[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-anthropic-builds-in-house-chip-team-claude-zh":3,"article-related-anthropic-builds-in-house-chip-team-claude-zh":31,"series-industry-dca20a4a-51fe-4918-815d-f6fe9736f54d":79},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":24,"views":28,"created_at":29,"published_at":30,"topic_cluster_id":11},"dca20a4a-51fe-4918-815d-f6fe9736f54d","anthropic-builds-in-house-chip-team-claude-zh","Anthropic 自建晶片團隊壓低 Claude 成本","\u003Cp data-speakable=\"summary\">\u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa> 已開始招募自研晶片團隊，目標是壓低 \u003Ca href=\"\u002Ftag\u002Fclaude\">Claude\u003C\u002Fa> 的推論成本。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> 這步棋很直白。Claude 吃掉的 Token 太多，算力帳單也跟著變大。公司一邊找晶片人才，一邊還繼續跟 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Ftpu\" target=\"_blank\" rel=\"noopener\">Google TPU\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Faws.amazon.com\u002Fmachine-learning\u002Finferentia\u002F\" target=\"_blank\" rel=\"noopener\">AWS Inferentia\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fdata-center\u002F\" target=\"_blank\" rel=\"noopener\">NVIDIA\u003C\u002Fa> 合作，路線很務實。\u003C\u002Fp>\u003Cp>這代表 Anthropic 想把成本控制得更細。對 LLM 公司來說，訓練很燒錢，推論也一樣燒。當產品開始大規模服務企業客戶，省下來的每一分推論成本，都會直接反映在毛利上。\u003C\u002Fp>\u003Cp>這種做法也很符合現在的 AI 產業節奏。模型公司不想只當軟體商，因為算力價格一波動，財務表現就跟著抖。自己養一支晶片團隊，至少能把部分命運握在手上。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>內容\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>公司\u003C\u002Ftd>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa>\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>目標\u003C\u002Ftd>\u003Ctd>降低 Claude 推論成本\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>既有算力來源\u003C\u002Ftd>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Ftpu\" target=\"_blank\" rel=\"noopener\">Google TPU\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Faws.amazon.com\u002Fmachine-learning\u002Finferentia\u002F\" target=\"_blank\" rel=\"noopener\">AWS Inferentia\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fdata-center\u002F\" target=\"_blank\" rel=\"noopener\">NVIDIA GPU\u003C\u002Fa>\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>核心背景\u003C\u002Ftd>\u003Ctd>Claude 服務大量 Token 推論\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>自研晶片，先看的是帳單\u003C\u002Fh2>\u003Cp>很多人一聽到自研晶片，就會聯想到硬派技術秀。但對 Anthropic 來說，這比較像財務工程。推論成本如果能降下來，Claude 的商業模式就更好算，也更能撐住企業級定價。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786449787150-5apv.png\" alt=\"Anthropic 自建晶片團隊壓低 Claude 成本\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這件事的重點很現實。AI 服務不是只比模型分數，還要比每次回答要花多少 GPU 時間。當 Token 量一大，任何 5% 到 10% 的成本差距，都會\u003Ca href=\"\u002Fnews\u002Fmmdiff-multimodal-feature-discovery-control-zh\">變成\u003C\u002Fa>很可觀的金額。\u003C\u002Fp>\u003Cp>Anthropic 不是第一家想碰這件事的公司。\u003Ca href=\"https:\u002F\u002Fai.google\u002Fdiscover\u002Four-models\u002Ftpu\u002F\" target=\"_blank\" rel=\"noopener\">Google\u003C\u002Fa> 很早就把 TPU 當成自家 \u003Ca href=\"\u002Ftag\u002Fai-\">AI 基礎設施\u003C\u002Fa>核心。