Microsoft unveils its Maia 300 AI chip
Microsoft's Maia 300 is built for reasoning models and AI agents, with production deployments planned for 2027.
Microsoft says Maia 300 is designed to become the next major step in its in-house AI infrastructure, with production deployments planned for 2027. The chip is intended for long-context reasoning, conversational workloads and multi-step agentic tasks, rather than general-purpose consumer hardware.1
What Microsoft is announcing
Microsoft says Maia 300 is its third-generation AI chip and will deliver 40% more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, the company claims up to seven times the rack-scale token throughput of Maia 200.1
The system uses eight HBM4 stacks, a 950-watt thermal design power and a multi-chiplet architecture built around TSMC's N2P process. Microsoft says the complete rack design can combine 96 accelerators, 27.6TB of HBM capacity and 1.7PB/s of aggregate HBM bandwidth.1
Why it matters
Microsoft currently relies heavily on external AI hardware, so Maia 300 gives Azure another way to control the cost and efficiency of serving its own models and products. Computerworld reports that the chip is expected to reduce Microsoft's dependence on Nvidia and AMD as more workloads move to Microsoft's own silicon.2
The announcement is also about scale, not a retail product. Microsoft says Maia 200 is already serving production workloads, while Maia 300 will begin production deployments in 2027. The company has not announced a consumer release or a standalone price.
Our opinion
Maia 300 is most interesting as infrastructure news, not as another chip launch for shoppers to compare. Microsoft's headline efficiency figures are promising, but they are still company claims made before independent testing and broad deployment. The real test will be whether Azure customers notice lower costs and better response times once the hardware is running at scale.
For now, the announcement shows Microsoft trying to own more of the AI stack. That should give it more control over performance and supply, while making its cloud ambitions less dependent on the companies that currently dominate accelerator hardware.