[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-microsoft-bare-metal-aks-ai-training-zh":3,"article-related-microsoft-bare-metal-aks-ai-training-zh":34,"series-industry-21f2afc6-8551-47be-8d10-639f01864016":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":26,"views":30,"created_at":31,"published_at":32,"topic_cluster_id":33},"21f2afc6-8551-47be-8d10-639f01864016","microsoft-bare-metal-aks-ai-training-zh","微軟把 AKS 推向 AI 訓練核心","\u003Cp data-speakable=\"summary\">微軟替 Azure Kubernetes Service 加上 bare metal、Fleet Manager 和 AI 部署工具，明顯是在把 AKS 往\u003Ca href=\"\u002Ftag\u002F企業-ai\">企業 AI\u003C\u002Fa> 訓練與推論場景推進。\u003C\u002Fp>\u003Cp>這次更新不是小修小補。微軟在 \u003Ca href=\"https:\u002F\u002Fbuild.microsoft.com\u002F\" target=\"_blank\" rel=\"noopener\">Microsoft Build 2026\u003C\u002Fa> 端出一串新東西，核心是 \u003Ca href=\"https:\u002F\u002Fazure.microsoft.com\u002Fen-us\u002Fproducts\u002Fkubernetes-service\" target=\"_blank\" rel=\"noopener\">Azure Kubernetes Service\u003C\u002Fa> 的 bare metal 能力。講白了，就是讓工作負載更直接碰到硬體。\u003C\u002Fp>\u003Cp>對 AI 團隊來說，這種改動很實際。延遲、\u003Ca href=\"\u002Ftag\u002Fgpu\">GPU\u003C\u002Fa> 存取、網路拓樸、每一層虛擬化開銷，都會影響訓練時間和雲端帳單。微軟也同步補上 \u003Ca href=\"https:\u002F\u002Fazure.microsoft.com\u002Fen-us\u002Fproducts\u002Fkubernetes-fleet-manager\" target=\"_blank\" rel=\"noopener\">Azure Kubernetes Fleet Manager\u003C\u002Fa>，還有 Ray 和\u003Ca href=\"\u002Fnews\u002Fextracted-prompts-turn-model-behavior-into-a-map-zh\">模型\u003C\u002Fa>部署工具，整套味道很明顯。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>功能\u003C\u002Fth>\u003Cth>狀態\u003C\u002Fth>\u003Cth>作用\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>AKS on Bare Metal\u003C\u002Ftd>\u003Ctd>Public preview\u003C\u002Ftd>\u003Ctd>直接存取硬體，減少虛擬化層\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Azure Kubernetes Fleet Manager\u003C\u002Ftd>\u003Ctd>已公布\u003C\u002Ftd>\u003Ctd>統一政策與工作負載配置\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Anyscale on Azure\u003C\u002Ftd>\u003Ctd>已推出\u003C\u002Ftd>\u003Ctd>託管 Ray 服務，處理分散式 AI 工作\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>AI Runway\u003C\u002Ftd>\u003Ctd>已推出\u003C\u002Ftd>\u003Ctd>Kubernetes 原生模型部署框架\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>bare metal 是這次最有感的改動\u003C\u002Fh2>\u003Cp>AKS on Bare Metal 是整場更新的主角。微軟說，這個功能現在是 public preview，目標是讓 AI 工作負載直接接觸硬體，不必再多經過一層虛擬化。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782540169706-uqoi.png\" alt=\"微軟把 AKS 推向 AI 訓練核心\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這件事聽起來很工程師，但其實很務實。訓練大\u003Ca href=\"\u002Fnews\u002Fopenai-latest-model-us-user-vetting-zh\">模型\u003C\u002Fa>很吃吞吐量，推論也很怕延遲。只要少一層開銷，常常就能換來更好的 GPU 利用率，甚至少一點雲端費用。\u003C\u002Fp>\u003Cp>微軟還特別提到 \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fdata-center\u002Fnvlink\u002F\" target=\"_blank\" rel=\"noopener\">NVLink\u003C\u002Fa> 和 \u003Ca href=\"https:\u002F\u002Fwww.rdma.org\u002F\" target=\"_blank\" rel=\"noopener\">RDMA\u003C\u002Fa>。這代表它不是只想把 Kubernetes 包裝得更漂亮，而是想讓 Azure 更像 \u003Ca href=\"\u002Ftag\u002Fai-\">AI 基礎設施\u003C\u002Fa>，不只是通用雲端。\u003C\u002Fp>\u003Cul>\u003Cli>bare metal 直接瞄準高效能 AI 訓練。\u003C\u002Fli>\u003Cli>NVLink 和 RDMA 對 GPU 叢集很重要。\u003C\u002Fli>\u003Cli>少一層虛擬化，通常就少一點延遲。\u003C\u002Fli>\u003Cli>對付昂貴加速器時，效率比口號更值錢。