ProductAnalysisMicrosoft announces Azure Maia 300 custom AI inference accelerator
What happened.Microsoft announced Azure Maia 300, its third-generation custom AI accelerator designed for high-density inference and multi-step agentic execution. The hardware uses a multi-chiplet architecture on TSMC's N2P node with 288 GB of HBM4 memory delivering 18 TB/s bandwidth per accelerator.
Why it matters.Cloud providers are optimizing inference silicon around low-precision arithmetic and high local memory bandwidth rather than merchant training clusters. Builders running long-context agent loops will see lower per-token serving costs as proprietary silicon takes over production traffic.
The technical specifications:
- Detail: Built on TSMC's N2P process with 288 GB of HBM4 memory delivering 18 TB/s peak bandwidth per accelerator.
- Detail: Standard racks house 96 Maia 300 chips within a 167-kilowatt power envelope, scaling to 12,288 accelerators per superpod.
- Detail: Rack-scale token throughput increases sevenfold compared to Maia 200, with volume deployment beginning in 2027.