Microsoft Unveils Maia 300 AI Chip to Cut GPU Reliance
Microsoft unveiled its Maia 300 AI chip, a custom processor arriving in 2027 designed to act as its primary inference engine and curb its dependence on third-party hardware.
Microsoft has unveiled the Maia 300, its third-generation custom silicon designed to handle the tech giant's heaviest cloud infrastructure workloads. Scheduled for production deployment in 2027, the new processor is set to serve as the primary AI inference engine across Microsoft's products and services. Constructed on TSMC's N2P process node, the Maia 300 utilizes a modular chiplet architecture that isolates compute, memory controller, and I/O duties to maximize the silicon area dedicated strictly to processing performance.
The hardware is tailored to handle specialized tasks such as long-context reasoning, latency-sensitive conversational experiences, and multi-step agentic workflows. Executive vice president Scott Guthrie stated that the chip is expected to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems. Furthermore, in a full 96-accelerator configuration, the system produces up to 7x the rack-scale token throughput of the previous-generation Maia 200. First announced in January, the Maia 200 is already powering live production workloads, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and internal foundational models.
This hardware evolution allows Microsoft to scale multi-gigawatt deployments while systematically reducing its dependence on external suppliers like Nvidia and AMD. Microsoft AI CEO Mustafa Suleyman highlighted that controlling the full technology stack from silicon to software is essential for matching rival labs like OpenAI, Google DeepMind, and Anthropic. For cloud practitioners and enterprise developers, this dedicated hardware layer offers significantly higher energy efficiency and enhanced throughput for next-generation AI applications running on Azure.
This is our own summary of reporting by The Verge AI