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Fleet Manager 管的是混合雲現實\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fazure.microsoft.com\u002Fen-us\u002Fproducts\u002Fkubernetes-fleet-manager\" target=\"_blank\" rel=\"noopener\">Azure Kubernetes Fleet Manager\u003C\u002Fa> 看起來沒那麼炫，但它解的是企業真問題。AI 叢集很少只待在單一區域。很多團隊會把工作拆到不同地點，原因可能是延遲、法規，或是容量不夠。\u003C\u002Fp>\u003Cp>這時候最麻煩的不是跑不動，而是不好管。政策要一致，工作負載要能移，還要有人知道哪個叢集在吃哪種資源。Fleet Manager 的定位，就是幫微軟補上這層中央控管。\u003C\u002Fp>\u003Cp>這種思路也跟 Kubernetes 的本質很搭。Kubernetes 本來就不是只有排程而已，它更像一個控制平面。只不過 AI 時代把這件事講得更直接：你不只要能跑，還要能管。\u003C\u002Fp>\u003Cblockquote>“A cluster is only as useful as the policy and automation around it,” said \u003Ca href=\"https:\u002F\u002Fkubernetes.io\u002F\" target=\"_blank\" rel=\"noopener\">Kubernetes\u003C\u002Fa> co-founder Brendan Burns in a 2024 CNCF interview.\u003C\u002Fblockquote>\u003Cp>這句話放在這次更新上很貼切。微軟不是在賣一個單獨的叢集功能。它是在賣一整套 AI 控管層，讓企業可以少一點手工維運。\u003C\u002Fp>\u003Cp>如果你看雲端市場，這套路其實很合理。\u003Ca href=\"https:\u002F\u002Faws.amazon.com\u002F\" target=\"_blank\" rel=\"noopener\">AWS\u003C\u002Fa> 和 \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002F\" target=\"_blank\" rel=\"noopener\">Google Cloud\u003C\u002Fa> 都在補 AI 能力。微軟的做法是把開源工具和託管服務黏在一起，讓企業有彈性，也有人幫忙收爛攤子。\u003C\u002Fp>\u003Ch2>Ray 和 AI Runway 是開發流程的補洞\u003C\u002Fh2>\u003Cp>微軟也把 \u003Ca href=\"https:\u002F\u002Fwww.anyscale.com\u002F\" target=\"_blank\" rel=\"noopener\">Anyscale\u003C\u002Fa> 帶進 Azure，做成託管的 \u003Ca href=\"https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002F\" target=\"_blank\" rel=\"noopener\">Ray\u003C\u002Fa> 服務。Ray 在分散式訓練、批次推論、資料處理都很常見。把它託管化，對團隊來說就是少掉很多安裝和維護成本。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782540168730-ipx3.png\" alt=\"微軟把 AKS 推向 AI 訓練核心\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>另一個是 \u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fai\" target=\"_blank\" rel=\"noopener\">AI Runway\u003C\u002Fa>。微軟把它定義成 Kubernetes-native 的模型部署框架。這個點很重要，因為很多 AI 專案不是卡在訓練，而是卡在上線。\u003C\u002Fp>\u003Cp>說得更白一點，Notebook 跑得動不代表 production 撐得住。模型版本、資源配置、流量切換、回滾機制，這些才是每天會咬人的地方。AI Runway 的價值，就是把這條路縮短一點。\u003C\u002Fp>\u003Cul>\u003Cli>Anyscale on Azure 適合分散式 AI 工作。\u003C\u002Fli>\u003Cli>Ray 對平行訓練和批次推論很常用。\u003C\u002Fli>\u003Cli>AI Runway 針對模型部署流程。\u003C\u002Fli>\u003Cli>這些工具都在補從實驗到上線的落差。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這次更新很像微軟的老套路\u003C\u002Fh2>\u003Cp>微軟一向很會做一件事：把熱門開源工具包進雲端服務，再加上企業要的治理和整合。這次 AKS 更新也是同一套邏輯。Kubernetes、Ray、AI 部署框架，全都圍著 Azure 轉。\u003C\u002Fp>\u003Cp>這種\u003Ca href=\"\u002Fnews\u002Fdeepseek-low-cost-chatbot-changed-ai-pricing-zh\">策略\u003C\u002Fa>的好處很明確。開發者不用全部重寫，企業也不用自己從零搭平台。壞處也很明確，就是你會越來越依賴雲端供應商的整體設計。這點很現實，沒什麼好粉飾。\u003C\u002Fp>\u003Cp>如果把這次更新拆開看，會發現它不是單點功能，而是三層一起補：硬體層、控制層、部署層。這種組合拳才是微軟真正想推的東西。\u003C\u002Fp>\u003Cul>\u003Cli>硬體層：bare metal AKS。\u003C\u002Fli>\u003Cli>控制層：Fleet Manager。\u003C\u002Fli>\u003Cli>部署層：Anyscale 和 AI Runway。\u003C\u002Fli>\u003Cli>整體方向就是把 Azure 做成 AI 作業系統。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>企業會在意的，其實只有三個數字\u003C\u002Fh2>\u003Cp>AI 基礎設施最後都會回到幾個數字。第一個是訓練時間。第二個是推論延遲。第三個是每次訓練或每千次推論的成本。其他名詞都可以先放旁邊。\u003C\u002Fp>\u003Cp>bare metal 的意義，就是看它能不能把這三個數字壓下來。Fleet Manager 的意義，就是看它能不能把跨區域和混合雲管理成本壓下來。