Microsoft has unveiled Maia 300, its third-generation AI accelerator, and plans multi-gigawatt deployments of its Maia systems starting in 2027. The company intends to make the chips its primary platform for running AI models behind Microsoft products and services.
Maia 200 is already serving production workloads, including Microsoft Copilot, and supporting the company’s MAI model-development team. Microsoft says GitHub Copilot and Microsoft Foundry will follow by the end of 2026. The existing systems run Microsoft’s models, OpenAI models and leading open-source families. Deployments span the United States, Europe and Australia.
Microsoft expects Maia 300 to generate 40 percent more tokens per megawatt than leading GPU-based systems. It did not identify the competing systems in that comparison. The figure is a forecast for the new platform, whose production deployments begin in 2027.
A full Maia 300 rack holds 96 accelerators within a 167-kilowatt power envelope. Microsoft’s technical account gives each accelerator 288 gigabytes of HBM4 memory and 18 terabytes per second of memory bandwidth. The company claims up to seven times Maia 200’s rack-level token throughput in that full configuration, which doubles the number of accelerators per rack.
The new design divides compute, memory-controller and input-output functions among separate chiplets. Its compute chiplets use TSMC’s N2P manufacturing process. Software controls where data is stored and moved, placing it near the part of the chip that needs it.
Microsoft also describes an Ethernet-based network connecting up to 128 racks, or 12,288 accelerators, in a single scale-up group. Within that group, workers processing a model’s input and workers generating its output can be assigned capacity separately. The system transfers cached model state directly between them.
The rack combines liquid and air cooling. Microsoft says liquid cooling removes more than 80 percent of its thermal load, including heat from accelerators and switching hardware. The design can connect to facility water or use a heat exchanger in an air-cooled data center.
Microsoft also plans to deploy other vendors’ hardware. Its Azure deployment agreement for AMD’s Helios systems covers the second half of 2026. Maia’s expansion is part of a capital program that Microsoft says will exceed $175 billion during the calendar year.
Electricity procurement is also expanding across the industry. Google’s 3,590-megawatt Constellation agreement includes 890 MW of additional nuclear capacity from reactor upgrades, with the first increase expected by 2028.
Microsoft plans to provide further details on its AI infrastructure strategy at Ignite, held in San Francisco and online from November 17 to 20.
Sources: Microsoft, Microsoft Azure, Reuters
–
By the Control Plane Editorial Team