Microsoft bets gigawatts on its own AI chip: Azure Maia scales up from 2027

Microsoft will scale its in-house Azure Maia accelerators to multi-gigawatt deployments starting in 2027, the company announced. It expects the new Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems.

What is new

Scott Guthrie, who oversees Azure, wrote that Maia is meant to become the main inference platform for Microsoft’s own products and services. The second generation, Maia 200, is already handling production workloads, including Microsoft Copilot, and supports the MAI team as it builds its next models. GitHub Copilot and Microsoft Foundry are due to follow by the end of this year. According to Microsoft, the chip serves OpenAI models, the company’s own MAI models and several leading open-source model families.

Maia 200 is being rolled out in the United States, Europe and Australia. Microsoft says it delivers up to 1.7x more tokens per watt than leading GPU infrastructure when running the AI features built into Teams, Copilot and Cowork.

Maia 300 is the third generation. Microsoft describes it as an inference-first design aimed at long-context reasoning, responsive conversational use and sustained agentic work that spans many steps. Production deployments begin in 2027. The main figures from Microsoft’s post:

• Eight HBM4 memory stacks per accelerator, giving 18.0 TB/s of peak bandwidth within a 950-watt TDP.
• A multi-chiplet design that separates compute, memory-controller and I/O functions, with the compute chiplets built on TSMC’s N2P process.
• Support for low-precision formats including an enhanced MXFP4, which Microsoft says cuts memory and bandwidth needs for both computation and KV-cache storage.
• A software-defined dataflow architecture and AI-assisted kernel optimization aimed at reducing data movement, plus on-die SRAM used as a software-managed scratchpad.
• A custom AI transport over Ethernet-based infrastructure, meant to scale from a single rack up to a whole datacenter.
• A rack of 96 accelerators drawing 167 kilowatts, with 27.6 TB of total HBM and 1.7 PB/s of combined memory bandwidth.

In that full 96-accelerator setup, Microsoft says Maia 300 reaches up to 7x the rack-scale token throughput of Maia 200.

Deep-dive details

A separate architecture post on the Azure blog adds more. Each accelerator carries 288 GB of HBM4. Against Maia 200, Microsoft lists 36% more compute FLOPs, 2.5x the memory bandwidth, 33% more capacity and 14% higher scale-up bandwidth, along with more than 2.5x performance per watt.

The rack holds 24 four-accelerator trays and doubles Maia 200’s accelerator density. According to the post, liquid cooling handles more than 80% of the thermal load. Racks can be linked through optical second-tier switches into a superpod of up to 128 racks, or 12,288 accelerators. Microsoft describes three latency domains: a direct pair, accelerators within a pod, and accelerators across pods.

Why it matters

Maia 300 is the third act of a program that has not run smoothly. In 2025, DatacenterDynamics reported, citing The Information, that Microsoft had pushed mass production of the chip then codenamed Braga, expected to ship as Maia 200, back by at least six months to 2026. The reported causes were unplanned design changes, staffing shortages and turnover. The first-generation Maia 100, according to the same reporting, was used mostly for internal testing rather than to power AI services. Microsoft now says Maia 200 is running Copilot in production, so the second generation appears to have arrived, although the delay reports and the announcement do not say how the final schedule compared with the original plan.

The scale-up is also larger than the announcement suggests. The Next Web reported in August, citing The Information, that Microsoft was in talks with TSMC for capacity to make more than 300,000 Maia 300 chips for delivery in 2027, and quoted a company manager describing a goal of “gigawatts’ worth” of Maia chips. Those were talks, not a confirmed order, and Microsoft has not published a unit target.

Earlier reports had suggested Microsoft would not match Nvidia on performance per watt until this third generation, as the same delay coverage noted, so the 40% claim is the first real test of that expectation. Microsoft still buys Nvidia hardware, and its stance toward Nvidia’s Vera Rubin generation will show whether Maia is a supplement or a replacement. The signals to follow are whether the 2027 volumes materialize on schedule, whether GitHub Copilot and Foundry move onto Maia 200 by the end of 2026, and what Microsoft discloses at Ignite next month.

What’s next

Maia 300 production deployments are due to begin in 2027, with GitHub Copilot and Foundry expected on Maia 200 before the end of 2026. Microsoft said it will share more about its strategy at Ignite 2026, held in San Francisco and online, next month. A technical paper on the Maia 300 architecture is also available from the company.