Microsoft has expanded its in-house AI infrastructure efforts, confirming that its Azure Maia 200 accelerator is now powering production workloads, including Microsoft Copilot. The company has also outlined plans for Azure Maia 300, the next-generation AI accelerator designed to improve inference performance, memory efficiency and large-scale AI deployment economics.
Microsoft Azure Maia 300 Specifications And Features
Microsoft says Maia 300 is designed for demanding AI workloads, including long-context reasoning, real-time conversational experiences and multi-step agentic tasks. These workloads require more than raw computing power, placing greater emphasis on memory bandwidth, data movement, networking and power efficiency.
According to Microsoft, Maia 300 is expected to deliver 40% more tokens per megawatt than leading GPU-based systems. In a full configuration with 96 accelerators, the company claims the system can achieve up to seven times the rack-scale token throughput of Maia 200. The figures are Microsoft's stated performance expectations and should be viewed in the context of its own system configurations and workload comparisons. Maia 300 uses an inference-first architecture intended to handle memory-intensive and compute-heavy AI workloads more efficiently. Its design combines specialised compute, high-bandwidth memory and networking within a system optimised for large-scale deployment. The accelerator features eight HBM4 stacks, delivering up to 18TB/s of peak memory bandwidth within a 950W thermal design power envelope. Microsoft says the architecture also uses on-die SRAM as a software-managed scratchpad to reduce pressure on high-bandwidth memory during AI workloads. The chip uses a multi-chiplet design that separates compute, memory-controller and input/output functions. Its compute chiplets are built using TSMC's N2P process node, which Microsoft says helps improve performance per watt.
Maia 300 also supports advanced low-precision formats, including enhanced MXFP4. These are intended to reduce memory and bandwidth requirements for model computation and key-value cache storage, potentially enabling larger context windows and higher inference throughput within the same power budget. Microsoft's Maia 300 rack system combines 96 accelerators within a 167kW power envelope. The configuration provides 27.6TB of HBM capacity and 1.7PB/s of aggregate HBM bandwidth across its high-speed scale-up domain. The system's networking architecture is designed to connect compute, memory and communication resources more closely. Microsoft says its high-bandwidth interconnects support tensor parallelism, while a custom AI transport built on Ethernet-based networking infrastructure enables communication across accelerators and racks. The company expects this rack-level design to support more simultaneous AI interactions and larger workloads within a data centre, potentially improving the economics of expanding AI services.
Microsoft Maia 200 Features
While Maia 300 remains on the roadmap, Maia 200 is already running Microsoft's production AI workloads. These include Microsoft Copilot, while the company's MAI team is using the accelerator to develop next-generation models. GitHub Copilot and Microsoft Foundry are expected to follow by the end of 2026.
Maia 200 supports OpenAI models, Microsoft's own MAI models and leading open-source model families. Microsoft says the accelerator delivers up to 1.7 times more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences compared with leading GPU infrastructure. The experience of operating Maia 200 at scale has also informed the development of Maia 300. Microsoft plans to use the two generations across coding assistance, cloud-scale model serving, model inference and data generation, as well as latency-sensitive applications such as Dragon Copilot.
Microsoft says its Azure Maia programme is becoming a central part of its broader AI infrastructure strategy. Starting in 2027, the company plans to scale the Maia family to multi-gigawatt deployments, with Maia positioned as a primary inference system for Microsoft products and services. The strategy reflects Microsoft's push to develop AI infrastructure around its own software, models and services instead of relying exclusively on third-party accelerators. By co-designing silicon, systems and software, the company aims to optimise performance for the workloads it runs at cloud scale.
Maia 300 And 200 Availability
Maia 200 is being deployed across the United States, Europe and Australia. Microsoft says the accelerator will support additional services, including GitHub Copilot and Microsoft Foundry, by the end of the year. Meanwhile, Maia 300 is scheduled to begin production deployments in 2027, as Microsoft prepares to scale its Maia infrastructure to multi-gigawatt deployments.
Microsoft is expected to share further details about its AI infrastructure strategy and adoption trends at Microsoft Ignite 2026 in San Francisco and online next month.
Microsoft's Maia roadmap highlights how AI infrastructure competition is increasingly shifting towards inference efficiency, memory bandwidth and the cost of serving AI at scale. Maia 200's move into production gives Microsoft an opportunity to evaluate its custom silicon against real workloads, while Maia 300's claimed efficiency and throughput gains point to a more ambitious expansion in 2027. The key test will be how these improvements translate into real-world performance, availability and costs for Microsoft's AI services and Azure customers.
