Microsoft will take Azure Maia to multi-gigawatt deployments and make it its primary inference system from 2027

Microsoft announced on October 12, 2026, in a post by Scott Guthrie, executive vice president of Cloud + AI, that it will scale its Azure Maia accelerator family to multi-gigawatt deployments starting in 2027. The goal is for Maia to become the primary inference system for the company's products and services. The post presents the Maia program as a foundational element of its AI infrastructure strategy. This summary is based solely on Microsoft's official blog post. The performance figures are the company's own claims and have not been independently verified.

Maia 200 is already in production. According to Microsoft, it serves production workloads, including Microsoft Copilot, and supports the MAI team in developing next-generation models. GitHub Copilot and Microsoft Foundry will be added before the end of the year. In these workloads, Maia 200 runs OpenAI models, Microsoft's own MAI models and open-source model families. It is being deployed in the United States, Europe and Australia. Microsoft says it delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for the AI features built into Teams, Copilot and the Cowork experiences.

The company emphasizes hardware-software co-design. Because silicon development takes years, its software and hardware teams work together from the outset to anticipate architectures and scale, translating that into more memory capacity, greater bandwidth and more accelerators operating as a single system. Experience with Maia 200 directly informed the design of Maia 300. Among the intended uses for both generations, it cites multi-step agentic work for GitHub, cloud-scale model serving for Foundry, inference and data generation for the MAI team, and low-latency interactions for Dragon Copilot, which aims to reduce clinicians' administrative burden.

Maia 300 is the program's third generation and is conceived as a complete system with an inference-first architecture. It is designed for long-context reasoning, latency-sensitive conversational experiences and sustained multi-step agentic tasks. Microsoft expects it to deliver 40% more tokens per megawatt than leading GPU-based systems. In the full 96-accelerator configuration, it says it achieves up to 7 times Maia 200's rack-scale token throughput. Production deployments of Maia 300 will begin in 2027, and Microsoft points to a technical paper for architectural details.

The post describes several key capabilities of Maia 300. The first is less data movement: according to Microsoft, moving data, rather than raw compute power, is increasingly the bottleneck in large-scale inference. It therefore combines a software-defined dataflow architecture with AI-assisted kernel optimization to maximize locality and utilization. The company credits this approach with a significant performance-per-watt advantage over traditional GPU architectures.

The second is memory bandwidth. Maia 300 uses eight HBM4 stacks delivering 18.0 TB/s of peak bandwidth within a 950-watt TDP. Microsoft presents this as the industry's best HBM bandwidth per watt. In addition, specialized kernels use on-chip SRAM as software-managed scratch memory, reducing pressure on the HBM.

The third is a chiplet architecture that separates compute, memory controller and I/O functions, with which the company aims to improve efficiency, scalability and cost structure. The compute chiplets are manufactured on TSMC's N2P node to maximize performance per watt. The fourth is support for advanced low-precision formats, including an enhanced MXFP4, which reduces memory and bandwidth requirements for both compute and KV cache storage. The result, it says, is larger context windows and higher throughput within the same power envelope.

The fifth is scale-up networking. Microsoft treats it as an extension of the dataflow architecture, in which operations span compute, memory and network as a single system. High-bandwidth interconnects between accelerators are optimized for tensor parallelism. A proprietary AI transport, built on Ethernet infrastructure, enables a unified topology that scales from a single rack to an entire data center.

At the rack level, Maia 300 brings together 96 accelerators in a 167-kilowatt envelope. It provides 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth within a high-speed scale-up domain. Microsoft argues that this density allows it to handle more simultaneous AI interactions and larger workloads per rack, improving data center capacity and the economics of expanding its AI services.

Strategically, Microsoft frames Maia within what it calls “Frontier Transformation.” It stresses that the future will not be defined by a single model or a single chip, and that choice of model and silicon is inherent to its vision. It says it can optimize end to end, from silicon to systems to agents, on an open, heterogeneous, model-diverse platform. The announcement does not replace GPUs. It sits within that heterogeneous platform and gives no volumes, investment figures or timelines beyond 2027.

The company adds that it will share more about its strategy, technology adoption trends and Frontier Transformation results next month at Microsoft Ignite 2026 in San Francisco and online. More details, photos and videos are available on the dedicated Maia 300 site.

The comparisons should be read with caution. The figures of 1.7x tokens per watt and 40% more tokens per megawatt are measured against “leading GPU infrastructure” or “leading GPU-based systems,” without the text naming the reference product or the methodology. The Maia 300 figures are also design expectations (“we expect”) for a system that is not yet in production.