Maia 200, unveiled January 26, 2026, is Microsoft's second-generation in-house accelerator and is built for inference: more than 140 billion transistors on TSMC 3 nm, over 10 PFLOPS of FP4, 216 GB of HBM3E at 7 TB/s and 272 MB of on-chip SRAM in a 750 W package. It scales to 6,144 accelerators over an Ethernet-based two-tier network. Microsoft says it is in production serving Microsoft 365 Copilot, Foundry, OpenAI GPT models and MAI models, deploying across the US, Europe and Australia. It is used inside Microsoft services rather than sold as a chip, and no public Azure instance had been announced as of November 2026.
规格参数与关键信息
每项数据均标注来源依据:厂商规格、厂商声明、媒体报道、估算或官方数据。"未公开披露"即字面含义,我们从不编造数值。
Overview
Designer
Microsoft
Announced
January 26, 2026Official
Microsoft blog
Status
In production in Azure for Microsoft servicesOfficial
Oct 2026; first region US Central (Iowa), then US West 3 (Arizona); Microsoft blog
Manufacturing
Process
TSMC 3 nmVendor spec
Microsoft blog, Jan 2026
Transistors
Over 140 billionVendor spec
Microsoft blog, Jan 2026
算力
FP4
>10 PFLOPSVendor spec
Per chip; Microsoft blog, Jan 2026
FP8
>5 PFLOPSVendor spec
Per chip; Microsoft blog, Jan 2026
内存
Memory
216 GB HBM3EVendor spec
Microsoft blog, Jan 2026
Memory bandwidth
7 TB/sVendor spec
Microsoft blog, Jan 2026
On-chip SRAM
272 MBVendor spec
Microsoft blog, Jan 2026
电力
SoC TDP
750 WVendor spec
Microsoft blog, Jan 2026
Physical
Scale-up
2.8 TB/s bidirectional per chip; 4 per tray; clusters up to 6,144Vendor spec
Ethernet-based two-tier network; Microsoft blog, Jan 2026
Cooling
Second-generation closed-loop liquid heat exchanger unitVendor spec
Microsoft blog, Jan 2026
Performance
Vendor comparison
~30% better performance per dollar than Microsoft's prior fleet hardware; 3x FP4 of Trainium3Vendor claim
Microsoft blog, Jan 2026
Efficiency
Up to 1.7x tokens per watt vs leading GPU infrastructure (Microsoft services)Vendor claim
Microsoft blog, 2026-10-12
Customers
Workloads
Microsoft 365 Copilot, Microsoft Foundry, OpenAI GPT-5.2, MAI models; GitHub Copilot by end of 2026Official
Microsoft blogs, Jan and Oct 2026
与技术栈其他层级的关联
本条目关联一览:点击分组可直接跳转。 +1 更多分组见下方。
数据来源
Microsoft: Maia 200, the AI accelerator built for inference (Jan 2026) ↗ · Microsoft: Azure Maia to multi-gigawatt deployments (Oct 2026) ↗
记录审核时间 2026-11-16; 查看 研究方法.