Microsoft unveiled Azure Maia 300 on October 12, 2026, announcing plans to scale its custom AI accelerator family to multi-gigawatt deployments starting in 2027. The company positions Maia as its primary internal inference system for Copilot, GitHub, and OpenAI models while maintaining third-party hardware investments.
On October 12, 2026, Microsoft introduced Azure Maia 300, marking the third generation of its in-house AI accelerator family. Executive Vice President of Cloud + AI Scott Guthrie announced that the company will scale the Azure Maia architecture to multi-gigawatt deployments starting in 2027, establishing Maia as the core inference system across Microsoft products and services.
Building Maia 300
Maia 300 departs from uniform memory structures by adopting a non-uniform memory access (NUMA) design across chiplets built on TSMC’s N2P process node. The architecture separates compute, memory-controller, and I/O functions to maximize efficiency for multi-step agentic tasks. Built for conversational AI and long-context reasoning workloads where models must digest massive amounts of text before replying, the chip features specialized kernel support.
The system introduces support for advanced low-precision formats, including enhanced MXFP4, which shrink memory footprints and data movement. These enhancements allow the hardware to handle larger models and extended context windows within a stable power envelope while utilizing eight HBM4 memory stacks.
Eight HBM4 memory stacks deliver up to 18 terabytes per second of peak memory bandwidth. According to technical specifications published by the company, this configuration yields performance leaps over previous hardware.
- HBM bandwidth increases by 2.5 times compared with Maia 200.
- Rack-scale token throughput reaches up to 7 times the capacity of Maia 200 in a full 96-accelerator configuration.
Deploying Maia 300
While Maia 300 is scheduled for production deployments in 2027, its predecessor is already embedded in daily operations. Azure Maia 200 is operational in data centers across the United States, Europe, and Australia, powering production workloads for Microsoft Copilot and supporting the MAI team in developing next-generation models.
GitHub Copilot and Microsoft Foundry are slated to incorporate Maia 200 by the end of 2026. Across these services, the second-generation silicon runs OpenAI models, proprietary MAI models, and leading open-source model families. Company data indicates that Maia 200 delivers up to 1.7 times more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot, and Cowork compared to leading GPU infrastructure.
The multi-gigawatt rollout planned for 2027 is designed to support more intensive applications, including multi-step agentic coding workflows for GitHub, cloud-scale model serving, and latency-sensitive Dragon Copilot healthcare interactions intended to reduce administrative burdens for clinicians.
Partnering With Hardware Vendors
Microsoft is using Maia primarily for internal workloads. Matt Kimball, vice president and principal analyst at Moor Insights & Strategy, noted that this internal focus changes the economic dynamics of the chip deployment.
The custom silicon strategy aims to trim infrastructure expenses. Microsoft’s annual capital expenditure plans exceed $175 billion, though industry estimates indicate custom chips like Maia will account for only about 10% of total AI chip spending by early 2027. Savings from the transition are expected to accumulate gradually rather than immediately.
Microsoft Unveils Maia 300 AI Chip to Reduce Dependence on Nvidia GPUsMicrosoft is taking another major step toward building its own artificial intelligence infrastructure with the introduction of the Azure Maia 300, a custom AI chip designed to reduce its reliance on Nvidia’s… pic.twitter.com/Z0fVWhRD6q
Despite the push into proprietary hardware, Nvidia and AMD remain central to Microsoft’s infrastructure, particularly for demanding AI training tasks. The reliance of Maia 300 on TSMC’s N2P node also ensures that foundry partnerships remain foundational whether chips originate from external vendors or in-house engineering teams.
More on Azure Maia
Microsoft unveils its new Azure Maia 300 AI chip, strengthening its push to power the next generation of AI infrastructure.Can Microsoft’s custom AI chips help it compete with NVIDIA? What do you think?#Microsoft #Azure #Maia300 #AI #ArtificialIntelligence #TechNews pic.twitter.com/OJWJKniaTW