Microsoft Aims for Multi-Gigawatt Capacity for Maia by 2027

Microsoft plans to scale its Maia chips to a multi-gigawatt level starting in 2027 and make it its primary inference system. Maia 200 is already powering Microsoft Copilot in several regions, while a Maia 300 is set to be announced for 2027 with claimed specifications and efficiency gains.

Announced Efficiency and Throughput Gains for Maia

According to Microsoft, Maia offers up to 1.7 times more tokens processed per watt than the leading GPUs used in Teams. For Maia 300, which is expected to launch in 2027, the company states that the number of tokens processed per megawatt would exceed that of leading GPU systems by 40%. A rack composed of 96 Maia 300 chips is expected to generate up to seven times the number of tokens produced by a rack of Maia 200. Each Maia 300 integrates eight HBM4 memory modules for a total of 288 GB, with a bandwidth of 18 TB/s and a reported power consumption of 950 watts.

What’s Already Working and What’s Coming by Year-End

The Maia 200 chip is already in use with Microsoft Copilot in the United States, Europe, and Australia. GitHub Copilot and Foundry are expected to adopt Maia by the end of the year. Maia supports OpenAI models, MAI, and open-source frameworks.

2027 Roadmap and Management's Stated Position

Microsoft intends to generalize the deployment of its Maia chips at a multi-gigawatt scale starting in 2027 and aims, according to Scott Guthrie, to make it the primary inference solution for its products. Maia 300, which will succeed Maia 200, is anticipated for 2027 with ambitions regarding long input processing, fast conversation management, and support for agentic tasks. Satya Nadella describes this strategy as “our biggest bet on our own silicon.” A technical report details the features of the Maia chips, and additional information is expected in November during the Ignite 2026 event.