Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs

Maia 300 is Microsoft's third-generation chip for artificial intelligence. Its main job is inference: running trained AI models to produce outputs for products and services. Microsoft designed it to improve the cost and efficiency of cloud AI, especially as models become larger and more demanding.

An AI chip performs the huge number of calculations needed by a model. Maia 300 is aimed at long-context reasoning, conversational services that need quick responses, and multi-step agentic tasks. In a full 96-accelerator configuration, it can deliver up to seven times Maia 200's rack-scale token throughput.

Maia 300 is not yet running production workloads. Microsoft plans to begin rolling it out in 2027. The company expects it to deliver 40 percent more tokens per megawatt than leading GPU-based systems, supporting Microsoft's plan to make Maia its primary AI inference system.

Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems. Tokens are the small units of text that AI systems process and generate. The comparison measures how much AI output the system can produce for a given amount of electricity.

This matters because cloud AI services consume substantial computing resources. Better tokens-per-megawatt efficiency can help Microsoft serve more requests without increasing power use at the same rate. Microsoft also says Maia 300 can deliver up to seven times the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration.

The 40 percent figure is an expectation, not a reported production result. Maia 300 is scheduled to start rolling out to production AI workloads in 2027. Its chiplet design, built on TSMC's N2P process, is intended to support these efficiency gains.

Using Maia 300 could reduce Microsoft's dependence on Nvidia and AMD because Microsoft would have an in-house system designed specifically for its own AI workloads. The expected efficiency improvement means the company could produce more tokens for each megawatt than leading GPU-based systems. It would therefore have greater control over capacity, performance, and deployment.

Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt. In a full 96-accelerator configuration, it says the chip provides up to seven times Maia 200's rack-scale token throughput. These figures could make Maia attractive for large Azure deployments and Microsoft services.

Microsoft does not say it will immediately stop using outside chips. Maia 300 is scheduled for production rollout in 2027. The company says it will scale the Azure Maia family to multi-gigawatt deployments and make Maia the primary AI inference system for its products and services.

Maia 300 is designed for newer AI workloads rather than only straightforward model responses. Microsoft highlights long-context reasoning, latency-sensitive conversational experiences, and sustained, multi-step agentic tasks. These workloads can require an AI system to process more information, respond promptly, and continue operating across several connected steps.

For example, a long-context system may need to work through a large amount of material before answering. A conversational service must keep response delays low. An agentic task may involve several actions in sequence. Maia 300's purpose is to provide the infrastructure for these patterns, while also handling the latest reasoning models.

The chip is a successor to Maia 200, which already runs production workloads including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and Microsoft's own AI models. Maia 300 is expected to begin production rollout in 2027, with efficiency and throughput improvements aimed at scaling these services.

Microsoft wants control of the full AI stack because it believes large-scale AI requires coordination from models and software down to silicon. Owning the platform can let the company design computing infrastructure around the needs of its own models and services. It also supports Microsoft's effort to compete with Google DeepMind, Axiom, and Paradox.

Microsoft AI CEO Mustafa Suleyman says the company must work across the full stack and co-design its platform with its models. That makes running Microsoft AI models on Maia a critical focus. The chip is therefore more than a separate hardware project; it is part of Microsoft's broader AI model and infrastructure strategy.

The immediate plan is to use Maia 300 as the primary AI inference system for Microsoft products and services. Microsoft expects to scale the Maia family to multi-gigawatt deployments starting in 2027. The article does not claim that every AI component will be made in-house.

Maia 300 uses a chiplet architecture instead of treating every chip function as one undivided block. The design separates compute, memory controller, and input/output functions. Compute handles model calculations, memory control manages movement of data, and I/O connects the chip with other system components.

This separation lets each function use technology best suited to its role. Microsoft says it also increases the silicon area dedicated to compute performance. In practical terms, the architecture can avoid making every part of the chip follow the same design priorities. That supports the efficiency gains Microsoft is targeting for AI inference.

The chip is built on TSMC's N2P process node. Microsoft links the chiplet design to Maia 300's expected performance and efficiency, including 40 percent more tokens per megawatt than leading GPU-based systems. The article does not provide separate efficiency figures for each chiplet.

Specialized AI chips are important because cloud services must run AI models repeatedly and at large scale. A chip designed around inference can prioritize the calculations, data movement, speed, and power efficiency that those workloads need. General-purpose processors remain useful, but they are not designed specifically for every AI operation.

Maia 300 illustrates this approach. Microsoft expects 40 percent more tokens per megawatt than leading GPU-based systems, and up to seven times Maia 200's rack-scale token throughput in a full 96-accelerator configuration. Its chiplet architecture separates compute, memory control, and I/O so each function can be optimized for its role.

The article presents this as part of Microsoft's move toward in-house infrastructure. Maia 200 already runs workloads such as Microsoft Copilot and GitHub Copilot. Maia 300 is planned for production in 2027 and is intended to become Microsoft's primary AI inference system.