A chip for 2027 and a commitment to scale
On Monday, October 12, 2026, Scott Guthrie, Microsoft's executive vice president of Cloud and AI, announced Maia 300, the third generation of Microsoft's in-house AI accelerator, and said Microsoft would scale the Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services. Maia 300 is due to begin production deployments in 2027, so nothing was shipping on the day. [1][2]
The same post gave an update on the previous generation, Maia 200. Microsoft said it was serving production workloads including Microsoft Copilot and supporting its MAI model team, with GitHub Copilot and Microsoft Foundry to follow by the end of the year, and that it runs OpenAI models, Microsoft's MAI models and open-source models. Microsoft claimed up to 1.7 times more tokens per watt than leading GPU infrastructure for AI features in Teams, Copilot and Cowork. [1]
What Microsoft published about the chip
Microsoft describes Maia 300 as inference-first, aimed at long-context, memory-heavy and decode-heavy work such as multi-step agents. It is a multi-chiplet design that splits compute from memory controllers and I/O, with compute chiplets on TSMC's N2P process. Eight HBM4 stacks give 18.0 TB/s of peak memory bandwidth within a 950-watt thermal design power. It supports low-precision formats including an enhanced MXFP4, and uses on-die SRAM as a software-managed scratchpad. [1]
A Maia 300 rack holds 96 accelerators within 167 kilowatts, with 27.6 TB of HBM and 1.7 PB/s of aggregate memory bandwidth, linked by an Ethernet-based custom transport. The Next Platform, writing the next day, reported the codename Clea, tied the disclosure to the start of the Open Compute Project Summit, and put the HBM4 capacity at 288 GB per package; its estimates of die area and SRAM are the author's own guesses. [1][3]
The performance claims
Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems, and a full 96-accelerator rack up to seven times the rack-scale token throughput of Maia 200. The post does not name the GPU systems in the comparison or publish the workloads behind either figure. Redmond Magazine reported both as Microsoft's expectations. [1][2]
The numbers are a forecast for hardware a year from deployment, measured against an unnamed baseline, so they describe Microsoft's target more than a result. What is concrete is the commitment: Microsoft put a date and a scale on moving its own inference off merchant GPUs, while its own models and OpenAI's run on the same Maia fleet. [1]
Sources
- Microsoft to expand Azure Maia system to multi-gigawatt deployments
Microsoft · Oct 12, 2026
- Microsoft Bets on Its Own Chips with New Maia 300 for AI
Redmond Magazine · Oct 12, 2026
- Microsoft’s “Clea” Maia 300 Inference XPU Is A Bandwidth Beast
The Next Platform · Oct 13, 2026