Microsoft Unveils Maia 300, the AI Chip It Wants Running Most of Its Inference by 2027

Microsoft is telling customers, and Nvidia, that the next wave of its AI answers will come off its own silicon. On Monday the company unveiled Maia 300, the third generation of its in-house AI accelerator, and said large-scale deployment starts in 2027. Scott Guthrie, executive vice president for Cloud + AI, wrote that the Azure Maia family will scale to multi-gigawatt deployments from that year and become “the primary AI inference system for Microsoft products and services.” Satya Nadella called it “our biggest bet yet on our own silicon.” Maia 200 is already in production. Maia 300 is the one meant to carry Copilot, Foundry, and the rest of the inference load. The efficiency claim is the headline number. Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration it delivers up to seven times the rack-scale token throughput of Maia 200. Those are company figures, not an independent bake-off. They are also the figures that justify spending less with Nvidia and AMD on the part of the workload that is answering questions rather than training the model. What Is on the Chip Maia 300 uses a chiplet design. Compute, memory controller, and I/O sit on separate dies. The compute chiplets are on TSMC’s N2P process, a two-nanometer-class node, which Microsoft says raises the share of silicon given to compute and lets each function use the process that suits it. Eight HBM4 stacks provide 288 gigabytes of high-bandwidth memory, the same capacity figure cited for Nvidia’s Rubin, with the memory bandwidth to feed long-context and multi-step agent work. A rack holds 96 accelerators inside a 167-kilowatt envelope, with 27.6 terabytes of aggregate HBM and 1.7 petabytes per second of aggregate HBM bandwidth across the scale-up domain. Jobs can span racks, so a workload that does not fill a rack does not leave accelerators idle. The design target is the current shape of use: long-context reasoning, conversational replies that cannot wait, and agent tasks that run for many steps. Training is still a GPU market. Inference — the token coming back to the user — is where Microsoft says it can substitute. Guthrie’s “primary inference system” is that substitution stated as a plan. How Much of the Spend Actually Moves Evercore ISI analysts, cited by CNBC, estimate Maia will be about 10 percent of Microsoft’s AI-chip spending by early 2027. That is not a flip of the supplier list. It is a start. Multi-gigawatt deployments and a 10 percent share can both be true if the total spend is large. The point of the Monday post is the direction: Maia 200 in production now, Maia 300 in production next year, and a stated intention that Microsoft products run on Microsoft inference chips. GitHub Copilot and Microsoft Foundry are due to move onto the current generation before the end of this year, which is the first production proof, not the 2027 one. Custom silicon is how the other hyperscalers have tried to cap the Nvidia bill. Google has TPUs. Amazon has Trainium. Microsoft’s first Maia generation was late and limited. The 300 is the generation the company is willing to call the primary system. Whether the 40 percent tokens-per-megawatt number survives a customer workload is the test that starts in 2027, not the test that was run for the blog post. Power is the other constraint. A 167-kilowatt rack and a multi-gigawatt campus plan land in the same market that has Samsung selling memory at record margins and utilities signing nuclear deals with Google. Efficiency per megawatt is not a benchmark vanity metric if the megawatts are the scarce input. Microsoft is arguing that its chip uses them better than a GPU for the inference job. The argument is internal until a third party measures it. What to Watch Ignite in San Francisco next month, where Microsoft said it will give more detail. Whether Copilot and Foundry actually move onto Maia 200 on the schedule Guthrie implied. The 2027 deployment pace against the “multi-gigawatt” and “primary system” lines. Any Nvidia comment on inference share, which is the revenue line this chip is aimed at. Independent tokens-per-watt comparisons once hardware is in a lab that is not Microsoft’s. Maia 300 is Microsoft’s third AI chip and the one it wants running most of its inference from 2027. The claims are 40 percent more tokens per megawatt than leading GPUs and seven times the rack throughput of the previous Maia. The spending share, on one analyst estimate, is still about 10 percent early that year. The bet is that the share rises. The silicon is the vehicle.