Microsoft announced on Monday, October 12, 2026, that it will expand its Azure Maia accelerator family to multi-gigawatt deployments starting in 2027 and that Maia will become the primary inference system for the company’s products and services. In an official blog post by Scott Guthrie, executive vice president of Cloud + AI, the company introduced Azure Maia 300, the third generation of the program, with production deployments expected to begin in 2027, and confirmed that Maia 200 is already in production serving Microsoft workloads.
According to Guthrie, Maia 200 is already handling production workloads including Microsoft Copilot and supporting the MAI team as it develops next-generation models, with GitHub Copilot and Microsoft Foundry to follow by year-end. Across those workloads, Maia 200 runs OpenAI models, Microsoft’s own MAI models and leading open-source model families. The company says co-engineering between hardware and software, informed by the scale of its artificial intelligence services, enables optimization for real-world enterprise applications and industry scenarios.
Maia 300 is presented as a complete system designed for emerging workloads: long-context reasoning, latency-sensitive conversational experiences and sustained multi-step agentic tasks. Microsoft expects Maia 300 to deliver 40% more tokens per megawatt compared with leading GPU-based systems. In a full 96-accelerator configuration, it says the system delivers up to seven times the rack-scale token throughput of Maia 200. CNBC, in a same-day report by Jordan Novet, stressed that the company intends to rely mainly on its own silicon to handle incoming inference requests.
The rack figures Microsoft published are specific: a Maia 300 rack system brings together 96 accelerators within a 167-kilowatt power envelope, delivering 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth across a high-speed scale-up domain. That doubles the number of chips per rack versus Maia 200, according to CNBC. Each Maia 300 accelerator uses eight HBM4 stacks providing 18.0 TB/s of peak memory bandwidth within a 950-watt TDP, and a multi-chiplet architecture that separates compute, memory-controller and I/O functions, with compute chiplets built on TSMC’s N2P process node.
The design emphasizes a software-defined dataflow architecture and AI-assisted kernel optimization to maximize locality and reduce data movement, an increasingly critical bottleneck in large-scale inference. Microsoft highlights advanced low-precision formats, including enhanced MXFP4, to cut memory and bandwidth needs for both computation and KV-cache storage, enabling larger context windows and higher throughput within the same power envelope. Scale-up networking is treated as an extension of that dataflow, with a custom AI transport over Ethernet infrastructure that unifies topologies from a single rack to datacenter scale.
The competitive context is an industry seeking alternatives to Nvidia GPUs without abandoning them entirely. CNBC notes that Microsoft introduced Maia 100 in 2023 and touted Maia 200 this past January; chief executive Satya Nadella told analysts in July that Maia 200 delivers 30% better performance per dollar than the latest-generation hardware in Microsoft’s fleet and was already handling OpenAI models. Finance chief Amy Hood said at a Goldman Sachs conference last month that use of the company’s own chips has helped with margin improvement over a decade.
Evercore ISI analysts, in a client note cited by CNBC, estimated that Maia will represent about 10% of Microsoft’s spending on AI chips by the first quarter of 2027, and recommended buying Microsoft shares. That figure does not come from Microsoft and should be read as a third-party estimate. Alistair Speirs, a Microsoft general manager, declined to provide a precise timeline for multi-gigawatt deployments and spoke of “optionality” in a fast-changing industry. Jim Kleewein, a technical fellow, said parts of Microsoft 365 Copilot should speed up with Maia 300.
What Maia is not also matters. Unlike Google’s tensor processing units or Amazon’s Inferentia and Trainium, Microsoft’s accelerators cannot be rented by the hour to other cloud customers, according to CNBC. Patrick Moorhead, chief executive of Moor Insights and Strategy, observed that nobody has tested them outside Microsoft yet. For model training, Speirs said Microsoft still partners with Nvidia and AMD. Nvidia chief executive Jensen Huang spoke last week at an event where Microsoft announced a Surface laptop containing an Nvidia processor, a sign that the commercial relationship continues.
Maia 200 is already deploying across the United States, Europe and Australia, and Microsoft says it delivers up to 1.7x more tokens per watt for built-in AI capabilities in Teams, Copilot and Cowork experiences compared with leading GPU infrastructure. Operating experience on Maia 200, Guthrie wrote, directly informed Maia 300’s design. Together, the two generations are meant to support additional services: multi-step agentic work for GitHub, cloud-scale model serving for Foundry, inference and data generation for the MAI team, and latency-sensitive interactions for Dragon Copilot in clinical workflows.
The bet has clear economic implications for the cloud. Speirs told CNBC that new architectures are needed for better use of available power, because power translates fairly directly into operational cost for a data center. A 40% gain in tokens per megawatt, if realized in production, would change the unit economics of Copilot and other high-volume services. At the same time, reliance on custom silicon concentrates execution risk: manufacturing delays at TSMC, performance shortfalls or software issues could slip 2027 schedules.
For the chip market, the announcement reinforces a visible trend: large hyperscalers design custom accelerators for inference, while Nvidia and AMD remain central for training and many general-purpose deployments. Microsoft frames Maia within what it calls Frontier Transformation — a heterogeneous, model-diverse platform open to silicon and model choice. The company says it will share more at Microsoft Ignite 2026 next month in San Francisco and online, and points to a technical paper and a dedicated Maia 300 site for architectural detail.
Open questions remain. There is no public figure for the exact gigawatts planned or for the share of Copilot traffic already running on Maia 200. It is also unclear when, if ever, Azure will offer Maia as a rentable instance to external customers. Evercore ISI’s estimate that Maia could reach about 10% of Microsoft’s AI chip spend by the first quarter of 2027 offers a provisional benchmark, but it does not replace Microsoft’s official financial disclosure.
Taken together, October 12 marks Microsoft’s shift from presenting Maia as a custom-silicon experiment to declaring an infrastructure ambition: multi-gigawatt scale, primary inference duty and energy efficiency measured in tokens per megawatt. If Maia 300 delivers in 2027 what was announced — 96 accelerators per rack, 7× Maia 200’s rack throughput and 40% more tokens per megawatt versus leading GPUs — the company will have tangibly reduced its relative dependence on external suppliers for serving AI at scale. Training, for now, remains shared ground with Nvidia and AMD.
