Microsoft introduced its third-generation Maia AI chip, claiming higher inference efficiency and throughput, with production deployment planned to begin in 2027.
Microsoft announced Maia 300, its third-generation in-house AI chip, on October 12, 2026. The company is designing it to handle inference for Microsoft products and services, including workloads involving long-context reasoning, conversational systems sensitive to response time, and multi-step AI agents. Microsoft says production use is expected to begin in 2027.
The company projects Maia 300 will deliver 40 percent more tokens per megawatt than leading GPU-based systems. Microsoft executive vice president Scott Guthrie also said a full configuration of 96 accelerators can provide up to seven times the rack-scale token throughput of Maia 200. Those figures are Microsoft’s stated expectations, rather than independent comparisons.
Maia 300 follows Maia 200, which Microsoft announced in January. The earlier chip is already used for production AI workloads across Microsoft Copilot, GitHub Copilot, Microsoft Foundry and the company’s own AI models. Guthrie described Maia 200 as the best-performing AI infrastructure in Microsoft’s fleet by real-world service metrics; the new chip is being developed for a broader set of demanding inference tasks.
A key design change is a chiplet architecture built using TSMC’s N2P process node. Microsoft’s Azure hardware and infrastructure team says the design separates compute, memory-controller and input/output functions, allowing each part to use technology suited to its role. The company has not provided further deployment details in the announcement.
Microsoft says it plans to scale its Azure Maia family to multi-gigawatt deployments starting in 2027, with Maia 300 intended as the primary inference system for its products and services. That would expand the role of Microsoft-designed silicon in its AI infrastructure, though the announcement does not specify how much the company’s reliance on Nvidia or AMD hardware will change.