Microsoft says Maia 200 serves workloads across 3 regions as Maia 300 targets 2027

What is new about Maia 200’s rollout?

Microsoft says Maia 200 is supporting production workloads, including Copilot, and running models from OpenAI, Microsoft MAI and open-source models. The company also says the chip is being deployed across the three areas named in its announcement. Those are company-reported details, not an independent count of the chips installed or a third-party check of the workloads.

Microsoft separately says it plans to bring Maia 200 to GitHub Copilot and Microsoft Foundry by the end of 2026. That is a future target for those named services, not confirmation that they have already moved to the chip. Microsoft’s references to current Copilot workloads and its future GitHub Copilot plan do not establish that all Copilot services run on Maia.

Microsoft announced its first Maia accelerator in 2023. GeekWire describes Maia 300 as the third generation of the company’s custom AI chip program. The Maia 200 update therefore changes the picture of the family’s current status, not the status of the Maia 300 product itself.

What workloads is Maia 300 designed for?

Microsoft describes Maia 300 as an inference-focused system for long-context, memory-intensive and decode-heavy workloads. Its examples include conversations where latency matters and sustained, multi-step tasks handled by AI agents. These are Microsoft’s intended uses, not evidence that the chip has delivered those services in production.

The company’s design description extends beyond the processor: it covers software, networking, racks and data-center infrastructure as well as the silicon. Microsoft identifies TSMC’s N2P process for the chip’s compute chiplets in its Maia 300 architecture overview. The sources describe a system built around those components, rather than a chip specification alone.

For a full rack with 96 Maia 300 accelerators, Microsoft lists a 167-kilowatt power envelope, 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth. HBM is the memory technology named in Microsoft’s specifications; these figures describe the company’s planned rack configuration.

Microsoft gives each accelerator 288 gigabytes of HBM, peak bandwidth of 18.0 terabytes per second and a 950-watt thermal design power rating. Those are stated design specifications, not independently measured results from Maia 300 hardware operating in production. The planned rack’s power envelope also does not, by itself, establish how much electricity a deployed system would use for a given workload.

How strong are Microsoft’s performance claims?

Microsoft says Maia 300 can deliver up to 40% more tokens per megawatt than leading GPU-based systems. It also says a full 96-accelerator Maia 300 rack can reach up to seven times Maia 200’s rack-scale token throughput. Both figures are comparisons made by Microsoft, not independently verified results in the reporting reviewed.

The GPU systems behind the 40% comparison are not identified in the prior coverage, and the materials reviewed do not provide independent, like-for-like testing. That leaves important context unresolved: which systems and workloads are being compared, and whether the claimed advantage holds under comparable conditions. The figures should be read as Microsoft’s projections for its design, not as demonstrated gains in deployed data centers.

Microsoft says experience operating Maia 200 informed Maia 300’s design. The reported Maia 200 rollout gives context for that statement: Microsoft says an earlier generation is supporting workloads in production. It does not test Maia 300’s claimed efficiency or throughput, or verify the company’s target for scaling the family.

Could Maia 300 reduce Microsoft’s reliance on Nvidia?

Microsoft’s plan to make Maia the primary inference system for its products and services could shift some of that work to its own chips if deployments reach the planned scale. But the announcement does not quantify what share of Microsoft’s inference workloads might move, how much Nvidia capacity the company will continue to use, or whether the plan would reduce costs.

GeekWire reports that Nvidia remains a key supplier to Microsoft’s data centers. It also reports plans for Microsoft to deploy systems based on Nvidia’s Vera Rubin and AMD’s Helios. That reporting points to a mix of suppliers and in-house silicon, rather than an announced end to Nvidia purchases.

The public plan concerns Microsoft products and services; the sources reviewed do not establish whether outside customers will be able to rent Maia 300 capacity directly through Azure. They also do not specify how many Maia 300 accelerators are planned or where the company expects to install them. The main milestones still ahead are the production deployments Microsoft says will begin in 2027 and independent, comparable evidence of the chip’s performance.