# Microsoft wants its own Maia 300 chip, not Nvidia’s GPUs, to run its AI from 2027 > Source: > Published: 2026-10-12 13:45:00+00:00 A prototype of Microsoft’s Azure Maia 300 AI accelerator, in a Microsoft press photo released on October 12, 2026. Image: Microsoft Microsoft wants its own chips, not someone else’s GPUs, to run most of its AI. The company [unveiled Maia 300](https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/) on Monday, the third generation of its in-house AI accelerator, and said it will scale the Maia family to “multi-gigawatt deployments” from 2027, making it “the primary inference system for Microsoft products and services”. Inference is the work of running AI models once they are trained: answering a Copilot question, writing code in GitHub Copilot, powering an agent. It is where most of the day-to-day cost of AI now sits, and Microsoft says Maia 300 is built for exactly the workloads that are growing fastest: long-context reasoning, quick back-and-forth conversation and multi-step agentic tasks. ## What Maia 300 promises Microsoft’s headline claim is that Maia 300 will deliver **40% more tokens per megawatt** than “leading GPU-based systems”, and that a full rack of 96 accelerators will produce up to seven times the token throughput of a Maia 200 rack. These are Microsoft’s own expectations for a chip that isn’t in production yet; it hasn’t named the GPU systems it compared against, and no independent tests exist. Production deployments start in 2027. The specifications Microsoft has published, in Executive Vice President Scott Guthrie’s announcement and a [technical deep dive](https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/), include: - **Chiplets on TSMC N2P:** compute chiplets are made on TSMC’s N2P process, with memory-controller and input/output functions split onto separate dies. Microsoft says this lifts compute by 36% over Maia 200. - **Memory:** eight HBM4 stacks per accelerator, giving 288 GB of capacity and 18 TB/s of bandwidth within a 950-watt power budget, which Microsoft says is 2.5 times the bandwidth of Maia 200. - **Racks:** 96 accelerators in a 167-kilowatt rack, double Maia 200’s density, with up to 128 racks linked into a “superpod” of 12,288 accelerators over Ethernet-based networking. - **Smaller numbers:** support for an enhanced MXFP4 format and sub-3-bit storage, so models use less memory and can handle longer contexts. Microsoft says the rack design as a whole delivers more than 2.5 times the performance per watt of Maia 200. ## Maia 200 is already running Copilot The announcement doubles as a progress report on Maia 200, which Microsoft [introduced in January](https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/). Guthrie wrote that it now serves production workloads including Microsoft Copilot, running OpenAI models, Microsoft’s own MAI models and open-source models, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. It is being deployed in the US, Europe and Australia, and Microsoft says it delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for AI features in Teams, Copilot and Cowork. Those efficiency numbers are again Microsoft’s own. In a video released with the announcement, Guthrie put it more bluntly: Maia 200 is the best performing AI infrastructure in our fleet, measured on real-world metrics for our AI services. Scott Guthrie, Executive Vice President, Cloud + AI, Microsoft ## ‘A platform that we own’ Microsoft AI chief Mustafa Suleyman, whose team builds the MAI models, tied the chip directly to the company’s superintelligence ambitions. He said lessons from running MAI models on Maia 200 “have gone straight into the design of the Maia 300”: At that scale, we need a platform that we own and that we co-design with our models. Mustafa Suleyman, CEO, Microsoft AI Microsoft did not say whether Azure customers will be able to rent Maia 300 directly, or how many chips it plans to build. It is a big change of tone from a company that has spent tens of billions of dollars on Nvidia hardware. Like Google with its TPUs and Amazon with Trainium, Microsoft is betting that a chip designed around its own models and services will run them more cheaply than general-purpose GPUs. The news comes as chipmakers race to meet AI demand: [AMD became a $1 trillion company](https://madrobot.blog/2026/09/21/amd-joins-1-trillion-club-chipmakers-rally-ai-demand/) last month, and Chinese labs such as DeepSeek are [building software to break Nvidia’s hold](https://madrobot.blog/2026/09/30/deepseek-huawei-ascend-open-source-tilelang-nvidia-cuda-alternative/) as well. Microsoft also [gave Copilot new agent powers on Windows](https://madrobot.blog/2026/10/07/microsoft-windows-copilot-local-files-actions-mxc-agent-sandbox-surface-laptop-ultra/) last week, the kind of workload Maia is meant to carry. Microsoft says it will share more at its Ignite conference in San Francisco next month. ## Why it matters If Maia becomes Microsoft’s main inference engine, one of Nvidia’s biggest customers will be running much of Copilot and its other AI services on its own silicon. The efficiency claims are untested until the chips ship in 2027, but the plan itself is a clear sign that the biggest AI companies want to own the hardware their models run on. Sources: [Official Microsoft Blog (Scott Guthrie)](https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/), [Azure Blog: Deep dive into Azure Maia 300 architecture](https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/), [Microsoft Maia 300 press kit and video transcripts](https://news.microsoft.com/maia-2026/).