Microsoft announces Maia 300 AI chip as primary inference system for products

Oct 12, 2026

Microsoft announced the Maia 300 on October 12, 2026, marking its third-generation custom AI chip designed to serve as the primary inference engine for its services. Starting deployment in 2027 at multi-gigawatt scales, Microsoft aims to deploy over 1 million chips in 96-accelerator racks. Built using TSMC's 2nm chiplet process, the Maia 300 offers 40% higher efficiency per megawatt than leading GPUs, reducing dependency on Nvidia while continuing model training partnerships.

Maia 300 announcement and optimization

▪Microsoft announced the Maia 300 custom artificial intelligence accelerator on October 12, 2026, positioning it as the primary AI inference system for its products and services
▪The Maia 300 chip is optimized for long-context reasoning, latency-sensitive conversational experiences, and sustained multi-step agentic tasks

Maia 300 performance and power efficiency

▪A full 96-accelerator rack configuration of Maia 300 delivers up to seven times the rack-scale token throughput of a Maia 200 rack
▪Microsoft claims the Maia 300 chip delivers 40% more tokens per megawatt compared to leading graphics processing unit systems

Technical specifications and hardware architecture

▪Microsoft expects to pack 96 Maia 300 chips into each data center rack, using approximately 167 kilowatts of power per rack
▪Each Maia 300 chip includes eight HBM4 memory stacks totaling 288 GB of memory with 18 TB/s of bandwidth and is rated at 950 watts
▪The Maia 300 architecture utilizes TSMC 2-nanometer N2P technology for its central compute chiplets and 3-nanometer technology for memory controller and I/O functions

Deployment plans and scaling Strategy

▪Microsoft Executive Vice President Scott Guthrie stated Microsoft will scale the Azure Maia family to multi-gigawatt deployments starting in 2027
▪Evercore ISI analysts estimated that Maia 300 chips will represent approximately 10% of Microsoft's AI chip spending by the first quarter of 2027
▪Reuters reported Microsoft plans to deploy potentially more than 1 million Maia 300 chips starting in 2027

Access to Maia chips

▪Unlike cloud computing accelerators from Google and Amazon, Microsoft does not rent its Maia silicon directly to external cloud customers
▪Moor Insights and Strategy CEO Patrick Moorhead stated that no organization outside of Microsoft has tested the Maia accelerators yet

Maia 200 performance and deployment

▪Microsoft's custom Maia 200 chip already runs workloads including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and OpenAI models on Azure cloud
▪Microsoft CEO Satya Nadella stated in July 2026 that the Maia 200 chip delivers 30 percent better performance per dollar than Microsoft's existing fleet hardware

Partnerships with third-party chipmakers

▪Microsoft general manager Alistair Speirs said Microsoft will continue to partner with Nvidia and AMD for training AI models, maintaining Nvidia GPUs across its existing data centers
▪Microsoft uses its own central processing units alongside processors from Intel and AMD for general-purpose computing workloads

Debatable claims

▪Multi-gigawatt AI infrastructure deployments undermine power grid sustainability
▪Unverified internal testing renders Microsoft's performance claims for Maia silicon unreliable
▪In-house custom silicon reduces cloud providers' dependence on Nvidia