Microsoft is preparing to move its in-house Azure Maia AI chip from a supporting role in the lab onto the main stage. In an article, executive Scott Guthrie revealed that the company plans to deploy Maia chips at multi-gigawatt scale starting in 2027, making them Microsoft's primary inference system across its product lines.
The current Maia200 is already running Microsoft Copilot in the US, Europe, and Australia, and GitHub Copilot and Foundry will follow by the end of the year. CEO Satya Nadella called this approach "our biggest bet on our own silicon." The chip is not picky — it can simultaneously handle inference for OpenAI, Microsoft's own MAI, and open-source models, and Microsoft released data showing that in Teams scenarios, its token throughput per watt is up to 1.7 times higher than leading GPUs, meaning it squeezes more intelligence out of less electricity.
The real headliner, the next-generation Maia300, will ship in 2027, targeting compute-hungry workloads such as long inputs, fast chat, and agent tasks. Microsoft claims its token output per megawatt is 40% higher than leading GPU systems; a rack packed with 96 chips can produce up to 7 times the tokens of a Maia200 rack. In terms of hardware specs, a single Maia300 packages 8 HBM4 memory stacks totaling 288GB with 18TB/s of bandwidth and a rated power draw of 950 watts — one chip carrying the entire appetite that used to be spread across a whole rack. A more detailed technical report has been released, and more details will be revealed at the Ignite 2026 conference in November.
When a cloud giant rolls out its own chips at gigawatt scale and explicitly declares them its primary inference system, what it clearly wants to escape is not just dependence on external GPUs, but rather — after welding models, chips, and cloud services into a self-reliant chain — a bill that no one else can calculate or hold hostage.