Overview
- Microsoft announced Maia 300 on Monday and said production deployments will begin in 2027 as it scales the Azure Maia family to multi‑gigawatt capacity.
- The company claims Maia 300 will deliver about 40% more tokens per megawatt than leading GPU systems and that a full 96‑accelerator rack can provide up to seven times the token throughput of Maia 200.
- Maia 300 uses a multi‑chiplet architecture with compute chiplets on TSMC’s advanced node, eight HBM4 stacks for high on‑package memory bandwidth, and dense racks rated at roughly 167 kilowatts each.
- Microsoft is negotiating high‑volume manufacturing — reports cite plans ranging from hundreds of thousands to over one million units — while saying Maia will be used mainly for inference and Nvidia and AMD will remain important for model training.
- Maia 200 is already serving Microsoft workloads such as Microsoft 365 Copilot and GitHub Copilot, but the efficiency and scale claims for Maia 300 are company targets that require independent verification and successful large‑scale manufacturing to change Azure economics.