Microsoft's new in-house chip will start replacing Nvidia for most AI tasks
The Maia 300 chip is set to become Microsoft's main engine for running AI services starting in 2027, cutting its reliance on Nvidia hardware.
What happened: Microsoft unveiled its third-generation Maia AI chip, the Maia 300, and said it plans to make it the main chip handling everyday AI requests across its products, rather than relying mostly on Nvidia's processors as it has for years. Executive Vice President Scott Guthrie said Microsoft will scale Maia to 'multi-gigawatt deployments' starting in 2027, calling it 'the primary AI inference system for Microsoft products and services.' Production begins in 2027, though the company has not given a firm date for when Maia 300 takes over most of the workload.
Why it matters: Nvidia's GPUs have been the default choice for running AI models industry-wide, and demand for them has outstripped supply for years. Microsoft, along with Amazon, Google, AMD, and Cerebras, has been building its own chips partly to get more capacity and partly to cut costs. Finance chief Amy Hood said using Microsoft's own silicon has already improved profit margins over the past decade. For Microsoft specifically, owning more of its AI hardware also supports its push to catch up with OpenAI, Anthropic, and Google in building competitive AI models of its own.
How it works, plainly: Maia 300 is built for 'inference' -- running already-trained AI models to answer real user requests -- rather than the separate, heavier job of training new models from scratch. It uses a new chiplet design, splitting compute, memory, and input-output functions onto separate pieces of silicon built on TSMC's N2P manufacturing process, which Microsoft says boosts the chip area devoted purely to computing. Microsoft says a full rack of 96 Maia 300 chips (double the Maia 200's count) can process up to seven times more AI 'tokens' and does so using 40% less power per unit of output than leading GPU systems, which should lower data-center operating costs.
The rollout: Maia 200, launched in January, already runs production traffic for Microsoft 365 Copilot, GitHub Copilot, Microsoft Foundry, and Microsoft's in-house models. Evercore ISI analysts estimate Maia chips could make up about 10% of Microsoft's AI chip spending by early 2027. Nvidia remains central for now: Microsoft still partners with Nvidia and AMD for training new models, and a new Surface laptop announced alongside Maia uses an Nvidia processor. Unlike Google's or Amazon's custom chips, Maia cannot be rented out to other companies -- it's for Microsoft's internal use only.