Microsoft Unveils Maia 300 Inference Chip With Larger Racks and Faster Memory

Microsoft has announced Maia 300, its latest custom AI chip for the Azure cloud. A Microsoft announcement page and an Azure engineering blog post, both dated 12 October 2026, describe a chip designed to run trained AI models. Microsoft says it targets agentic workloads, where a model plans steps, calls tools and reads long inputs.[1][2]

The engineering post holds most of the detail. It describes a new memory layout, a doubled rack and a larger network, and compares them with Maia 200, the previous generation. The figures are Microsoft’s. No independent tests, price or customer date has been published.[2]

Microsoft’s announcement page is titled "Microsoft’s latest custom chip: Maia 300". It carries renders of the chip, a server and a rack, and lists talks by Scott Guthrie, executive vice president of the Cloud + AI Group. The page has little technical text.[1]

The Azure blog post says Maia 300 is built as a complete system, with software, silicon, racks, networking and data centres designed together. Microsoft says agentic workloads change the load on inference hardware, and it names multi-step execution, long-context decoding and tool use. Its stated aim is more useful tokens per watt and per dollar.[2]

Data Center Knowledge reports that Microsoft expects Maia 300 in production in 2027 and plans multi-gigawatt Azure Maia deployments that year. We could not open Microsoft’s own post on that timetable, so treat the 2027 dates as third-party reporting.[3]

Microsoft says Maia 300 places compute on TSMC’s N2P process and moves memory control and input-output functions onto separate dies. The post credits this layout with 36% more compute operations, 2.5 times the memory bandwidth and 14% more scale-up bandwidth. Each accelerator has eight HBM4 stacks, giving 288 GB of capacity and 18 TB/s of bandwidth.[2]

At rack level, 96 accelerators sit in 24 trays of four, inside a 167-kilowatt rack that Microsoft says doubles Maia 200’s density. Liquid cooling removes more than 80% of the rack’s heat. A rack can connect to facility water or use a heat-exchanger unit in air-cooled data centres, which widens where it can be installed.[2]

A two-tier network links up to 128 racks into a superpod of 12,288 accelerators, using Ethernet-based transport. Racks use electrical links inside and optics between racks. Microsoft says software can treat the superpod as one system.[2]

Microsoft says a full 96-accelerator rack delivers up to seven times the token throughput of Maia 200, and more than 2.5 times the performance per watt. Both claims compare Maia 300 with Microsoft’s earlier chip, not with a rival product. The text of the post does not say which models or workloads produced the numbers.[2]

Data Center Knowledge adds that Microsoft expects 40% more tokens per megawatt than leading GPU-based systems, another Microsoft comparison. It notes that Microsoft still relies on Nvidia and AMD GPUs and, unlike Amazon and Google, does not offer its chip to customers as an Azure instance.[3]