Microsoft unveils Maia 300 AI chip, due in production in 2027
The accelerator is built for inference, when a model answers. Comparisons with rival GPUs are Microsoft's own estimates, with no independent test.
For people who use Copilot or pay for AI services on Azure, Maia 300 changes nothing for now: Microsoft has not given a price, regions or a date for customers, and production for deployment would begin only in 2027, according to SiliconANGLE. The chip is the third generation of AI accelerators Microsoft designs in-house and was unveiled on Oct. 12.
Microsoft described it on the Azure blog, in a post by Rani Borkar and Saurabh Dighe. It is an inference chip, meaning it handles the stage where an already trained model generates answers, a cost that grows with every user of an assistant.
What it packs
The specifications come from the Azure blog and SiliconANGLE's write-up:
- Four compute chiplets, pieces of silicon mounted together in a single package, made by TSMC on a 2-nanometer-class process.
- Eight HBM4 memory modules with 288 GB in total and 18 TB/s of bandwidth per accelerator.
- Liquid-cooled racks of 96 accelerators drawing 167 kW, grouped into "superpods" of up to 12 288 accelerators on a custom Ethernet-based network.
The memory matches the capacity of Nvidia's Rubin accelerator, but with about 20% lower bandwidth, according to SiliconANGLE.
The numbers are Microsoft's
Microsoft says Maia 300 delivers more than 2.5 times the performance per watt of Maia 200, its previous chip, and that a full rack produces up to 7 times as many tokens as one with Maia 200. It also estimates 40% more tokens per megawatt than "leading GPU-based systems," without naming which one, according to SiliconANGLE.
None of those figures comes from an independent test. The 7x figure compares against the company's own earlier chip, so it does not show that Maia 300 beats an Nvidia GPU, and the 40% figure does not specify which system it was measured against.
Still unanswered: whether Azure will offer Maia 300 to outside customers or Microsoft will use it only in its own services, and whether any savings will reach the price of those services. The Azure blog says more details will come at Microsoft Ignite 2026.
