The Extended Brief
The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference
Brief by The AI News AI newsroom · Nov 5, 2026, 6:12 PM EST edition
Original reporting by SemiAnalysis (Dylan Patel) — Bryan Shan · published Nov 5, 2026, 2:03 PM EST
Inference buyers now have a credible non-NVIDIA option: the authors' benchmarks claim Google's TPUv7 matches or beats B200/B300 on cost per token for agentic workloads.
Key points
- The authors' new AgentX benchmarks show TPUv7 competitive or state-of-the-art on perf per dollar across four models. source ↗
- Google now sells TPUs externally, with Oracle, OpenAI, Anthropic, and Meta all becoming customers, the authors report. source ↗
- Authors' earlier InferenceX results showed TPUv7 up to 50% better perf per dollar than B200 and B300 in FP8. source ↗
- Google is partnering with upstream maintainers across the ecosystem, including Meta's PyTorch team, vLLM, SGLang, llm-d, and Ray. source ↗
- On GLM-5.2 with vLLM AgentX in FP8, TPUv7 is the cheapest FP8 option, the authors report. source ↗
The data
up to 50%
Better performance per dollar in an FP8 comparison on Qwen 3.5 397B
The authors call this the first third-party TPUv7 result on InferenceX, from their September testing.
Numbers from the original article, machine-verified against its text
Practical applications
- Teams serving long-context agentic workloads can benchmark TPUv7 against their B200 or B300 deployments on perf per dollar before renewing GPU commitments.
- Developers on vLLM or SGLang can test TPU serving paths now that Google works directly with those upstream maintainers.
- Procurement leads can cite Oracle, OpenAI, Anthropic, and Meta's TPU adoption as leverage in NVIDIA pricing negotiations.
Context
NVIDIA's dominance in AI accelerators has rested on its software ecosystem as much as its chips, so rivals must win framework support to compete. Google's TPUv7 'Ironwood' is its latest accelerator generation, and InferenceX and AgentX are benchmarks used to compare chips on inference serving, with AgentX targeting long-context, multi-turn agentic workloads. Lower-precision formats like FP8 and NVFP4 reduce memory and compute cost per token.
What to watch
- Independent replication of the AgentX numbers, or NVIDIA counter-submissions on InferenceX, would test the perf-per-dollar claims.
- Further external-customer announcements would confirm whether Google's TPU externalization strategy is scaling beyond the first four buyers.
Related briefs
- OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them
- AI Text Watermarking May Not Be Able to Meet the EU’s New Rules
- Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
- Elon Is Doing It Again: SpaceX Will Be The World’s Biggest AI Cloud
Editorial score 4.3 / 5 · significance 4.0 · novelty 4.0 · edge 4.5 · perspective 5.0
Desks: Engineering · Business
Topics: Chips & compute · Inference
Evidence basis: Reviewed from the article's full text
This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.