ai · 2026-10-15
- ASML raises 2026 forecast to €45B on 45% EUV demand surge but warns of tool shortages through 2028 — CurieuxExplorer · 2026-10-15
- NVIDIA to showcase neural rendering, physical AI and robotics at SIGGRAPH Asia 2026 — tomaszbednarz · 2026-10-15
- Reddit user builds 256GB VRAM local AI rig with 4x CMP 170HX cards in a chicken-coop shed — sevendaysworth · 2026-10-15
- Microsoft's MAI-Image-2.6 and Flash land on Vercel AI Gateway with one API key — lmoroney · 2026-10-15
- Inworld opens up its consumer-AI inference API, citing tens of trillions of tokens a month — thetripathi58 · 2026-10-15