Microsoft targets multi-gigawatt scale for Maia 300
Microsoft says its Maia 300 inference system will reach multi-gigawatt deployments in 2027, challenging Nvidia’s role in its largest AI workloads.
Microsoft has announced Maia 300, the next generation of its in-house AI accelerator, and said it plans to scale the Azure Maia family to multi-gigawatt deployments beginning in 2027. The announcement turns a custom-chip program into a central part of Microsoft’s infrastructure strategy rather than a limited experiment.
Designed around inference
Microsoft describes Maia 300 as an inference-first system for long-context reasoning, latency-sensitive conversations and sustained agentic workloads. A full rack is planned to combine 96 accelerators within a 167-kilowatt power envelope, with 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth.
The company claims Maia 300 will deliver 40% more tokens per megawatt than leading GPU-based systems and up to seven times the rack-scale token throughput of Maia 200. Maia 200 is already serving Microsoft production workloads, including Copilot, while later deployments are expected to support GitHub Copilot, Microsoft Foundry and Dragon Copilot.
Microsoft’s design emphasizes data movement and memory efficiency as much as raw compute. The chip uses HBM4, a multi-chiplet architecture and lower-precision formats intended to reduce bandwidth and power requirements for large models and long context windows. Microsoft says Maia 300 will begin production deployments in 2027.
What changes for the market
The announcement does not mean Microsoft is abandoning Nvidia. Azure will continue to offer a broad mixture of hardware, and the Maia system is being built for Microsoft’s own workloads. But the scale is strategically meaningful: multi-gigawatt deployment gives Microsoft a way to optimize cost, power and software behavior around its own services instead of accepting every constraint of general-purpose GPU infrastructure.
The open question is execution. Maia 300 is still a future production system, and Microsoft’s performance-per-watt comparisons are company claims rather than independent benchmarks. Even so, the plan signals that custom inference silicon is moving from a defensive procurement tactic to a core competitive layer for hyperscalers running agents at global volume.