[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-anthropic-custom-ai-inference-chips-claude-zh":3,"article-related-anthropic-custom-ai-inference-chips-claude-zh":32,"series-industry-0ffd4803-74bb-43f8-a5be-0a5681ecc049":81},{"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":25,"views":29,"created_at":30,"published_at":31,"topic_cluster_id":11},"0ffd4803-74bb-43f8-a5be-0a5681ecc049","anthropic-custom-ai-inference-chips-claude-zh","Anthropic 也要自己做 AI 晶片","\u003Cp data-speakable=\"summary\">\u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa> 正在為 \u003Ca href=\"\u002Ftag\u002Fclaude\">Claude\u003C\u002Fa> 研發自家推論晶片，目標是降低對 \u003Ca href=\"\u002Ftag\u002Fnvidia\">Nvidia\u003C\u002Fa> GPU 的依賴，並把推論成本壓下來。\u003C\u002Fp>\u003Cp>這件事很直白。AI 公司開始算硬體帳了。Anthropic 想把 Claude 的推論成本壓低，因為租 GPU 太貴，尤其是大量 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa> 流量進來時。\u003C\u002Fp>\u003Cp>這也不是單純的內部實驗。Anthropic 牽涉到\u003Ca href=\"\u002Fnews\u002Ftest-time-harnesses-weak-model-transfer-zh\">模型\u003C\u002Fa>、雲端、晶片供應鏈，還可能碰上 \u003Ca href=\"https:\u002F\u002Fwww.samsung.com\u002Fsemiconductor\u002F\" target=\"_blank\" rel=\"noopener\">Samsung\u003C\u002Fa> 這類製造夥伴。它的方向很清楚，就是把推論這筆錢握回自己手上。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>數字\u003C\u002Fth>\u003Cth>意義\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>ASIC co-design 市占集中度\u003C\u002Ftd>\u003Ctd>約 95%\u003C\u002Ftd>\u003Ctd>顯示 custom AI 晶片設計市場很集中\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.broadcom.com\u002F\" target=\"_blank\" rel=\"noopener\">Broadcom\u003C\u002Fa> backlog\u003C\u002Ftd>\u003Ctd>730 億美元\u003C\u002Ftd>\u003Ctd>代表需求很強，案子排到很後面\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Broadcom 預估 AI 晶片年營收\u003C\u002Ftd>\u003Ctd>2027 年底超過 1000 億美元\u003C\u002Ftd>\u003Ctd>顯示 custom silicon 已經是大生意\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.marvell.com\u002F\" target=\"_blank\" rel=\"noopener\">Marvell\u003C\u002Fa> 預估 co-design 營收\u003C\u002Ftd>\u003Ctd>2026 年超過 110 億美元\u003C\u002Ftd>\u003Ctd>說明推論晶片合約的價值不低\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>Anthropic 為什麼要碰晶片\u003C\u002Fh2>\u003Cp>先講結論。推論才是燒錢主戰場。訓練模型很貴，但訓練完之後，真正持續出帳單的是推論。Claude 每多回一個 token，就多一點成本。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786645974543-pram.png\" alt=\"Anthropic 也要自己做 AI 晶片\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>對 Anthropic 來說，這個壓力很現實。企業客戶、政府合約、\u003Ca href=\"\u002Ftag\u002Fagent\">Agent\u003C\u002Fa> \u003Ca href=\"\u002Fnews\u002Fopen-generative-ai-github-studio-breakdown-zh\">工作\u003C\u002Fa>流，全都會把 token 消耗拉高。當使用量變大，通用 GPU 的成本就會越來越刺眼。\u003C\u002Fp>\u003Cp>自家晶片的好處很直接。它可以針對 Claude 的推論型態做優化。像是記憶體存取、延遲、功耗、批次處理，都能一起調。這種優化，租現成 GPU 很難做到。\u003C\u002Fp>\u003Cul>\u003Cli>推論工作負載可針對模型形狀做最佳化。\u003C\u002Fli>\u003Cli>自訂晶片可把總持有成本壓低，報導提到最多可降 65%。\u003C\u002Fli>\u003Cli>Agent 系統通常吃掉更多 token。\u003C\u002Fli>\u003Cli>企業部署越多，GPU 帳單越難看。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這場晶片戰，早就不是新鮮事\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Ftpu\" target=\"_blank\" rel=\"noopener\">Google\u003C\u002Fa> 很早就做 TPU。\u003Ca href=\"https:\u002F\u002Fwww.amazon.com\u002Fawslabs\u002F\" target=\"_blank\" rel=\"noopener\">Amazon\u003C\u002Fa> 有 Trainium 和 Inferentia。\u003Ca href=\"https:\u002F\u002Fwww.meta.com\u002F\" target=\"_blank\" rel=\"noopener\">Meta\u003C\u002Fa> 有 MTIA。\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fai\" target=\"_blank\" rel=\"noopener\">Microsoft\u003C\u002Fa> 有 Maia。