Microsoft Unveils Maia 300 as Its Primary AI Inference Chip

Microsoft Unveils Maia 300 as Its Primary AI Inference Chip

Microsoft has introduced the Maia 300, its third-generation AI inference chip. The company first entered the custom silicon space with an AI inference chip in 2023. In January 2026, it launched the Maia 200, which runs OpenAI models and Microsoft’s own MAI models. The Maia 200 already supports production workloads, including Copilot, and is designed for next-generation MAI models. The new Maia 300 brings further gains in efficiency and is set to become Microsoft’s main AI inference platform.

Azure Maia Programme Expansion

Microsoft announced the update in a press release on Monday. The company is expanding its Azure Maia programme and increasing its use of in-house silicon for AI inference. The Maia 200 is already deployed in select markets across the US, Europe, and Australia. It delivers up to 1.7 times more tokens per watt than comparable systems. The chip currently powers AI features in Microsoft Teams, Copilot, and Cowork. It runs both OpenAI models and Microsoft’s MAI models. Support for GitHub Copilot and Microsoft Foundry is planned by the end of the year.

Maia 300 Specifications and Performance

The Maia 300 is claimed to deliver 40 percent more tokens per megawatt than comparable GPU-based inference systems. It will serve as Microsoft’s primary AI inference platform and will power the company’s in-house AI products and services. The chip is optimised for long-context, memory-intensive, and decode-heavy workloads. It balances compute, memory bandwidth, networking, and power efficiency at cloud scale.

The Maia 300 uses a software-defined dataflow architecture with AI-assisted kernel optimisation for efficient data movement. It offers higher HBM bandwidth per watt through eight HBM4 stacks, providing 18TB/s of peak memory bandwidth within a 950W TDP. The chip is built on TSMC’s N2P process node. It features 288GB of HBM capacity, a two-tier Ethernet scale-up with AI Transport, and a pod size of 96 Maia chips per rack.

Microsoft’s latest chip strengthens its push toward custom silicon for large-scale AI inference. The Maia 300 is positioned to handle growing production demands across its AI services.