Microsoft announced its third‑generation Maia 300 AI inference chip on Oct. 12, 2026. The company says the chip is built to run most of its cloud AI services, reducing reliance on Nvidia and AMD GPUs. Maia 300 will begin rolling out to production workloads in 2027.
What makes Maia 300 different?
The chip uses a chiplet design on TSMC’s N2P process. Compute, memory controller and I/O are split onto separate pieces of silicon, which lets more area be dedicated to raw compute. Microsoft claims this architecture lets the chip deliver 40% more tokens per megawatt than leading GPU‑based systems. In a full 96‑accelerator rack, it can provide up to seven times the token throughput of the earlier Maia 200.
Maia 300 is aimed at newer AI workloads such as long‑context reasoning, low‑latency chat and multi‑step agentic tasks. The chip is already part of Microsoft’s internal stack, supporting products like Microsoft Copilot and GitHub Copilot.
Why it matters
If the efficiency claims hold up, Azure customers could see lower costs for AI services that run on Microsoft’s own hardware. The shift also means Microsoft will have more control over the hardware‑software stack, which could speed up feature rollouts for its AI products.
Why it matters
Azure users may benefit from cheaper AI inference as Microsoft’s own chips become the primary hardware. For developers, tighter integration between Microsoft’s models and its silicon could mean faster updates and new features for services like Copilot.