連 \u003Ca href=\"https:\u002F\u002Fopenai.com\u002F\" target=\"_blank\" rel=\"noopener\">OpenAI\u003C\u002Fa> 也在碰自製推論硬體。\u003C\u002Fp>\u003Cp>原因很簡單。AI 規模一大，長期租 GPU 就像一直刷高額信用卡。大公司會想要自己的晶片，因為它們有自己的模型、自己的資料中心，也有自己的 serving stack。\u003C\u002Fp>\u003Cp>Anthropic 這步走得不算早，但也不算奇怪。它的 Claude 已經不是小型 demo，而是實際上線的產品。只要 token 流量夠大，自建晶片就會變成一個很務實的選項。\u003C\u002Fp>\u003Cblockquote>“The chips will allow Claude to run faster and more efficiently at the scale its customers need,” Anthropic told \u003Ca href=\"https:\u002F\u002Fwww.businessinsider.com\u002F\" target=\"_blank\" rel=\"noopener\">Business Insider\u003C\u002Fa>.\u003C\u002Fblockquote>\u003Ch2>Samsung、Broadcom、Marvell 各有什麼角色\u003C\u002Fh2>\u003Cp>如果 Anthropic 真找上 \u003Ca href=\"https:\u002F\u002Fwww.samsung.com\u002Fsemiconductor\u002Ffoundry\u002F\" target=\"_blank\" rel=\"noopener\">Samsung Foundry\u003C\u002Fa>，那代表它想要的是可量產的製造能力。Foundry 這一關很重要，因為設計出來不代表能順利量產，尤其是 AI 晶片對良率和封裝要求都很高。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786645968680-sq6o.png\" alt=\"Anthropic 也要自己做 AI 晶片\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>設計端則是另一個世界。custom ASIC 這塊，現在幾乎被 \u003Ca href=\"https:\u002F\u002Fwww.broadcom.com\u002F\" target=\"_blank\" rel=\"noopener\">Broadcom\u003C\u002Fa> 和 \u003Ca href=\"https:\u002F\u002Fwww.marvell.com\u002F\" target=\"_blank\" rel=\"noopener\">Marvell\u003C\u002Fa> 包走。兩家加起來大約吃下 95% 的 co-design 市場，這數字很誇張，也很現實。\u003C\u002Fp>\u003Cp>這些數字透露出一件事。AI 晶片不是只有 Nvidia 一家在賺。真正的供應鏈上游，還有設計服務、封裝、製造、互連和記憶體。每一層都有人收錢，而且都不便宜。\u003C\u002Fp>\u003Cul>\u003Cli>Broadcom 和 Marvell 合計約占 95% 的 ASIC co-design 市場。\u003C\u002Fli>\u003Cli>Broadcom backlog 達 730 億美元。\u003C\u002Fli>\u003Cli>Broadcom 預估 2027 年底 AI 晶片年營收超過 1000 億美元。\u003C\u002Fli>\u003Cli>Marvell 預估 2026 年 co-design 營收超過 110 億美元。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>對 Nvidia 和 Claude 使用者的影響\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fdata-center\u002Fgpu-cloud-computing\u002F\" target=\"_blank\" rel=\"noopener\">Nvidia\u003C\u002Fa> 不會馬上被踢下場。訓練大模型還是很吃它的 GPU，而且 CUDA 生態系還是很多團隊的預設解法。\u003C\u002Fp>\u003Cp>但推論這一段，壓\u003Ca href=\"\u002Fnews\u002Fanthropic-riot-deal-ai-compute-billions-zh\">力已\u003C\u002Fa>經很明顯。模型訓練完之後，成本會在上線階段持續累積。誰能把每次回應的成本壓低，誰就有空間把價格打得更兇。\u003C\u002Fp>\u003Cp>對 Claude 使用者來說，最直接的變化可能是更便宜的服務費，或更穩的延遲表現。對 Anthropic 來說，這代表它有機會把毛利守住，不會被硬體成本吃掉太多。\u003C\u002Fp>\u003Cp>但自家晶片也有風險。設計費、驗證費、封裝費、軟體適配，全都很貴。只有當工作負載夠大，而且模型行為夠穩，這筆帳才划算。Anthropic 顯然覺得 Claude 已經到這個門檻。\u003C\u002Fp>\u003Ch2>這背後其實是 AI 經濟學\u003C\u002Fh2>\u003Cp>很多人把 AI 只看成模型能力競賽，我覺得這看太淺。真正決定勝負的，常常是單位成本。每 100 萬 token 要花多少錢，才是企業會盯的數字。\u003C\u002Fp>\u003Cp>這也是為什麼大家都想碰硬體。當模型變大、流量變多、應用變複雜，軟體優化會碰到上限。到了那一步，晶片就不是加分題，而是成本題。\u003C\u002Fp>\u003Cp>Anthropic 現在的動作，等於在回答一個很務實的問題：當 Claude 的使用量繼續上升時，誰來接住那張 GPU 帳單。它選擇自己下場做晶片，答案很清楚，也很有壓力感。\u003C\u002Fp>\u003Ch2>接下來要看什麼\u003C\u002Fh2>\u003Cp>我會先看三件事。第一，Anthropic 是否公開更多晶片規格。第二，Samsung 是否真的接下製造。第三，這批晶片能不能在推論成本上打贏現成 GPU。\u003C\u002Fp>\u003Cp>如果這三件事都成立，Anthropic 就不只是做模型公司。它會變成一個同時懂軟體和硬體成本的 AI 供應商。這條路很硬，但很符合現在的 AI 現實。\u003C\u002Fp>\u003Cp>接下來幾季，重點就是看 Claude 的推論價格有沒有下來。只要價格和延遲有改善，這筆投資就算走對了。