Microsoft has announced Maia 300, the third generation of its in-house artificial intelligence (AI) accelerator, and said it will scale its Azure Maia systems to multi-gigawatt deployments beginning in 2027. The company also said the previous chip, Maia 200, is now in production and serving Microsoft Copilot.

The plans were set out on October 12 in a post on the Official Microsoft Blog by Scott Guthrie, executive vice president of Cloud + AI, and in a technical deep dive on the Azure blog by Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, corporate vice president of systems planning and architecture. Microsoft said Maia 300 will begin production deployments in 2027 and described the Maia family as its main inference system, the hardware that runs trained models to answer requests.

What is in the chip

According to Microsoft, Maia 300 uses a multi-chiplet design. Compute is placed on chiplets built on TSMC’s N2P process, while memory-controller and input/output functions sit on separate dies. Each accelerator carries eight HBM4 high-bandwidth memory stacks, providing 288 GB of capacity and 18 TB/s of peak bandwidth, within a 950-watt thermal design power (TDP). The chip uses on-die SRAM as a software-managed scratchpad and supports what Microsoft calls enhanced MXFP4 and sub-3-bit storage formats for low-precision inference.

A Maia 300 rack holds 96 accelerators within a 167-kilowatt envelope, with 27.6 TB of HBM capacity and 1.7 PB/s of aggregate memory bandwidth. Scale-up networking runs over a custom AI transport on Ethernet, with electrical links inside a rack and optical links between racks, and Microsoft said this two-tier network can build a superpod of up to 12,288 accelerators. Liquid cooling removes more than 80% of a rack’s thermal load, according to the Azure post, with air cooling handling memory, storage and other components.

The performance claims

Microsoft said it expects Maia 300 “to deliver 40% more tokens per megawatt compared to leading GPU-based systems.” It did not name the GPU systems used as the baseline. Against its own predecessor, the company claimed up to seven times the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration, 2.5 times the memory bandwidth, 33% more memory capacity and a 36% increase in compute.

For Maia 200, Microsoft said the chip is deploying across the United States, Europe and Australia and runs OpenAI models, Microsoft’s own MAI models and leading open-source model families. The company said it delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork, with GitHub Copilot and Microsoft Foundry due to follow by the end of the year.

Microsoft said Maia 300 is aimed at long-context reasoning, latency-sensitive conversational experiences and multi-step agentic tasks, naming GitHub coding experiences, Foundry model serving, MAI model inference and data generation, and Dragon Copilot as uses.

Why it matters

“Like other hyperscalers, Microsoft continues to rely heavily on Nvidia and AMD GPUs for AI workloads,” Data Center Knowledge noted in its coverage. The publication also pointed out that, unlike Amazon Web Services with its Trainium and Inferentia chips or Google Cloud with its TPUs, Microsoft has not made Maia available to customer workloads through Azure instances, and its blog post did not say whether that could change. Microsoft framed the announcement around output per unit of power, measured in tokens per watt and per megawatt, rather than raw compute.

Matt Kimball, vice president and principal analyst at Moor Insights & Strategy, told Data Center Knowledge that because Maia mainly serves Microsoft’s internal workloads, the competition is less a traditional chip war than a race to deliver services faster and more efficiently. “That’s good for the customer because they’re getting better quality of service, and it’s good for the cloud provider because they’re doing it at lower cost, which means they’re more profitable,” he said. Kimball described the progression this way: “Maia 100 was a proof of concept.”

The caveats

All of the performance figures are Microsoft’s own claims or projections. The 40% figure for Maia 300 is an expectation rather than a measured result, the company has not published independent benchmarks, and the Azure post gives no workload or methodology details for its comparisons. The 40% tokens-per-megawatt figure for Maia 300 and the 1.7 times tokens-per-watt figure for Maia 200 also use different baselines, so they cannot be compared directly. Microsoft did not give a total gigawatt figure for the planned rollout.

What to watch

Microsoft said it will share more about its strategy at Microsoft Ignite 2026 in San Francisco and online, which the Azure site lists for November 17 to 20. Production deployments of Maia 300 are scheduled to begin in 2027.

Corrections and tips: email the editor using the address on the Contact page.