{
 "generated": "2026-11-06",
 "sha": "78bbd39a8828",
 "node": {
  "id": "event:sig-0103",
  "type": "Event",
  "name": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代",
  "slug": "sig-0103",
  "kind": "signal",
  "kind_zh": "行业信号",
  "date": "2026-11-06",
  "summary": "11-05 发布。此前（9 月）他们只有单轮 InferenceX 结果，这次放出第一批官方 AgentX（长上下文、多轮、真实 agentic 服务）跨四个模型的结果：GLM-5.2 vLLM FP8 下 TPUv7 是最便宜的 FP8 选项，在 60 到 300 tok/s/user 区间给出 B200/B300/MI355X 的 1.2 到 2.7 倍 tokens/美元（但 NVIDIA 的 NVFP4 仍更强，代价是精度损失）；Qwen3.5 配 llm-d 分离式服务时 TPUv7 在所有交互档位都胜过 GB300 FP8，对 GB300 FP4 也能在同等质量下打平；Kimi K3 上是 GB200/B200/B300/MI355X 的 1.15 到 2.5 倍；连为 GPU 协同设计的 Meta Muse Spark Open 在 SGLang NVFP4 下 TPU 也拿到有竞争力的性价比。更关键的是软件侧：TPU 已经和 Meta PyTorch 团队、vLLM（Inferact、RedHat）、SGLang（RadixArk）、llm-d、Ray 等上游维护者打通，提交方里出现了 OSS 生态开发者而非只有谷歌自己。结论是 NVIDIA 历史上的软件生态护城河在推理这一侧第一次出现可行替代。",
  "url": "https://newsletter.semianalysis.com/p/tpu-strikes-the-king",
  "evidence": [
   {
    "kind": "own",
    "label": "本站页面",
    "url": "https://xiaopingfeng.com/graph/signals/#sig-0103",
    "text": "AI 行业信号流（独立研究机构）：SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代"
   },
   {
    "kind": "media",
    "url": "https://newsletter.semianalysis.com/p/tpu-strikes-the-king",
    "text": "信号的原文出处（SemiAnalysis）"
   }
  ],
  "entry": "derived",
  "category": "research",
  "category_zh": "独立研究机构",
  "source_name": "SemiAnalysis",
  "source_id": "org-semianalysis",
  "links": [
   {
    "label": "SemiAnalysis",
    "url": "https://newsletter.semianalysis.com/p/tpu-strikes-the-king"
   }
  ],
  "degree": 0
 },
 "page": "https://xiaopingfeng.com/graph/e/event/sig-0103/",
 "out": [
  {
   "type": "placed_in",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代出现在 Landscape 的「独立研究机构」",
   "direction": "out",
   "id": "layer:ai-industry-signals--5",
   "name": "独立研究机构",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103",
     "text": "AI 行业信号流（独立研究机构）：SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代"
    }
   ]
  },
  {
   "type": "involves",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代涉及SemiAnalysis",
   "direction": "out",
   "id": "org:semianalysis",
   "name": "SemiAnalysis",
   "role": "issuer",
   "note": "信源：SemiAnalysis",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103",
     "text": "AI 行业信号流（独立研究机构）：SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代"
    }
   ]
  },
  {
   "type": "mentions",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代提到Google",
   "direction": "out",
   "id": "org:google",
   "name": "Google",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "名称匹配：「谷歌」出现在信号的标题里",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103"
    }
   ]
  },
  {
   "type": "mentions",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代提到Kimi",
   "direction": "out",
   "id": "app:kimi",
   "name": "Kimi",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "名称匹配：「Kimi」出现在信号的正文里",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103"
    }
   ]
  },
  {
   "type": "mentions",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代提到Meta",
   "direction": "out",
   "id": "org:meta",
   "name": "Meta",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "名称匹配：「Meta」出现在信号的正文里",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103"
    }
   ]
  },
  {
   "type": "mentions",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代提到Moonshot AI",
   "direction": "out",
   "id": "org:moonshot",
   "name": "Moonshot AI（月之暗面）",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "名称匹配：「Kimi」出现在信号的正文里",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103"
    }
   ]
  },
  {
   "type": "mentions",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代提到NVIDIA",
   "direction": "out",
   "id": "org:nvidia",
   "name": "NVIDIA",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "名称匹配：「NVIDIA」出现在信号的正文里",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103"
    }
   ]
  },
  {
   "type": "mentions",
   "sentence": "SemiAnalysis 首发 TPUv7 的长上下文多轮 agentic 实测：谷歌 TPU 在推理侧已是 CUDA 护城河的可行替代提到PyTorch",
   "direction": "out",
   "id": "org:pytorch",
   "name": "PyTorch",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "名称匹配：「PyTorch」出现在信号的正文里",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0103"
    }
   ]
  }
 ],
 "in": []
}