◆ Daocaijing · 稻草财经 AI 投研

SA:谷歌TPU:在AI推理市场争夺一席之地

稻草财经 AI 研报速读 · 来源:海外投行 · 2026-11-07
一句话看懂:SemiAnalysis认为TPUv7在多款模型的FP8 Agent推理基准上已达到每美元性能领先或可比,谷歌正借软件生态外化策略争夺推理市场。

稻草财经 视角 · 独立逻辑线拆解

1. TPU外部化战略:从内部芯片到多客户产品

事实:Since then, we have been proven right over and over again with Oracle, OpenAI, A;partnering with upstream maintainers across the entire ecosystem, including the

2. AgentX首批TPUv7官方基准结果

事实:delivering 1.2x to 2.7x the tokens per dollar of B200, B300, and MI355X from 60;it is 1.67x better perf per dollar than even FP4 GB300 NVL72;NVIDIA NVFP4 still beats TPUv7 FP8, but with accuracy / quality loss

3. NVIDIA的反制与组织层面挑战

事实:their first GB300 FP4 counter submission is still 1.64x worse per dollar than TP;have been stuck in “internal NVIDIA legal review” for three months;they overtuned decode at the expense of prefill, hurting overall end-to-end perf

4. Meta Muse Spark Open:对TPU不友好的模型仍具竞争力

事实:TPUv7 is competitive for decode performance per dollar even when using the cost;Muse Spark Open’s 64 head dim, which means that for long-context attention, 25%

5. GLM-5.2/5.3与FP4格局: cheapest FP8,但FP4 GPU仍领先

事实:GB300 and GB200 FP4 deliver 3.6x and 3.5x the tokens per dollar of TPUv7;we will need to wait until Q1 2027 for TPUv8 with FP4 support;With TPUv7 delivering 1.15x to 2.5x the tokens per dollar of GB200, B200, B300,

6. Qwen3.5 397B与llm-d分解式服务:FP8口径全面超越GB300

事实:Qwen3.5 397B became the first model Google submitted to AgentX;certain major hyperscalers/AI lab use Qwen3.5 397B and multi trillion parameter;TPUv7 with llm-d beats GB300 at every interactivity, from >2x at high throughput;routing to the exact data-parallel rank raised decode cache hit rates from 56% t

核心逻辑

证据边界在于:上述结论均基于SemiAnalysis自家的InferenceX/AgentX基准及其TPUv7测试提交,在FP8同精度对比中TPUv7多款模型每美元性能领先或可比;但NVIDIA NVFP4在多数速度点仍提供更高每美元token数(如GLM场景下GB300 FP4为TPUv7的3.6倍),只是伴随精度下降。结论成立条件是外部客户口径的成本基准与FP8对FP8的比较。

主要风险

展开完整深度解读(报告综述·更多利好与风险)

报告综述

本材料为SemiAnalysis于2026年11月5日发布的报告《Google's TPU Stakes Its Claim on Inference》,作者包括Bryan Shan、Dylan Patel等。报告指出,自2025年11月该机构首次披露谷歌对外销售TPU以来,Oracle、OpenAI、Anthropic、Meta已相继成为TPU客户。本次报告发布首批TPUv7(Ironwood)在长上下文、多轮Agent推理基准AgentX上的官方结果,覆盖四个模型:GLM-5.2 FP8下TPUv7在60至300 tok/s/user区间提供B200、B300、MI355X的1.2倍至2.7倍每美元token数;Qwen3.5 397B FP8下每美元性能达FP4 GB300 NVL72的1.67倍;Kimi K3下为GB200、B200、B300、MI355X的1.15倍至2.5倍;Meta Muse Spark Open结果可比。报告同时指出NVIDIA FP4在多数场景仍领先,但存在精度损失,且NVIDIA首个FP4反制提交在端到端上仍落后。

本文为 稻草财经 对海外投行公开研报的 AI 解读与整理 · 解读置信度 86%,不含研报原文;完整跟踪与实时行情请在 稻草财经 终端查看。

在 稻草财经 终端看实时行情与更多解读 →