Microsoft plans to make its homegrown Maia chips the primary inference system for its products and services as the company moves to reduce the cost and power demands of running AI at massive scale.
The company said that it will begin multi-gigawatt deployments of its Azure Maia family in 2027, when its third-generation Maia 300 accelerator is scheduled to enter production. Microsoft said the chip is designed specifically for inference – the computing work required when AI models respond to users – including long-context reasoning, conversational AI and multi-step agentic tasks.
Reuters reported that Microsoft could deploy more than 1 million Maia 300 chips in 2027 for its internal AI services, part of an effort to improve efficiency and reduce its reliance on outside chip suppliers including Nvidia.
Microsoft said Maia 300 will deliver 40% more tokens per megawatt than leading GPU-based systems, although it did not identify the competing systems. A rack containing 96 accelerators will consume about 167 kilowatts and provide up to seven times the rack-scale token throughput of the previous-generation Maia 200, according to the company.
Maia 200 is already running production workloads including Microsoft Copilot and is being deployed across the U.S., Europe and Australia. Microsoft said GitHub Copilot and Microsoft Foundry will follow by the end of the year.
The shift toward proprietary chips gives Microsoft greater control over the cost and energy efficiency of running AI services as inference demand grows. It also follows similar efforts by hyperscale cloud rivals Google and Amazon, which have developed their own AI accelerators while continuing to use chips from Nvidia and other suppliers.