\u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002F\" target=\"_blank\" rel=\"noopener\">OpenAI\u003C\u002Fa> 也長期依賴雲端與硬體夥伴。只是現在大家都更清楚，光靠外部供應商，成本談判空間有限。\u003C\u002Fp>\u003Cul>\u003Cli>自研晶片能針對 Claude 的推論流程做優化。\u003C\u002Fli>\u003Cli>外部 TPU 與 GPU 仍會保留，風險比較分散。\u003C\u002Fli>\u003Cli>成本下降後，企業客戶定價更好談。\u003C\u002Fli>\u003Cli>人才招募本身就是長期投入訊號。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>晶片團隊不是裝飾品\u003C\u002Fh2>\u003Cp>晶片團隊一旦成立，就不是找幾個硬體工程師來湊數。它牽涉架構、編譯器、記憶體配置、模型切分，還有和雲端供應商對接的實務問題。這些東西都很硬，也很花時間。\u003C\u002Fp>\u003Cp>對 AI 公司來說，最難的地方常常不是設計一顆晶片，而是把模型真的跑在上面。Claude 的工作負載和一般影像晶片完全不同。它要處理長上下文、即時回應，還要兼顧穩定性。\u003C\u002Fp>\u003Cp>因此，Anthropic 就算真的做出自家晶片，也不會一下子甩開現有供應商。比較合理的路線，是先讓部分推論工作轉過去，再慢慢擴大。這種節奏雖然不炫，但比較不會翻車。\u003C\u002Fp>\u003Cblockquote>「The only way to discover the limits of the possible is to go beyond them into the impossible.」— Arthur C. Clarke\u003C\u002Fblockquote>\u003Cp>這句話常被拿來形容技術嘗試。放在 Anthropic 身上，其實更像是在講經濟邏輯。做晶片不是為了帥，是為了把每次模型回應的成本壓低。\u003C\u002Fp>\u003Cp>而且這條路很吃團隊整合能力。晶片、模型、雲端、供應鏈，少一塊都不行。這也是為什麼多數 AI 公司最後都會回到一個老問題：自己做，還是跟別人買。\u003C\u002Fp>\u003Ch2>跟 Google、AWS、NVIDIA 的關係更像多線下注\u003C\u002Fh2>\u003Cp>Anthropic 目前看起來不是要切斷合作，而是增加籌碼。\u003Ca href=\"\u002Ftag\u002Fgoogle\">Google\u003C\u002Fa> TPU 擅長大規模 AI 計算，AWS 提供雲端彈性，\u003Ca href=\"\u002Ftag\u002Fnvidia\">NVIDIA\u003C\u002Fa> GPU 仍是市場主流。把這三條線都留著，談判空間就大很多。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786449785902-i38t.png\" alt=\"Anthropic 自建晶片團隊壓低 Claude 成本\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種策略在 AI 產業很常見。公司不想把命運綁死在單一硬體供應商身上，因為供貨、價格、排程都可能出問題。尤其當模型需求暴衝時，誰有現成算力，誰就有話語權。\u003C\u002Fp>\u003Cp>從產品角度看，Anthropic 也需要這種彈性。Claude 會跑在不同場景，從聊天到企業內部工具都有。不同工作負載，適合的硬體也不一樣。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Ftpu\" target=\"_blank\" rel=\"noopener\">Google TPU\u003C\u002Fa>：適合大規模矩陣運算。\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Faws.amazon.com\u002Fmachine-learning\u002Finferentia\u002F\" target=\"_blank\" rel=\"noopener\">AWS Inferentia\u003C\u002Fa>：偏向雲端推論成本控制。\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fdata-center\u002F\" target=\"_blank\" rel=\"noopener\">NVIDIA GPU\u003C\u002Fa>：生態完整，工具成熟。\u003C\u002Fli>\u003Cli>自研晶片：針對 Claude 工作負載做客製化。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這股風潮其實早就開始了\u003C\u002Fh2>\u003Cp>AI 公司自建硬體能力，已經不是新聞。Google 有 TPU，Meta 有自家 AI 晶片計畫，Amazon 也推 Inferentia 和 Trainium。大家都在做同一件事：把最貴的算力成本往內收。\u003C\u002Fp>\u003Cp>原因很簡單。模型越大，推論越頻繁，外部硬體的成本壓力就越明顯。當產品進入商業化階段，算力不再只是工程問題，而是營運問題。\u003C\u002Fp>\u003Cp>Anthropic 的新動作，放在這個脈絡裡就很合理。它不是在追求炫技，而是在補齊基礎設施。這種補法很無聊，但很有用。\u003C\u002Fp>\u003Cp>接下來可以觀察兩件事。第一，Anthropic 會不會公開更多晶片職缺。第二，Claude 的推論效率有沒有出現可量化改善。只要這兩點有進展，這支團隊就\u003Ca href=\"\u002Fnews\u002Ftts-evaluators-miss-more-than-naturalness-zh\">不只\u003C\u002Fa>是備案，而是成本戰的核心工具。\u003C\u002Fp>\u003Ch2>結論很簡單：Anthropic 在買時間，也在買控制權\u003C\u002Fh2>\u003Cp>我覺得這步棋很務實。短期內，Anthropic 還是得靠雲端與既有晶片夥伴供貨。長期來看，自研晶片能讓它在成本、效能、排程上多一層主動權。\u003C\u002Fp>\u003Cp>如果你在看 AI 產業，這類消息值得盯。