Microsoft announces Maia 200 inference chip built on TSMC 3nm with 216GB HBM3e

Microsoft/Maia 200

Microsoft unveiled Maia 200, a custom inference accelerator on TSMC 3nm. The chip features 216GB HBM3e, native FP4/FP8 cores, and 750W TDP. Microsoft claims 3x FP4 throughput versus Amazon Trainium3 and 30% better cost-per-performance than prior generation. Deployed to Azure US Central.

• TSMC 3nm process with 140+ billion transistors and 216GB HBM3e memory at 7 TB/s.
• 10+ petaFLOPS FP4, 5+ petaFLOPS FP8 performance at 750W TDP.
• Targets inference cost-per-token economics for production AI workloads.
• Deployed to Azure US Central in January 2026; Azure US West 3 expansion planned.