Microsoft Plans a Million Maia 300 Chips to Run Its Own AI
Microsoft says it may deploy more than a million Maia 300 chips in 2027 to run its own AI services. The chiplet design claims 40 percent more useful computing per megawatt than leading GPU systems.
Key points
- SAN FRANCISCO: Microsoft announced the Maia 300 on Monday, October 12, the third generation of its own AI chip, and said it plans to deploy potentially more than a million of them next year.
- Scale is the story.
- The chip targets inference, the work of running a trained model to answer a question, write code or handle a business task.
SAN FRANCISCO: Microsoft announced the Maia 300 on Monday, October 12, the third generation of its own AI chip, and said it plans to deploy potentially more than a million of them next year. The chips will run Microsoft's internal AI services, the work behind products such as Microsoft 365 Copilot and GitHub Copilot, and OpenAI models delivered on Azure. Microsoft unveiled the chip at the OCP Global Summit in San Jose.
Scale is the story. A deployment above one million chips moves Maia from an experiment to core infrastructure. Microsoft expects capital spending to top 175 billion dollars this calendar year, and a large share of that money has gone to Nvidia. Maia 300 is built to keep more of that spending in house.
Built for inference, built from pieces
The chip targets inference, the work of running a trained model to answer a question, write code or handle a business task. Nvidia still leads in training, the heavier job of building models from scratch. Microsoft designed Maia 300 for the running side, and says it becomes the primary inference system for its products and services from 2027.
The design differs from Nvidia's approach. Nvidia builds one large processor ringed by high bandwidth memory. Microsoft built Maia 300 from chiplets, smaller specialized pieces bonded together. Key compute chiplets use TSMC's 2 nanometer process. Cheaper 3 nanometer chiplets handle moving data through the system.
At rack level, the trade shows in the numbers. A Maia 300 rack holds 96 chips and draws about 167 kilowatts. Nvidia's GB300 rack holds 72 graphics processors and draws about 142 kilowatts. More chips, more power. Microsoft says the return is about 40 percent more useful computing per megawatt than leading GPU based systems. The company did not name the systems it compared against, so treat that figure as a claim to be tested, not a result.
A line of chips
Maia 100 arrived in 2023. Maia 200 was announced in January 2026 and now runs production work, including Copilot products and OpenAI models on Azure. Executive vice president Scott Guthrie said Maia 300 uses an inference first design tuned for long context and memory heavy work, and that a 96 chip rack delivers seven times the rack level token throughput of Maia 200. The chips are reserved for Microsoft's own services rather than rented out as outside capacity.
Why it matters
Every big cloud company is now designing its own silicon, and each design chips at Nvidia's hold on AI computing. Custom chips also change the cost math behind AI products. If Microsoft can run Copilot features on its own hardware at a lower cost per answer, the pressure on AI prices across the industry rises. The same week, chip markets fell hard on news of a Chinese memory advance. The hardware race is moving on several fronts at once.
What happens next
The deployment lands in 2027, in what Microsoft calls multi gigawatt installations. Watch for the first independent readings on the 40 percent efficiency claim once outside users can test the systems, and for Nvidia's response on inference pricing.