Microsoft plans multi-gigawatt Maia deployments for 2027

Microsoft plans to begin scaling its Maia family to multi-gigawatt deployments in 2027 and make Maia the primary inference system for its own products and services. The company also plans to start production deployments of Maia 300 that year, according to its October 12, 2026 announcement.

Maia 200 is serving production workloads

Microsoft says Maia 200 is already serving production workloads, including Microsoft Copilot, across the United States, Europe and Australia. The company also names OpenAI models, Microsoft MAI models and leading open-source model families as workloads running on Maia 200. It said GitHub Copilot and Microsoft Foundry were planned to follow by the end of 2026.

That is the operating baseline for the next step: Maia 200 is in production, while Maia 300’s production rollout remains a 2027 plan.

Maia 300 scales from accelerator to superpod

Maia 300 is designed as an inference-first system for long-context reasoning, latency-sensitive conversations and multi-step agent tasks. Microsoft describes a multi-chiplet design that separates compute, memory-controller and I/O functions; its compute chiplets use TSMC’s N2P process.

Each accelerator has eight HBM4 stacks, with 288 GB of high-bandwidth memory (HBM) and peak bandwidth of 18 TB/s. Microsoft’s architecture deep dive describes how those accelerators fit into a larger system: a rack holds 96 accelerators across 24 four-accelerator trays, within a 167-kW power envelope.

The scale-up domain—the connected group of accelerators working together—can reach 12,288 accelerators across as many as 128 racks. Microsoft describes an Ethernet-based AI network with direct accelerator links, switched connections within racks and optical links between racks. Liquid cooling handles more than 80% of a rack’s thermal load, while air cooling serves memory, storage and other components.

Maia 200 and Maia 300: published specifications

Microsoft’s technical pages list these shared specifications for the two generations. The Maia 200 figures come from the company’s architecture article.

The capacity and bandwidth figures describe memory on each accelerator; the scale-up figures describe the largest connected accelerator group. Maia 300’s 96-accelerator rack and its 12,288-accelerator scale-up domain are therefore different levels of the system.

What Microsoft’s performance claims measure

For a full 96-accelerator Maia 300 rack, Microsoft claims up to seven times Maia 200’s rack-scale token throughput. Separately, the company claims Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems. Both are Microsoft’s published comparisons, but they use different metrics and baselines: one measures rack-scale throughput against Maia 200; the other compares tokens per unit of power with GPU-based systems.