Microsoft Maia 300: Azure’s Own AI Chip Becomes Microsoft’s Main Inference System From 2027

Microsoft Maia 300, the third generation of the company’s in-house artificial intelligence (AI) chip, will enter production use in 2027, when Microsoft starts expanding its Azure Maia chips to deployments measured in gigawatts and makes them the main system for running AI inside its own products and services.

That work, known as inference, means running trained models to answer requests, and Microsoft expects the new chip to generate 40% more tokens for every megawatt of power than leading systems built on graphics processing units (GPUs). Chief executive Satya Nadella described the move on social media as the company’s biggest bet yet on its own silicon.

Efficiency measured per megawatt

Tokens are the small units of data, mostly fragments of text, that an AI model reads and writes, so tokens per megawatt measures how much output a system produces for each unit of power it draws. Microsoft says a full Maia 300 rack of 96 accelerators delivers up to seven times the token throughput of a rack built on the current Maia 200. Power has become a central limit on AI expansion, and Microsoft is also building a roughly 2 gigawatt Pecos campus in Texas with its own power plant.

Inside Maia 300

Azure chief Scott Guthrie set out the specifications in Microsoft’s announcement. Each chip carries eight stacks of high-bandwidth memory (HBM4) with 18.0 terabytes per second of peak memory bandwidth, within a thermal design power of 950 watts. The compute chiplets are made on the N2P process of Taiwan Semiconductor Manufacturing Company (TSMC), and a multi-chiplet design separates compute, memory control and input and output functions.

A Maia 300 rack combines 96 accelerators within 167 kilowatts, with 27.6 terabytes of high-bandwidth memory and 1.7 petabytes per second of combined memory bandwidth. Microsoft says the chip is designed for long-context reasoning, conversational services that need fast replies and multi-step agentic tasks, and that it supports enhanced MXFP4, a four-bit microscaling floating-point format that cuts memory and bandwidth needs. Networking between accelerators runs on a custom AI transport over Ethernet-based infrastructure.

Maia 200 in production today

Maia 200 already handles live Microsoft work, Microsoft Copilot among it, and helps the company’s Microsoft AI (MAI) team build its next models. It runs models from OpenAI, from the MAI team and from the main open-source families, and GitHub Copilot and Microsoft Foundry are scheduled to move onto it before the year ends. Microsoft says the chip is being rolled out in the United States, Europe and Australia, and that it produces up to 1.7 times as many tokens for each watt as leading GPU infrastructure when running the AI features built into Teams, Copilot and Cowork.

Where Nvidia stands

Microsoft’s comparison point, leading GPU-based systems, is a market Nvidia leads. Microsoft frames Maia as one part of a mixed platform, arguing that no single model or chip will define what comes next. Nvidia has been using its own balance sheet to support demand for its systems, including its Ohio backstop for an AI campus serving OpenAI.

What it means for the cost of AI

Microsoft says the density of Maia 300 racks will let it handle more AI conversations at once from each rack and make its AI services cheaper to expand. The cost of serving each request is part of what businesses pay when they buy AI through the cloud, from retailers to banks and financial technology companies.

Microsoft says it will share more about its strategy at its Ignite 2026 conference, held in San Francisco and online next month, and has published a technical paper with further architectural details of Maia 300.

Editor’s note

Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.