Microsoft’s AI chip targets 40% more tokens per megawatt in 2027

What it claims

Microsoft says Maia 300, its third-generation in-house accelerator, will become its primary inference system across its products from 2027. The headline figures from Scott Guthrie are 40% more tokens per megawatt than "leading GPU-based systems" and up to 7x Maia 200's rack-level token throughput in a 96-accelerator rack. Training stays on Nvidia and AMD. Guthrie's stated plan is to grow the Azure Maia family into multi-gigawatt deployments starting in 2027.

Evidence

Every figure below is a vendor or analyst claim reported second-hand through The Verge and CNBC. The page has no independent measurement.

• Rack scaling: CNBC reports the Maia 300 rack holds twice as many accelerators as the Maia 200 rack. With 2x the chips, most of the claimed 7x gain must come from per-chip and system improvements rather than density. This inference is ours; the page gives no per-chip figure.
• Predecessor track record: Maia 200 was introduced in January 2026. It already serves production traffic for Microsoft Copilot, GitHub Copilot, Microsoft Foundry and Microsoft's own models, and CNBC reports it also runs OpenAI models. In July, Satya Nadella said it delivered 30% better performance per dollar than the latest-generation hardware in Microsoft's fleet.
• Spend share: Evercore ISI estimates Maia will be about 10% of Microsoft's AI-chip spending by Q1 2027. That means GPUs remain the large majority of the fleet in the near term.
• Workload target: long-context reasoning, latency-sensitive conversation and multi-step agent tasks. Jim Kleewein says parts of Microsoft 365 Copilot should get faster.

How to use it

You cannot use it directly. Microsoft's accelerators are not available for other companies to rent by the hour. The only way an outside engineer reaches Maia is indirectly, through Microsoft services that run on it. The page names Foundry and the Copilot products as running on Maia 200 today.

Limits

• The "leading GPU systems" baseline is not named. No model, context length, batch size or precision is given for the tokens-per-megawatt or throughput claims, so you cannot compare them with your own numbers.
• There are no specs: no memory capacity or bandwidth, no FLOPS, no interconnect details. The only process detail is a chiplet design on TSMC N2P with separate compute, memory-controller and I/O dies.
• Maia 300 is a pre-production announcement; production workloads start in 2027.
• The article states it was produced with AI assistance and reviewed by editors. It relays quotes and does not report independently.