Overview
- Microsoft unveiled the Maia 300 on Monday, October 12, and said production rollouts will begin in 2027 as it plans to scale the Maia family to multi‑gigawatt deployments.
- Maia 300 uses a chiplet architecture with TSMC N2P 2‑nanometer compute chiplets, separate 3nm I/O chiplets, eight HBM4 stacks, and custom dataflow and networking to reduce data movement within the system.
- The company claims Maia 300 will deliver about 40% more tokens per megawatt than leading GPU systems and that a 96‑accelerator rack will offer up to seven times the rack‑scale token throughput of Maia 200.
- Microsoft intends Maia to be its primary inference platform while continuing to use Nvidia and AMD hardware for model training, and the Maia accelerators are initially for internal use only.
- Large‑volume manufacturing talks with TSMC reportedly target hundreds of thousands to over a million units for 2027 but real‑world efficiency, supply constraints and packaging capacity will determine how quickly Microsoft can hit those targets.