By Rebecca Stone

Microsoft claims its new Maia 300 AI chip will deliver 40% more tokens per megawatt than leading GPU-based systems. The company unveiled the third-generation processor on Monday.

Scott Guthrie, Microsoft's executive vice president for cloud and AI, announced the figure in a blog post dated October 12. The chip targets the latest reasoning AI models and agent workloads. Microsoft says the design should lower operating costs for inference at scale.

"We expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems," Guthrie said.

What is the deployment timeline?

Microsoft plans to deploy the Maia 300 after the turn of the year. The schedule positions the chip for production use in early 2027 infrastructure cycles.

The predecessor, the Maia 200, launched in January. It now supports Microsoft Copilot workloads. GitHub Copilot and Microsoft Foundry will follow by the end of 2026, the company said.

The rollout marks a shift in Microsoft's silicon strategy, moving inference (the phase where a trained model generates output) from merchant GPUs to first-party hardware.

How far will Maia scale?

Guthrie committed to gigawatt-scale capacity. "We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services," he said.

That timetable gives enterprise buyers roughly a year of overlap. Maia 200 workloads expand through 2026 while Maia 300 capacity ramps behind them.

What does this mean for Nvidia and AMD?

The Verge reported that the Maia 300 will reduce Microsoft's dependence on AI chips from Nvidia and AMD. Microsoft is shifting more workloads onto its own silicon.

For CIOs, the implications are indirect but concrete. Microsoft controls the full inference stack — chip, firmware and cloud scheduler — for the systems that will run Copilot, GitHub Copilot and Foundry. Vendor lock-in deepens as those services move to first-party hardware.

Key claims at a glance

• Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems, per Microsoft.
• Maia 200 launched in January; GitHub Copilot and Microsoft Foundry support arrives by end of 2026.
• Multi-gigawatt Azure Maia deployments begin in 2027.
• Maia becomes the primary inference system for Microsoft products and services.

Microsoft did not disclose pricing, power draw or per-chip performance figures beyond the efficiency claim. The company also did not specify which GPU systems served as the comparison baseline.

Enterprise customers running Azure AI services will see the change through pricing and capacity rather than procurement. Microsoft absorbs the hardware layer; the efficiency gain appears in operating cost per token, the unit that measures model output volume.

The move follows a broader industry pattern. Cloud providers are designing inference silicon in-house to control cost at scale. Microsoft's commitment to multi-gigawatt capacity signals volume production, not a limited trial.

Guthrie's statements frame Maia as infrastructure, not a product SKU. The 2027 milestone is the one to watch.