Microsoft unveils Azure Maia 300, its own AI chip that will become the backbone of its systems from 2027

The new chip and its performance

On Monday, 12th, Microsoft unveiled the Azure Maia 300, the third generation of its artificial intelligence processor. From 2027 onwards, the chip will serve as the primary platform for the company’s products and services. Microsoft estimates that the system will deliver over 40% more tokens per megawatt compared to leading GPU-based solutions.

What does tokens per megawatt mean?

This metric measures the amount of content processed or generated by AI relative to the electrical energy consumed. A higher value indicates that the infrastructure can accomplish more work with the same power.

Focus on inference and AI agents

The Maia 300 was designed to prioritize inference, the phase where a trained model processes new information to generate responses or predictions. The chip has also been optimized to handle extensive contexts, tasks requiring large memory, and complex response generation processes. These capabilities are essential for AI agents that can perform multi-step activities, such as querying information, using tools, and iterating until a task is completed.

Integration and control strategy

Microsoft designs the chip, software, networks, racks, and data centers as a unified system to execute agent tasks at cloud scale. This approach provides greater control over cost and performance, while ensuring that every dollar invested delivers more intelligence. The company maintains partnerships with Nvidia and AMD and is preparing its infrastructure for the Vera Rubin generation while expanding the use of Maia.

Context and timeline

The launch comes approximately nine months after the introduction of the Maia 200, which already runs models from OpenAI, Microsoft, and other companies. In November 2023, Microsoft announced the first chips in this strategy: the Maia 100 for AI tasks and the Cobalt 100 for general-purpose computing in Azure.

Frequently Asked Questions

• What is the Azure Maia 300?

  It is the third generation of Microsoft’s proprietary AI chip, which will become the company’s primary AI processing platform from 2027.

• What is the announced efficiency gain?

  Microsoft claims the system will deliver over 40% more tokens per megawatt compared to leading GPU-based systems.

• What does tokens per megawatt mean?

  It is the ratio between the volume of content processed by AI and the energy consumed, indicating higher productivity per unit of energy.

• What is inference?

  It is the stage where a trained AI model processes new information to generate a response or prediction.

• Will Microsoft stop using Nvidia chips?

  The strategy combines proprietary chips with products from other vendors. The company remains a partner of Nvidia and AMD and is preparing its infrastructure for the Vera Rubin generation.