With Maia 300, Microsoft is accelerating its in-house AI chip strategy

Microsoft unveils Maia 300, its new AI chip announced to be 40% more efficient than competing GPU systems, with a massive deployment planned as early as 2027.

Microsoft is taking a new step in the development of its own artificial intelligence infrastructure. On October 12, the company unveiled Maia 300, the third generation of its internally designed accelerators, accompanied by a large-scale deployment program starting in 2027. Satya Nadella's stated ambition is to make the Maia family the primary inference system used by Microsoft products and services, with installations representing several gigawatts of electrical power. An alternative to traditional GPU infrastructures

With Maia 300, Microsoft seeks to improve the energy efficiency of its data centers as AI computing demands continue to grow. According to the company's projections, the new generation is expected to produce 40% more tokens per megawatt compared to leading GPU-based systems.

The chip is specifically designed for inference, which is the execution of AI models once they are trained. It notably targets conversations requiring fast responses, long-context processing, and agents capable of executing multi-step operations.

Microsoft also announced a token throughput up to seven times higher than that of Maia 200, at the scale of a full rack of 96 accelerators. This comparison applies to entire hardware configurations, rather than the performance of an isolated chip. An architecture designed for future generations of AI

Maia 300 relies on an architecture composed of multiple chiplets, whose compute units are manufactured using TSMC's N2P process. It embeds eight HBM4 memory modules, offering a maximum bandwidth of 18 TB/s for a specified thermal design power of 950 watts.

A full rack combines 96 accelerators, 27.6 TB of HBM memory, and a power density of 167 kW. Microsoft has also developed a data flow system designed to limit unnecessary transfers between compute units, memory, and network connections.

The company details this architecture in a technical publication dedicated to Maia 300. Maia 200 already powers Microsoft services

This new generation builds on the experience gained with Maia 200, currently deployed in the United States, Europe, and Australia. These accelerators already execute workloads related to Microsoft Copilot and models developed by the MAI team, as well as models from OpenAI and open-source models.

Microsoft claims up to 1.7 times more tokens per watt for Maia 200 than leading GPU infrastructures on certain Teams, Copilot, and Cowork features. GitHub Copilot and Microsoft Foundry are also scheduled to integrate this infrastructure by the end of 2026. Microsoft wants to regain control over its AI infrastructure

By developing its own accelerators, Microsoft seeks to better manage the costs, power consumption, and performance of its AI services. This strategy could also reduce its dependence on GPU suppliers, notably NVIDIA and AMD, without ending the use of their technologies.

The first production deployments of Maia 300 are planned for 2027. If the announced gains are confirmed at scale, Microsoft would have an additional lever to absorb the growth in usage of Copilot, language models, and autonomous agents. The challenge is no longer just having high-performing models, but being able to run them simultaneously for millions of users with controlled energy consumption and costs.