First third-party TPU inference benchmark: Google Ironwood up to 50% better perf/$ than B200
dylan522p · x · 2026-09-08
SemiAnalysis and the InferenceX team published the first open third-party inference benchmarking of Google TPUs, running daily across many models and scenarios. Key takeaways:
- Headline result: In apples-to-apples tests, TPUv7 Ironwood delivers up to 50% better performance per dollar than B200/B300, holding an advantage across much of the Pareto curve.
- Economics: The report analyzes both Google's internal TCO and the external TCO a real customer pays.
- Context: Anthropic is the biggest TPU user, committed to over 1 million chips (400k+ direct purchases, 600k+ rented via GCP), and will surpass DeepMind's usage by 2029.
- Trend: Rapid externalization of the TPU stack (Ironwood, TPUv8i) is eroding NVIDIA's CUDA moat.
More from Venture
- Polymarket gives Oura's IPO 41% odds of closing above $20 billion — Polymarket · 2026-09-08
- Bittensor Explained for Normies: 21M TAO Cap, 128 Subnets, $251K Paid Daily — markjeffrey · 2026-09-08
- AI companion Tolan tops 100K paying subscribers, mostly 30-40s non-tech users — ajaymehta · 2026-09-08
- Josh Elman: AI consumer apps' moats are still network effects, not models — illscience · 2026-09-08
- Gary Marcus: Model Commodification Is the Worst Case for Trillions in Data Center Capital — GaryMarcus · 2026-09-08
- Building an app has never been easier — selling one has never been harder — nikvassev · 2026-09-08