Microsoft plans multi-gigawatt deployments of its own Maia AI chips from 2027

On 12 October 2026 Microsoft said it will scale its own Azure Maia AI chips to multi-gigawatt deployments starting in 2027, and make Maia the primary inference system for Microsoft products and services. Scott Guthrie, who runs Microsoft's cloud and AI group, wrote it in an official Microsoft blog post. A gigawatt is 1,000 megawatts, a measure of how much electricity these chips draw.

Inference is the work of answering: every time Copilot replies or an AI agent takes a step, a chip runs the model. Training builds the model. Microsoft is saying most of the answering for its own products will move onto chips it designs itself.

It also unveiled Maia 300, which begins production deployments in 2027. Microsoft expects it to deliver 40% more tokens per megawatt than leading GPU (graphics chip) systems. Tokens are the small chunks of text an AI reads and writes, so tokens per megawatt is how much AI work you get for the same power. Its main compute parts are made by TSMC on N2P, one of its newest chipmaking processes. The current Maia 200 already serves Copilot and runs OpenAI models, Microsoft's own MAI models and open-source models.

Why we file it. When one of the biggest AI spenders says it will build multiple gigawatts of its own chips, that is a capacity plan, not just a product launch. Over time it could also move more of Microsoft's chip spending to its own designs, though Microsoft hasn't said it will spend less on GPUs.

How we read it. Microsoft goes from gray to green on the map: a bullish capacity signal. It is still a plan. Microsoft gave no gigawatt figure, cost or site list, and a Microsoft manager declined to give a timeline, CNBC reported. Microsoft still works with NVIDIA and AMD for training, and the efficiency figure is its own expectation, not an outside test.

Why it matters to you: Copilot and other Microsoft AI features you use would run more and more on Microsoft's own chips, which Microsoft says get more AI work out of the same power.

What would change this: a stated gigawatt figure or capex line, slips in the 2027 production start, or Microsoft pulling back to merchant GPUs for inference.