Microsoft on Oct. 12, 2026 introduced Azure Maia 300, the third generation of its in-house AI accelerator, and said it will scale the Maia family to multi-gigawatt deployments starting in 2027. In a post on the Official Microsoft Blog, Scott Guthrie, executive vice president of Cloud + AI, said that expansion would be "making Maia the primary inference system for Microsoft products and services." The post gives no gigawatt total and names no data center sites for the new system.

For facility operators, the more consequential details sit in a companion architecture deep dive on the Azure blog by Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, corporate vice president for systems planning and architecture. It lays out a rack design built around a 167-kW envelope, liquid cooling for most of the heat load and 50-V power distribution.

The rack: 96 accelerators in 167 kW

According to the architecture post, a Maia 300 rack holds 96 accelerators spread across 24 trays of four accelerators each, double the accelerator density of Maia 200. The design drops separate host trays: each tray carries its own host processor, memory, storage and cooling, and six switch ASICs sit in three liquid-cooled switch trays.

Each accelerator pairs compute dies built on TSMC's N2P process with eight stacks of HBM4 memory, which the deep dive puts at 288 GB and 18 TB/s of bandwidth per chip. Guthrie's post gives each chip a thermal design power of 950 W and lists 27.6 TB of high-bandwidth memory and 1.7 PB/s of aggregate memory bandwidth per rack. Racks connect through a two-tier, Ethernet-based scale-up network that, per the deep dive, links up to 128 racks into a superpod of up to 12,288 accelerators.

Cooling: liquid for most of the heat, two facility options

The architecture post is specific about where the heat goes. "Liquid cooling removes more than 80% of the rack's thermal load from accelerators, voltage regulators, and switch ASICs," Borkar and Dighe wrote. Air cooling handles the memory, storage and other ancillary components.

Microsoft describes two ways to integrate the racks into a building. They can connect directly to facility water where it is available, or they can use a heat-exchanger unit for deployment in air-cooled data centers.

Power: 50-V busbar and dual feeds

On the electrical side, the deep dive lists a 50-V busbar, dual power feeds, high-capacity power shelves and an integrated rack automatic transfer switch, which Microsoft says keeps a rack available when one feed is offline for servicing. The 167-kW figure is the rack's power envelope, not a site-level demand figure.

Microsoft's performance claims

The efficiency numbers attached to the launch are Microsoft's own. The deep dive says Maia 300 delivers more than 2.5x the performance per watt of Maia 200 and up to 7x its rack-scale token throughput in the full 96-accelerator configuration. It also claims 2.5x the memory bandwidth, 33% more memory capacity, 36% more compute FLOPs and 14% more scale-up bandwidth than the previous generation.

Guthrie's post frames the comparison against GPUs in power terms, claiming 40% more tokens per megawatt than leading GPU-based systems. It makes a similar claim for the current generation, saying Maia 200 delivers up to 1.7x more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork.

Where Maia fits

The previous generation, Maia 200, is in production serving Microsoft Copilot and is deploying across the United States, Europe and Australia, according to Guthrie's post, which says GitHub Copilot and Microsoft Foundry are set to follow by the end of the year. The launch was organized around a Microsoft Source feature page that links the architecture post, an analyst-quotes document and video transcripts, and features Guthrie, Borkar, Dighe, Microsoft AI CEO Mustafa Suleyman and Technical Fellow Jim Kleewein.

What Happens Next

Microsoft says Maia 300 production deployments begin in 2027, and both blog posts say the company will share more at Microsoft Ignite 2026 in San Francisco and online in November. Microsoft has not published a total capacity target or the data center sites, utilities or regions that will host Maia 300 racks, so no site-level permits, interconnection filings or construction milestones tied to the system are yet on the public record.

Sources