下一波真正拉開差距的，可能不是誰模型分數高 2 分，而是誰能把每 1,000 萬次推論的成本壓得更低。那才是能\u003Ca href=\"\u002Fnews\u002Fdutch-government-llm-benchmark-values-zh\">不能\u003C\u002Fa>賺錢的分水嶺。\u003C\u002Fp>","Anthropic 開始招募自研晶片團隊，目標是壓低 Claude 的推論成本，同時仍維持 Google TPU、AWS 與 Nvidia 的算力合作。","www.forbes.com","https:\u002F\u002Fwww.forbes.com\u002Fsites\u002Fjonmarkman\u002F2026\u002F08\u002F06\u002Fanthropic-enters-the-ai-chip-race-with-in-house-chip-team\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786449787150-5apv.png","industry","zh","863e1343-f4e7-4749-a7d6-9620d798b8f9",[17,18,19,20,21,22,23],"Anthropic","Claude","自研晶片","推論成本","AI 基礎設施","TPU","NVIDIA GPU",[25,26,27],"Anthropic 正在招募自研晶片團隊，重點是降低 Claude 的推論成本。","公司仍會保留 Google TPU、AWS Inferentia 與 NVIDIA GPU 等外部算力來源。","AI 公司的競爭正在從模型能力，延伸到硬體與成本控制。",0,"2026-08-11T12:02:34.763912+00:00","2026-08-11T12:02:34.755+00:00",{"tags":32,"relatedLang":38,"relatedPosts":42},[33,34,36],{"name":20,"slug":20},{"name":17,"slug":35},"anthropic",{"name":18,"slug":37},"claude",{"id":15,"slug":39,"title":40,"language":41},"anthropic-builds-in-house-chip-team-claude-en","Anthropic Builds Its Own Chip Team for Claude","en",[43,49,55,61,67,73],{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"3ca7a087-abb7-4140-bcb6-f98ad55cb63b","5-banking-workflow-patterns-sas-viya-governed-zh","5 個 SAS Viya 銀行工作流模式","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786476769335-f9r4.png","2026-08-11T19:32:21.346196+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"0c7bc6ae-ecd8-4fa4-a643-0b646a4faf86","wall-street-backs-nvidia-ai-financing-push-zh","華爾街把 AI 基建當資產來融資","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786474977335-bst1.png","2026-08-11T19:02:34.11344+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"320bd82e-735e-4ab6-a57f-704be321a90a","claude-code-5-alternatives-2026-08-zh","Claude Code 不再一家獨大，5 個替代選擇更穩","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786471381965-yk9h.png","2026-08-11T18:02:32.694634+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"729066f1-3d80-4bef-9ac5-4e05b6b6c152","anthropic-macquarie-gic-data-centers-zh","Anthropic 砸錢蓋專用資料中心","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786433580513-mfwp.png","2026-08-11T07:32:29.317071+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"1293526e-2e6e-4fb3-9106-ed43965ac821","2027-ai-capex-reorders-narrative-zh","2027年AI资本开支重排叙事的5个关键信号","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786384978906-p709.png","2026-08-10T18:02:30.380276+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":13},"5972deb9-e5b9-458a-ae6e-6949652e872c","openai-anthropic-take-80-percent-ai-50-funding-zh","OpenAI、Anthropic 吃掉 AI 50 資金","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786366986001-3yt3.png","2026-08-10T13:02:34.99251+00:00",[80,85,90,95,100,105,110,115,120,125],{"id":81,"slug":82,"title":83,"created_at":84},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":86,"slug":87,"title":88,"created_at":89},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]