跳到主要內容

Maia 300評測:微軟自研AI晶片,單年要出貨百萬顆 | Maia 300 Review: Microsoft's AI Chip Bets on 1M Units

By Kit 小克 | AI Tool Observer | 2026-10-13

🇹🇼 Maia 300評測:微軟自研AI晶片,單年要出貨百萬顆

Maia 300 是微軟在 2026 年 10 月 12 日 OCP Global Summit 上公開的第三代自研 AI 晶片,目標很明確:擺脫對 Nvidia、AMD 的依賴,把 Azure 的推論主力換成自己的硬體。微軟宣稱這顆晶片的效能是「每百萬瓦可處理的 token 數比主流 GPU 系統多 40%」,96 顆晶片組成的機架吞吐量則是前代 Maia 200 的 7 倍。

為什麼現在做自己的晶片

這不是微軟第一次嘗試自研晶片,但 Maia 300 是第一次喊出「要當 Azure 主要推論系統」的目標 —— 不是輔助,是主力。背後邏輯很現實:Nvidia GPU 又貴又缺貨,Amazon、Google、Meta 早就在走自研晶片這條路,微軟現在才追上並不奇怪,真正的問題是規模能不能做出來。

根據外媒報導,微軟已經在跟 TSMC 談產能,第一批目標是 30 萬顆,長期野心是衝到 單年超過百萬顆。量產預計從 2027 年開始,屆時會支撐微軟多個 GW(gigawatt)規模的資料中心部署。

Maia 300 的真實意義:不是炫技,是省錢

  • 推論優先架構:Maia 300 針對長上下文、記憶體密集、decode-heavy 的工作負載設計 —— 說白了就是為 Copilot、Teams 這類已經在跑的產品量身訂做,不是拿來訓練下一代大模型。
  • 降低對單一供應商的曝險:Nvidia 一家獨大的局面讓所有雲端巨頭都緊張,自研晶片是議價籌碼,也是供應鏈保險。
  • 效能數字要打折扣看:「40% 更省電」是微軟自己公布的數字,沒有第三方基準測試佐證,實際生產環境的表現如何,要等明年真正大規模上線才知道。

小克的看法

自研晶片這條路,Google TPU 走了快十年才真正成熟,Amazon Trainium 也磨了好幾代。微軟現在才喊出「單年百萬顆」的目標,代表的是決心,不是現狀——真正考驗在 2027 年量產之後,晶片能不能撐住大規模生產環境,而不是發布會上的 PPT 數字。對台灣的晶片供應鏈(特別是 TSMC)來說,這又是一張穩定的大單,但對 Nvidia 的護城河來說,這是又一塊磚被敲掉。

好不好用,試了才知道。


🇺🇸 Maia 300 Review: Microsoft's AI Chip Bets on 1M Units

Maia 300 is Microsoft's third-generation custom AI chip, unveiled on October 12, 2026 at the OCP Global Summit. The goal is blunt: cut dependence on Nvidia and AMD, and make Azure's own silicon the primary inference engine instead of a side project. Microsoft claims the chip delivers 40% more tokens per megawatt than leading GPU-based systems, with a 96-chip rack configuration hitting 7x the throughput of the previous Maia 200.

Why Microsoft Is Building Its Own Chips Now

This isn't Microsoft's first attempt at custom silicon, but Maia 300 is the first time the company has explicitly said it wants this to be the primary AI inference system for its products, not a backup option. The logic is simple: Nvidia GPUs are expensive and supply-constrained, and Amazon, Google, and Meta have already been running their own custom chip programs for years. Microsoft catching up isn't surprising; whether it can scale is the real question.

According to reporting, Microsoft is in talks with TSMC for manufacturing capacity, targeting an initial 300,000 units with an ambition to eventually scale past 1 million units in a single year. Production is slated to ramp starting in 2027, supporting multi-gigawatt data center deployments.

What Maia 300 Actually Means, Cost Control Not a Flex

  • Inference-first architecture: Maia 300 is built for long-context, memory-intensive, decode-heavy workloads, tuned for running Copilot and Teams at scale, not training the next frontier model.
  • Reducing single-vendor risk: Nvidia's dominance makes every hyperscaler nervous. Custom silicon is both a bargaining chip and supply-chain insurance.
  • Take the efficiency numbers with a grain of salt: The "40% more efficient" figure comes from Microsoft itself, with no independent benchmark attached. Real-world performance at scale won't be known until deployment actually ramps next year.

Kit's Take

Google took nearly a decade to make TPUs genuinely competitive, and Amazon's Trainium chips went through several generations before they mattered. Microsoft setting a "1 million units a year" target now signals intent, not current reality, the real test comes after 2027 production actually ramps, not the slide deck at a conference. For Taiwan's chip supply chain, especially TSMC, this is another stable mega-order. For Nvidia's moat, it's one more brick knocked loose.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code