LuminaBench
← Lumina Journal
AI News2 min read

Microsoft plans multi-gigawatt Maia rollout from 2027

Maia 200 is already running Copilot workloads. Maia 300 follows next year, with 96 accelerators per rack and a focus on getting more AI output from each megawatt.

Lumina
Maia 300 is due to enter production deployments in 2027.Source: Microsoft’s announcementCredit: Microsoft
Microsoft’s Azure Maia 300 accelerator, photographed against a dark blue background with blue and purple lighting.
Open original image ↗

Microsoft plans to make Maia its primary system for running AI models across its products and services. The company announced on 12 October that deployments of its in-house accelerator family will reach multi-gigawatt scale starting in 2027.

Some of that shift is already happening. Microsoft says Maia 200 is serving production workloads for Copilot and supporting its MAI model team. GitHub Copilot and Microsoft Foundry are due to follow by the end of 2026.

Maia 300 doubles the accelerators per rack

The next generation fits 96 accelerators into a 167-kilowatt rack, twice the accelerator count of Maia 200. Each Maia 300 has 288 GB of HBM4 memory and 18 TB/s of memory bandwidth, according to Microsoft’s technical breakdown.

Microsoft has split the chip into separate compute, memory-controller and I/O dies, with compute built on TSMC’s N2P process. The design lets software place data near the compute that needs it, cutting the distance data travels. That matters for the long-context and multi-step agent workloads Microsoft is targeting.

Microsoft’s specifications for Maia 300. The 950 W figure is the chip’s TDP; the complete 96-accelerator rack has a 167 kW power envelope.Source: Microsoft’s announcementCredit: Microsoft
Microsoft’s Maia 300 specifications: N2P compute chiplets, 13.8 FP4 PFLOPS, 6.9 FP8 PFLOPS, 950 W chip TDP, eight HBM4 stacks, 288 GB HBM, 18 TB/s memory bandwidth and 96 accelerators per rack. A labelled diagram identifies compute, I/O and memory dies.
Open original image ↗

The networking extends beyond one rack. Microsoft describes a shared Ethernet-based network connecting up to 12,288 accelerators across 128 racks. Software can divide a job across racks and move work to spare accelerators elsewhere in the system if a tray fails.

Microsoft’s performance claims

Microsoft expects Maia 300 to produce 40% more tokens per megawatt than leading GPU-based systems. The announcement does not name the GPU configuration used for that comparison.

Separately, it claims up to seven times Maia 200’s token throughput per rack in the full 96-accelerator configuration. That is a comparison of complete racks, with twice as many accelerators in the new one. Both figures come from Microsoft; production deployments of Maia 300 are scheduled to begin in 2027.

Sources & references

  1. Microsoft: Maia expansion and the 2027 rollout ↗

    12 October 2026. Announcement by Scott Guthrie, including deployment plans and Microsoft’s performance claims. Hero image and specifications graphic: Microsoft.

  2. Microsoft’s Maia 300 architecture breakdown ↗

    12 October 2026. Technical account by Rani Borkar and Saurabh Dighe. Specifications and comparisons are Microsoft’s own figures.

What to read next