Microsoft Will Run Its AI Stack Mostly on Its Own Silicon: Maia 300 Goes Multi-Gigawatt in 2027

Microsoft picked the OCP Global Summit in San Jose on October 12 to make its boldest infrastructure claim yet: starting in 2027, the company will scale its Azure Maia chip family to multi-gigawatt deployments and make it "the primary AI inference system for Microsoft products and services." The announcement came from Scott Guthrie, the executive vice president who runs Microsoft's cloud and AI group, in a blog post timed to the show.

The chip at the center of the plan is Maia 300, Microsoft's third-generation in-house accelerator. Its predecessor, Maia 200, was only announced in January, but it is already running production workloads across the company, including Microsoft 365 Copilot, with GitHub Copilot and Microsoft Foundry scheduled to follow by the end of 2026. Microsoft says Maia 200 runs OpenAI models, its own MAI models and leading open-source models, is deployed in the United States, Europe and Australia, and delivers up to 1.7 times more output per watt than leading GPU-based systems for the AI features built into Teams, Copilot and Cowork.

Maia 300 is pitched at the workloads that are becoming the expensive part of AI: long-context reasoning, latency-sensitive conversation, and sustained multi-step agentic tasks. Microsoft says it will deliver 40% more tokens per megawatt than leading GPU-based systems, and that a full rack of 96 accelerators — twice the chip count of a Maia 200 rack — will deliver up to seven times the rack-scale token throughput of Maia 200. The published specifications put the accelerator at eight HBM4 stacks with 18 TB/s of peak memory bandwidth and a 950-watt thermal design power, and a full rack at 27.6 TB of HBM within a 167-kilowatt power envelope. Compute chiplets are manufactured on TSMC's N2P process, with separate chiplets for memory control and I/O. All of those figures are company claims and have not been independently verified.

The engineering story underneath the numbers is chiplets and data movement. In a separate post, Microsoft's engineers said a major focus was cutting the power wasted moving data back and forth inside servers — a cost that grows as models handle longer contexts and longer agent sessions. The racks are liquid-cooled by default, but Microsoft says they can also run in older data centers cooled by air, which matters for retrofitting capacity rather than building new shells.

The strategic reading is simpler: inference is where hyperscaler money is heading, and Microsoft wants to stop renting that layer from Nvidia. Nvidia still draws more than half of its quarterly revenue from hyperscalers, and Microsoft is the largest of them. Amazon, Google and Meta are all building their own accelerators, and Amazon and Google now rent their custom chips to third parties. OpenAI and Anthropic are designing silicon too. Microsoft is late to that party relative to Google's TPU program, but it is arriving with the largest commercial software estate to feed.

There are real limits to the claim. Microsoft has not said when the multi-gigawatt expansion completes, and outside customers still cannot rent Maia accelerators by the hour the way they can Google's TPUs or Amazon's Trainium. Model training remains a Nvidia-and-AMD affair. Analysts at Evercore ISI, in a report to clients last week, estimated Maia will account for only about 10% of Microsoft's AI chip spending by the first quarter of 2027. Microsoft shares rose more than 1% intraday on the news.

For the industry, the announcement marks the point where "buy vs. build" stops being a hedge and becomes the default plan at the biggest buyer of GPUs. The number to watch in 2027 is not the marketing ratio of tokens per megawatt but whether Maia 300 racks actually absorb Copilot, Foundry and Microsoft's own MAI traffic at scale — because every token Microsoft moves onto its own silicon is a token Nvidia no longer sells.