Microsoft announces Azure Maia 300 AI chip for 2027 production

Microsoft announced the Azure Maia 300 on 12 October 2026, its third-generation AI accelerator designed to reduce reliance on Nvidia chips. Production begins in 2027, with 96 Maia 300 chips placed into each data center rack, about twice the number used for Maia 200. The announcement marks Microsoft's push to run its own AI inference workloads on in-house silicon.upi.com

The Maia 300 matters because it represents Microsoft's most ambitious effort yet to reduce dependence on Nvidia and AMD chips for AI workloads. Unlike previous generations, Microsoft positions the Maia 300 as the primary AI inference system for its products and services. Scott Guthrie, executive vice president of Microsoft's cloud group, said the company expects Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems.

What the announcement changes

In a full 96-accelerator configuration, the Maia 300 delivers up to 7x the rack-scale token throughput of Maia 200, according to Guthrie. Microsoft plans to scale the Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary AI inference system for Microsoft products and services. The efficiency gains will allow Microsoft to rely mainly on its in-house AI chips rather than those from Nvidia and AMD.

Microsoft uses a new chiplet architecture to achieve its efficiency gains. Built on TSMC's N2P process node, the Maia 300 separates compute, memory controller, and I/O functions. The Azure hardware and infrastructure team said this increases the silicon area dedicated to compute performance and allows each function to use the technology best suited to its role. The design includes two I/O chiplets, four chiplets with memory controllers, and eight HBM4 stacks with 72 GByte capacity each.

How the chip is built

Mustafa Suleyman, Microsoft AI CEO, said the company must work across the full stack down to the silicon to compete with Google DeepMind, OpenAI, and Anthropic. Suleyman said running Microsoft AI models on top of the Maia chip is a critical focus for the team. Microsoft has relied on Nvidia's GPU technology, as have many industry leaders, with Nvidia GPUs used in data centers and AI model training for OpenAI and Anthropic.theverge.com

Microsoft is not yet completely abandoning Nvidia technology. In its announcement of its new line of Microsoft Surface laptops, specifications include Nvidia processing power. Microsoft also works with technology by AMD. Amazon Web Services and Google Cloud have each already developed multiple generations of their own AI chips, including AWS Inferentia and Trainium, and Google TPU. Microsoft has been in the Maia race for about three years.

What happens next

As of 12 October 2026, no third-party companies have tested the Maia 300. Microsoft will begin production on Maia 300 units in 2027. The company intends to place 96 Maia 300 chips into each of its data center racks, about twice as many as the Maia 200 chips. Maia 200 systems were announced with 10 Pflops at FP4 accuracy.

Microsoft plans to equip its Azure data centers with Maia 300 pods starting in 2027, with total power consumption reaching several gigawatts. At approximately 1000 watts per chip, this translates to several million accelerator chips. For every million Maia 300 chips, Microsoft must also purchase 288 Petabyte of HBM4 capacity, representing several billion US dollars in investments.