Microsoft unveils Maia 300, its third-generation in-house AI chip: production starts in 2027 while Maia 200 already runs Copilot workloads

Start with what the official post actually says. It is signed by Scott Guthrie, Microsoft's executive vice president for Cloud + AI, titled 'Microsoft to expand Azure Maia system to multi-gigawatt deployments', and dated October 12, 2026. Three things are central. First, from 2027 Microsoft will scale the Azure Maia family to multi-gigawatt deployments and make Maia the primary inference system for its products and services. Second, the second-generation Maia 200 is already in production, serving Microsoft production workloads including Microsoft Copilot and supporting the MAI team as it develops next-generation models, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. Third, Maia 300's stated targets are the three workload types emerging today: long-context reasoning, latency-sensitive conversational experiences, and sustained multi-step agentic tasks. The timeline is explicit: Maia 300 will begin production deployments in 2027.

Maia 200's rollout deserves its own paragraph. Microsoft says it is deploying across the United States, Europe and Australia, delivering up to 1.7x more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences compared with leading GPU infrastructure. One easily missed sentence matters more than it looks: across these workloads, Maia 200 runs OpenAI models, Microsoft's own MAI models, and leading open-source model families. The point is that in-house silicon does not mean in-house models — the chip is closer to a common substrate, and the models on top can still come from several places. Microsoft also stresses that its software and hardware teams collaborate early to translate requirements into greater memory capacity, higher bandwidth and more accelerators operating as a single system, and that this co-engineering is already delivering measurable efficiency gains — with Maia 300's design directly shaped by running real workloads on Maia 200.

The technical specification is the hardest part of the post. Maia 300 is described as an inference-first architecture optimised for long-context, memory-intensive and decode-heavy AI workloads. On memory: eight HBM4 stacks provide 18.0 TB/s of peak memory bandwidth inside a 950-watt TDP, which Microsoft calls industry-leading HBM bandwidth per watt. On architecture: a new multi-chiplet design separates compute, memory-controller and I/O functions, with compute chiplets built on TSMC's N2P node to maximise performance per watt, plus on-die SRAM used as a software-managed scratchpad to relieve pressure on HBM. On precision: the design targets advanced low-precision formats including enhanced MXFP4, cutting memory and bandwidth requirements for computation and KV-cache storage in exchange for larger context windows. At rack level: 96 accelerators fit inside a 167-kilowatt power envelope, providing 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth, with networking treated as a seamless extension of the software-defined dataflow architecture and an Ethernet-based unified topology that scales from a single rack to datacenter scale. The two relative figures Microsoft offers: 40% more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200 in a 96-accelerator configuration.

How to read it. First, check the reference points behind the numbers: the 40% comparison is against what Microsoft calls leading GPU-based systems, and the 7x is a rack-scale comparison with the same vendor's previous generation, Maia 200. Neither is a third-party benchmark against a specific GPU, and the post says 'we expect'. Second, the language is unusual: calling a chip the primary inference system for a company's products and services is a roadmap-level commitment, not a technology preview. The Verge's headline the same day read 'Microsoft's new AI chip makes it less reliant on Nvidia's GPUs', which is outside media reading the same official announcement. Third, Microsoft says it will share more on strategy and adoption next month at Microsoft Ignite 2026 in San Francisco and online. For ordinary users, little changes in the short term, but if Maia really carries the main inference load from 2027 as planned, the cost per unit of inference behind products like Copilot is the real story.