Microsoft unveils Maia 200 inference chip, live in US Central

Microsoft said Maia 200, its in-house accelerator built for AI inference, is now running in its US Central datacenter near Des Moines, Iowa, with US West 3 near Phoenix next. The TSMC 3nm chip is rated above 10 petaFLOPS at FP4 and carries 216GB of HBM3e memory. Microsoft claims about 30% better performance per dollar than the latest hardware in its fleet, which is a vendor claim. For a trader, lower inference cost per token is a margin lever that reduces reliance on Nvidia GPUs.

• Over 10 petaFLOPS FP4 and over 5 petaFLOPS FP8 per chip, 750W envelope
• 216GB HBM3e at 7 TB/s plus 272MB on-chip SRAM
• Claimed about 30% better performance per dollar than latest fleet hardware
• US Central live; US West 3 (Phoenix) next