Microsoft Scales Maia to Multi-Gigawatt and Puts Own Silicon First for Inference

Compute / Network
Microsoft said it will scale the Azure Maia family to multi-gigawatt deployments starting in 2027 and make Maia the primary inference system for its own products and services. Maia 200 is already in production. Maia 300 is expected to deliver about 40 percent more tokens per megawatt than leading GPU-based systems.
By Shashi Bellamkonda · October 12, 2026
2027
multi-gigawatt Maia deployments begin
40%
more tokens per megawatt vs leading GPUs (Microsoft)
7x
rack-scale token throughput vs Maia 200
96
accelerators per 167 kW rack
Owning the silicon is as strategic as it gets. Without the silicon and the hardware there is no software.

Microsoft said on October 12 that it will scale the Azure Maia family to multi-gigawatt deployments starting next calendar year and make Maia the primary inference system for its own products and services. Scott Guthrie, executive vice president of Cloud and AI, posted the plan on the official Microsoft blog. Satya Nadella called it the company's biggest bet yet on its own silicon.

Maia 200 is already in production and deploying across the United States, Europe, and Australia. It is serving real workloads, including Microsoft Copilot, and Microsoft says it delivers up to 1.7 times more tokens per watt for built-in artificial intelligence features in Teams, Copilot, and Cowork versus leading graphics processing unit infrastructure.

Maia 300, the third generation, begins production deployments in 2027. Microsoft expects it to deliver about 40 percent more tokens per megawatt than leading graphics processing unit-based systems. In a full 96-accelerator configuration it delivers up to seven times the rack-scale token throughput of Maia 200. The design targets long-context reasoning, latency-sensitive conversational experiences, and multi-step agentic workloads.

A Maia 300 rack packs 96 accelerators inside a 167-kilowatt power envelope. That is double the accelerator density of Maia 200. The rack holds 27.6 terabytes of high-bandwidth memory capacity and 1.7 petabytes per second of aggregate bandwidth. Each accelerator has eight high-bandwidth memory 4 stacks for 288 gigabytes of capacity and roughly 18 terabytes per second of bandwidth, with a 950-watt thermal design power. Compute dies use TSMC's N2P process in a chiplet layout that separates memory-controller and input-output functions. Scale-up networking supports superpods up to 12,288 accelerators.

Microsoft is still running a mixed fleet. The company said it will continue to bring the latest innovations from across the ecosystem into its data centers while making Maia the primary path for its own products and services. Those products include Copilot, Teams, GitHub Copilot, Microsoft Foundry, and the MAI model line.

Owning the silicon is as strategic as it gets. Without the silicon and the hardware there is no software. Microsoft is moving in the right direction by putting its own accelerator at the center of the inference path it uses for the products customers already run. The efficiency numbers matter because power and rack space set the cost of serving more Copilot sessions and more agent steps from the same data center footprint.

Ask your Microsoft team when Maia capacity will be available for the specific Copilot and agent workloads your organization runs, and whether the efficiency gain is visible in the rate card or only inside Microsoft's own services. The public plan is clear. The timing for any given customer workload is the open question.

Related reading: Lumentum Says AI Optical Parts Are Sold Out Through Early 2029 and The Chip Stack Has Never Been Wider, or Narrower.

For the chief information officer: confirm with your Microsoft account team which inference workloads will move to Maia first and what the measured tokens-per-watt gain means for your own usage charges. Then ask whether your current reserved capacity or private link arrangements need any change when the primary path shifts.

Disclaimer: This blog reflects my personal views only. Content does not represent the views of my employer, Info-Tech Research Group. AI tools may have been used for brevity, structure, or research support. Please independently verify any information before relying on it.