Microsoft says its Maia 200 AI accelerator is already running production workloads, including Copilot, and plans to begin production deployment of Maia 300 in 2027. The company wants Maia to become its primary system for inference—the computing used to run trained models and generate responses—which could shift some work away from Nvidia GPUs.
On Monday, Oct. 12, Microsoft announced plans to scale Maia deployments to multiple gigawatts starting in 2027. The expansion plan is part of the company’s effort to build AI infrastructure around its own services. Microsoft’s performance figures are company comparisons and projections, not independent benchmark results.
What work is Microsoft’s Maia chip designed to do?
AI computing includes model training and inference. AFP describes inference as running trained models to answer questions and carry out tasks, and Microsoft says it designed Maia primarily for that work.
Microsoft says Maia 200 is serving production workloads that include OpenAI models, its own MAI models and leading open-source model families. The company also says Maia 200 is being deployed in the United States, Europe and Australia.
By the end of 2026, Microsoft expects GitHub Copilot and Microsoft Foundry to join the workloads already running on Maia 200. Those planned additions would bring the company’s chip into more of its AI products and services.
Microsoft says Maia 300 is designed for long-context reasoning, conversations where response time matters, and sustained, multi-step tasks by AI agents. The company describes its design as balancing computing, memory bandwidth, networking and power efficiency.
How could Maia reduce Microsoft’s reliance on Nvidia GPUs?
If Microsoft moves more inference work to Maia, it could rely less on Nvidia GPUs for that part of its AI infrastructure. That would be a shift in where some tasks run, not a replacement of Nvidia hardware across Microsoft’s systems.
The distinction between training and inference helps explain the competition. AFP describes Nvidia as holding a near-lock on chips used to train AI models, while chipmakers and cloud companies compete to supply inference hardware. Its report on Microsoft’s announcement names Nvidia, AMD, Amazon and Google among companies introducing products for inference.
Microsoft’s stated ambition is to make Maia its primary inference system for the company’s products and services. Designing its own accelerator could give Microsoft more control over the hardware supporting those workloads and let it direct some computing away from a third-party supplier.
That also concentrates more infrastructure decisions within Microsoft: the company can determine which of its workloads run on Maia as it scales the system. The plan is not just to develop a chip, but to expand Maia deployments to multiple gigawatts.
What do Microsoft’s Maia performance figures show?
Microsoft says Maia 200 can deliver up to 1.7 times more tokens per watt than leading GPU infrastructure for selected built-in AI features in Teams, Copilot and Cowork. The comparison is tied to those specified features.
Free newsletter
Get the morning briefing
Start each day with the stories that matter and why — a short, free email from our newsroom.
For Maia 300, Microsoft projects 40% more tokens per megawatt than leading GPU-based systems. The company also says a full Maia 300 rack with 96 accelerators could deliver up to seven times the rack-scale token throughput of Maia 200.
Microsoft lists Maia 300’s peak memory bandwidth as 18.0 terabytes per second and its thermal design power as 950 watts. It gives a power envelope of 167 kilowatts for a rack containing 96 accelerators. The efficiency and throughput figures are company claims or projections, not independently verified comparisons across different systems and workloads.
The scale of the planned buildout makes electricity part of the story alongside efficiency. It also reports that rising AI-related electricity demand is putting pressure on power grids and increasing costs for technology companies building data centers.
More output per unit of power does not necessarily mean lower total electricity use if computing capacity expands. Microsoft’s efficiency claims need to be considered alongside its plan to scale Maia to multiple gigawatts.
Who makes Maia, and what will the rollout mean?
Microsoft designs the Maia system, but it relies on outside manufacturing for a key part of the next generation. The company says Maia 300’s compute chiplets will use TSMC’s N2P process.
That division of work shows how power in AI computing is distributed. Microsoft is seeking greater control over the hardware behind its services, while TSMC supplies the process used to manufacture Maia 300’s compute chiplets. Maia could reduce Microsoft’s reliance on Nvidia for some inference workloads without eliminating other supplier dependencies.
Microsoft says Maia 300 production deployments will begin in 2027, alongside the start of the broader expansion. The rollout will put the company’s projected efficiency gains to a real-world test and show how much inference its Maia systems handle.
Comments
Comments are written by readers. They are not reporting or opinion from The Wells Post.
Share your view on this story. Criticise ideas and public records, not other readers.
Most comments appear right away; some wait for a moderator first.
Community guidelines
More in our terms and privacy policy.
No comments yet. Start the conversation.