Microsoft unveils Maia 300 in-house AI chip, plans to deploy over a million units next year to cut compute costs微软推出 Maia 300 自研 AI 芯片，明年拟部署超百万颗以降低算力成本

Microsoft unveiled a new generation of its in-house AI chip on Monday local time and plans to deploy more than a million units next year to support its internal AI business. Called Maia 300, it is the company’s third-generation custom silicon, built specifically for AI data operations such as code generation and carrying out complex business tasks on a user’s behalf. The previous Maia generation already runs Microsoft’s own models as well as OpenAI models hosted on Azure. The launch lands as cloud providers and AI companies including OpenAI rush out custom chips aimed at breaking Nvidia’s grip on the AI compute market. Architecturally Microsoft took a different route from Nvidia, whose accelerators pair one giant compute die with high-bandwidth memory: Maia 300 uses a chiplet design that packages several specialised dies together. The compute chiplets are made on TSMC’s 2nm process, while the dies handling data movement and in-system scheduling use a cheaper 3nm process. A single Maia 300 rack holds 96 chips and draws around 167kW, versus roughly 142kW for Nvidia’s latest GB300 rack with 72 GPUs. Microsoft says Maia 300 delivers about 40% more useful compute per megawatt than “mainstream GPU systems”, without naming the comparison product. Rani Borkar, president of Azure hardware systems and infrastructure, told Reuters: “By optimising the whole stack, this chip can deliver more intelligence per watt, per unit cost and per byte of memory.” Microsoft plans multi-gigawatt-scale deployment of Maia 300 starting next year, and says capital spending this calendar year is expected to exceed $175 billion.