Ray 和 AI Runway 的意義，則是看它能不能縮短交付時間。\u003C\u002Fp>\u003Cp>如果這三件事都做得到，AKS 就不再只是容器平台。它會變成企業 AI 的操作層。這也是微軟這波更新最值得盯的地方。\u003C\u002Fp>\u003Ch2>接下來要看的是實測，不是簡報\u003C\u002Fh2>\u003Cp>我覺得這次更新方向對了，但真正的考驗還沒來。微軟現在講的是能力，市場接下來要看的是數字。像是 GPU 利用率、訓練吞吐量、部署時間，還有實際省下多少成本。\u003C\u002Fp>\u003Cp>如果 Azure 之後能拿出 bare metal AKS 的實測結果，這套故事就更完整。反過來說，如果沒有硬數據，大家很快就會把它當成另一版雲端包裝。\u003C\u002Fp>\u003Cp>對開發者來說，現在最實際的動作很簡單：先看你目前的 AI 工作流，有多少時間花在基礎設施，而不是模型本身。那個比例越高，就越值得關注這次 AKS 更新。\u003C\u002Fp>","微軟在 Azure Kubernetes Service 加入 bare metal、Fleet Manager 和 Ray 服務，直接把 AKS 往 AI 訓練與推論場景推進，重點放在效能、治理和部署流程。","techgig.com","https:\u002F\u002Ftechgig.com\u002Famp\u002Fnews\u002Fcloud-infrastructure\u002Fmicrosoft-enhances-azure-kubernetes-service-for-ai-workloads\u002F131954659",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782540169706-uqoi.png","industry","zh","45baf201-87d8-460d-9f9b-7c8cf48e0f52",[17,18,19,20,21,22,23,24,25],"Microsoft","Azure Kubernetes Service","AKS","bare metal","AI training","inference","Kubernetes","Ray","Fleet Manager",[27,28,29],"AKS on Bare Metal 把 AI 工作負載更直接接到硬體。","Fleet Manager 補的是混合雲與多叢集治理。","Anyscale on Azure 和 AI Runway 針對分散式工作與部署流程。",0,"2026-06-27T06:02:26.373585+00:00","2026-06-27T06:02:26.364+00:00","166ad762-eb22-4df8-83f5-f344194edd88",{"tags":35,"relatedLang":38,"relatedPosts":42},[36],{"name":17,"slug":37},"microsoft",{"id":15,"slug":39,"title":40,"language":41},"microsoft-bare-metal-aks-ai-training-en","Microsoft adds bare metal AKS for AI training","en",[43,49,55,61,67,73],{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"576f1de0-bbf9-4a91-96bf-a1bf6ff4c67c","us-model-curbs-security-deals-not-bans-zh","美國應以安全協議解除模型管制，而非一刀切禁令","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782658969312-tf30.png","2026-06-28T15:02:19.927898+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"c1d71ae5-dabd-4778-8326-7645316004c2","meta-replacing-moderators-with-ai-to-cut-costs-zh","Meta 用 AI 取代審核員，省錢先上","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782653576451-arn6.png","2026-06-28T13:32:29.737246+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"08c94bd8-e6b6-4328-82ff-bee0a7cef126","meta-ai-moderation-push-is-the-wrong-tradeoff-zh","Meta 把 AI 用在內容審核上，這筆交換不划算","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782652669314-in2k.png","2026-06-28T13:17:21.733509+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"6ad43bed-fc6b-4bc6-a728-38362a29ffec","meta-ai-content-moderation-human-reviews-zh","Meta 內容審核轉向 AI 的 5 個關鍵","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782651773409-llaq.png","2026-06-28T13:02:22.855907+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"304a413f-48b5-4e03-ae02-805b048d6023","2026-xiang-liang-zi-liao-ku-dui-bi-10-kuan-zen-me-xuan-zh","2026 向量資料庫對比：10 款怎麼選","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782649107089-ppzf.png","2026-06-28T12:17:57.21576+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":13},"74273718-6aab-49c4-8c0c-3fc4c534757d","aaas-new-protocol-gives-ai-agents-legal-backup-zh","AAA 讓 AI 代理先補上法律紀錄","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782646375320-jxqg.png","2026-06-28T11:32:29.265409+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"]