Chips
Microsoft's own chip for running its AI won't arrive until 2027
Maia 300 is the third generation of Microsoft's AI chip. The company expects to get 40% more text per megawatt than with leading graphics processors, but that is its own forecast for a chip that isn't working yet.
The essentials
4 according to the source · 3 open questions
- According to the sourceMicrosoft plans to start deploying Maia 300, the third generation of its own AI chip, in 2027.12
- According to the sourceMicrosoft expects Maia 300 to generate 40% more text per megawatt than leading systems based on graphics processors.12
- According to the sourceAccording to Microsoft, Maia 200, the previous generation, is already running in production and serving Microsoft Copilot.1
- According to the sourceMicrosoft wants the Maia family to be the main inference system for its products and services, meaning the one that runs models that have already been trained.12
Why it matters
Microsoft is stressing how much text it gets from each megawatt, more than raw power. If the forecasts hold, it could handle more requests with the same energy; it can be read as a step toward relying less on graphics processors from other companies. But these are Microsoft's own figures for a chip that is not yet in service.
The details
Microsoft has announced Maia 300, a chip it designed for its own data centers1. It is built for inference, the moment when a model that has already been trained responds to what it is asked. According to the company, it is aimed at reasoning with a lot of context, fast responses and agents that chain many steps together1. Each rack holds 96 chips and, Microsoft says, delivers up to 7 times the performance of a Maia 200 rack1.
The previous generation, Maia 200, is already in use. According to Microsoft, it serves Microsoft Copilot, and before the end of the year it will do the same for GitHub Copilot and Microsoft Foundry1. It runs models from OpenAI, Microsoft's own models and open-source ones, and it is being installed in the United States, Europe and Australia1. The efficiency forecast comes from Scott Guthrie, executive vice president of the cloud division, according to The Verge2.
What we don't know
- Whether anyone outside Microsoft has checked the efficiency figures: for now, only the company is providing them.
- Whether it will change its purchases of graphics processors from other companies, and by how much.
- When in 2027 the deployment will begin, and whether its cloud customers will notice any change in price.
Background
We already reported that Microsoft's new Surface Laptop Ultra has an Nvidia chip for running AI on the device itself. Maia is something else: chips that Microsoft designs for its data centers.
Sources
- 1Microsoft to expand Maia chip for the AI era
- 2Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs
Was this story useful? Yes Not really