Microsoft plans to bring its Maia 300 chip into production AI workloads in 2027. The company describes the third-generation in-house processor as a primary system for inference—the computing used to run AI models and respond to requests across Microsoft products and services.
The announcement signals a larger role for Microsoft-designed chips in its cloud infrastructure. Maia 300 is intended to support newer AI tasks while improving the efficiency of running workloads, according to the company.
A chip designed for demanding AI tasks
Microsoft says Maia 300 is built for long-context reasoning, conversational services where response time matters, and agentic tasks that require sustained, multiple steps. These uses place different demands on AI infrastructure, from handling extensive context to maintaining a sequence of actions.
The chip follows Maia 200, which Microsoft says is already running production AI workloads across Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and the company’s own AI models. Microsoft executive vice president Scott Guthrie called Maia 200 the best-performing AI infrastructure in the company’s fleet by real-world service metrics.
Maia 300 is positioned as a complete AI system for the workloads emerging now. That framing suggests Microsoft is designing the processor around how its services use models, rather than treating it as a standalone component.
Microsoft’s efficiency and throughput targets
Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems. Tokens are units of text processed or produced by an AI model, so the measure connects energy use with model output.
In a full 96-accelerator configuration, Microsoft says the new chip can provide up to 7x the rack-scale token throughput of Maia 200. These are company projections, and the source does not provide independent measurements or detailed conditions for the comparisons.
The figures point to the goals behind the design: process more AI output for the energy used and increase the amount of work handled across a rack. If achieved in production, those improvements could help Microsoft serve its own AI services at greater scale.
A custom architecture for different chip functions
Maia 300 uses a chiplet architecture built on TSMC’s N2P process node. Its compute, memory controller, and I/O functions are separated, allowing each part to use technology suited to its role, according to Microsoft’s Azure hardware and infrastructure team.
Microsoft says the design also increases the silicon area devoted to compute performance. Separating functions gives the company a way to tailor the architecture to the different needs of processing, memory management, and communication within the system.
This approach is part of Microsoft’s effort to improve efficiency across the full AI system. The company connects the chip’s design to both cloud workloads and the needs of its own AI models.
Building more of Microsoft’s AI stack in-house
Microsoft says it will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, with Maia serving as the primary AI inference system for its products and services. That plan would make the company more reliant on its own chips for inference, while reducing its dependence on processors from Nvidia and AMD.
The company also sees Maia as important to its model development. Microsoft AI CEO Mustafa Suleyman said the company needs to work across the full stack, down to silicon, and described running Microsoft AI models on Maia as a critical focus.
That close connection between chips and models is central to the announcement. A processor designed alongside the company’s AI work could give Microsoft more control over how its models run in Azure and its products. The extent of that shift will become clearer as Maia 300 reaches production workloads in 2027.