Microsoft will scale its Azure Maia system to multi-gigawatt deployments starting in 2027. Maia 200 is now in production for Microsoft Copilot and other services. GitHub Copilot and Microsoft Foundry will follow by end of year. Maia 200 runs OpenAI, Microsoft MAI and open-source models. Maia 300 is designed for long-context reasoning and agentic tasks. Maia 300 delivers 40% more tokens per megawatt than leading GPU systems. A 96-accelerator Maia 300 rack delivers up to 7x the throughput of Maia 200. Maia 200 already deploys in the US, Europe and Australia. Maia 200 offers 1.7x more tokens per watt for Teams, Copilot and Cowork. Maia 300 production deployments begin in 2027.
What changed
Maia 200 was the primary inference system for limited Microsoft workloads.
What it unlocks
Running long-context reasoning, latency-sensitive chat and multi-step agent tasks at cloud scale.
- 40% more tokens per megawatt vs GPU systems
- 7x rack-scale token throughput vs Maia 200
- 1.7x tokens per watt for Teams, Copilot, Cowork
- 18.0 TB/s peak memory bandwidth per chip
What you need to act on it
- Azure cloud access
- Microsoft AI services usage
Sources