Full Breakdown
Microsoft Schedules Unveiling of Maia 300, Its Third-Generation AI Inference Chip
By Drooid · · How we work
Core Event
Microsoft will introduce the third-generation Maia 300 AI accelerator on October 12, 2026. The chip is slated to become the primary inference system for Azure and services such as Microsoft 365 Copilot, GitHub Copilot and Microsoft Foundry. Production deployments are planned for 2027, with a “multi-gigawatt” rollout across Microsoft data centres.
Background & Context
Microsoft’s AI stack has relied on Nvidia GPUs for training and inference. Rising demand and a shift toward diversified silicon have driven the company to build its own chips. The first in-house accelerator, Maia 100, debuted in 2023, followed by Maia 200 on January 26, 2026. Maia 200 already handles production workloads for Copilot, delivering up to 1.7 × more tokens per watt than leading GPU infrastructure. Maia 300 is an “inference-first” design that reduces data movement and improves power efficiency, lessening dependence on external GPU suppliers.
Timeline
| Date | Event |
|---|---|
| January 26, 2026 | Launch of Maia 200, built on TSMC’s 3-nm process. |
| October 12, 2026 | Planned announcement of Maia 300 and the multi-gigawatt deployment plan. |
| 2027 | Start of production-scale deployments of Maia 300 in Azure data centres. |
| November 20, 2026 | Further details expected at the Ignite conference in San Francisco. |
Data & Statistics
- Efficiency claim: Maia 300 will deliver 40 % more tokens per megawatt than leading GPU-based systems.
- Rack configuration: Each rack will house 96 accelerators, draw 167 kW, and provide 27.6 TB of HBM with 1.7 PB/s aggregate bandwidth.
- Throughput improvement: A full-rack setup is projected to achieve up to seven times the token throughput of a Maia 200 rack.
- Memory architecture: Eight stacks of HBM4 deliver a peak bandwidth of 18 TB/s within a 950-W TDP.
- Chiplet design: Compute chiplets use TSMC’s 2-nm N2P process; separate chiplets handle memory-controller and I/O functions.
- Target workloads: Long-context reasoning, latency-sensitive conversational AI, and multi-step agentic tasks.
Official Statements & Responses
- Scott Guthrie, executive vice president of Microsoft’s cloud group, called Maia 300 a “complete intelligent system” for emerging AI workloads and said the Azure Maia family will scale to multi-gigawatt deployments starting in 2027.
Verbatim Quotes
- “We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems,” — Scott Guthrie
- “We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary AI inference system for Microsoft products and services,” — Scott Guthrie
- “Maia 200 is the best performing AI infrastructure in our fleet, measured in real world metrics for AI services,” — Scott Guthrie
- “We need new architectures that give us better performance and better use of the power we have,” — Alistair Speirs, general manager
What’s Next
Microsoft indicated that additional details will be shared at the Ignite conference in San Francisco later in November. The company also plans to expand Maia 300 production with TSMC, targeting more than 300,000 chips for the 2027 rollout. No dates have been set for completing the multi-gigawatt expansion or for offering Maia 300 to external Azure customers.