\u003C\u002Fp>","Anthropic 正在為 Claude 研發自家推論晶片，目標是降低對 Nvidia GPU 的依賴，並把推論成本壓下來。","www.tomshardware.com","https:\u002F\u002Fwww.tomshardware.com\u002Ftech-industry\u002Fanthropic-to-build-its-own-co-designed-custom-ai-accelerator-for-inferencing-workloads-samsung-reported-to-be-partnering-with-the-claude-ai-maker-for-manufacturing",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786645974543-pram.png","industry","zh","bf672b43-7312-47e4-aadf-9d09e153446f",[17,18,19,20,21,22,23,24],"Anthropic","Claude","AI 晶片","推論晶片","Nvidia","Samsung Foundry","Broadcom","Marvell",[26,27,28],"Anthropic 正在做自家推論晶片，核心目標是壓低 Claude 的硬體成本。","custom silicon 的競爭已經很擁擠，Broadcom 和 Marvell 幾乎主導這個市場。","如果晶片計畫成功，Claude 的推論價格和延遲都有機會改善。",1,"2026-08-13T18:32:29.543527+00:00","2026-08-13T18:32:29.524+00:00",{"tags":33,"relatedLang":40,"relatedPosts":44},[34,36,38],{"name":21,"slug":35},"nvidia",{"name":17,"slug":37},"anthropic",{"name":18,"slug":39},"claude",{"id":15,"slug":41,"title":42,"language":43},"anthropic-custom-ai-inference-chips-claude-en","Anthropic is building custom AI chips for Claude","en",[45,51,57,63,69,75],{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"1c5f0bb7-ea6e-413d-8e72-b02f20316ef4","astra-fangman-yanfa-5-ge-guan-jian-xin-hao-zh","Astra放慢研发的5个关键信号","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786647774714-k0m5.png","2026-08-13T19:02:29.945739+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"e0d0476a-e3d5-4bf7-995b-e5416f5e392a","moka-ai-hrms-three-layer-architecture-zh","Moka AI 不是锦上添花，HRMS 选型正在转向三层架构","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786611776096-cy66.png","2026-08-13T09:02:28.103219+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"3640e980-0be1-408e-b5fe-a1ad42022b61","ai-hardware-rally-short-sellers-august-2026-zh","AI硬件反攻，8月空頭更難做","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786584781842-0bny.png","2026-08-13T01:32:36.671063+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"f1349fae-7cbd-49a0-9d6a-ceaa65b51c1e","pixel-11-launch-live-google-reveals-zh","Pixel 11 發表會重點整理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786581163428-ivyj.png","2026-08-13T00:32:20.178966+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"07408fa7-809f-49e2-873d-427cb10c0c15","claude-invisible-watermark-right-direction-zh","Claude隱形水印是正確方向，不是暴政","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786579370713-ic6h.png","2026-08-13T00:02:25.128006+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":13},"d8c7a2ca-123e-49bf-a1c2-df856e63fce8","dockers-latest-releases-security-compose-zh","Docker 這 5 版先看安全與 Compose","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786557772075-zir4.png","2026-08-12T18:02:27.019175+00:00",[82,87,92,97,102,107,112,117,122,127],{"id":83,"slug":84,"title":85,"created_at":86},"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":88,"slug":89,"title":90,"created_at":91},"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":93,"slug":94,"title":95,"created_at":96},"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":98,"slug":99,"title":100,"created_at":101},"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":103,"slug":104,"title":105,"created_at":106},"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":108,"slug":109,"title":110,"created_at":111},"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":113,"slug":114,"title":115,"created_at":116},"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":118,"slug":119,"title":120,"created_at":121},"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":123,"slug":124,"title":125,"created_at":126},"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":128,"slug":129,"title":130,"created_at":131},"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"]