Maia 200 by Microsoft: status and timeline

Milestones

Next · Deployed in Azure US Central

Maia 200 designed and taped out2025Complete.
Announced at Microsoft event26 Jan 2026Complete.
Deployed in Azure US CentralTarget Jan 2026Current milestone.

Most important updates

Microsoft announces Maia 200 inference chip built on TSMC 3nm with 216GB HBM3e26 Jan 2026

Upcoming

Jan 2026Deployed in Azure US Central (next)

Current obstacles

Production yield and volumeTSMC 3nm is mature but custom design requires yield tuning and volume ramp.
Software ecosystemRequires PyTorch, Triton, and custom compiler optimization to realize hardware potential.

Physics limits

Lower precision limits model sizeFP4/FP8 good for inference but quantization artifacts grow with model size.
Power scales with token throughputReal power depends on utilization. Low load reduces efficiency-per-token.

How it works

3 parts

Tensor cores

Native FP4 and FP8 arithmetic

Computes in lower precision, reducing memory and power while maintaining inference accuracy.

Memory system

High-bandwidth HBM3e

216GB at 7 TB/s keeps large models fed. Data movement engines optimize token generation.

Power efficiency

750W at peak compute

Designed for datacenters with liquid cooling, targeting cost-per-token for production.

Spec sheet

Spec	Maia 200
Manufacturing process	TSMC 3nm^R (reported)
Transistor count	140+ billion^R (reported)
HBM3e capacity	216 GB^R (reported)
Memory bandwidth	7 TB/s^R (reported)
On-chip SRAM	272 MB^R (reported)
TDP	750 W^R (reported)
FP4 peak performance	10+ petaFLOPS^R (reported)
FP8 peak performance	5+ petaFLOPS^R (reported)
FP4 vs Trainium3	3x higher^R (reported)

^R reported by the company

Update log

1 update

Mon 26 Jan

Major: Microsoft announces Maia 200 inference chip built on TSMC 3nm with 216GB HBM3ePressHardware

About Microsoft

The team behind Maia 200

Microsoft

AI models, cloud computing, Azure, GitHub Copilot

Microsoft develops software, cloud, and hardware across computing and AI. It operates Azure and GitHub. In June 2026, Microsoft launched frontier AI models in-house, gaining freedom after its OpenAI partnership ended in April 2026.

Microsoft launched seven frontier AI models June 2026, trained in-house from scratch with zero competitor distillation.
GitHub Copilot reached 50M users by July 2026; 1 in 3 PRs now involve an AI agent.
Maia 200 (3nm, 216GB HBM3e) ships to Azure US Central, claims 3x FP4 vs Trainium3.

microsoft.com (opens microsoft.com)Company profile →

Founded: 197551 yrs
Headquarters: United States
Status: Public
Valuation: $3.5Tmarket capOct 2026
Works in: AIFrontier modelsMaterialsComputing
Coverage: 5 programs · 6 updateslatest 9d agochecked 10 Oct
Partners: OpenAI (opens openai.com)GitHub (opens github.com)
Suppliers: NVIDIA (opens nvidia.com)TSMC (opens tsmc.com)
People: Satya Nadella (opens en.wikipedia.org)CEOKevin Scott (opens en.wikipedia.org)Chief Technology Officer

Spec

Maia 200

Manufacturing process

TSMC 3nm^R (reported)

Transistor count

140+ billion^R (reported)

HBM3e capacity

216 GB^R (reported)

Memory bandwidth

7 TB/s^R (reported)

On-chip SRAM

272 MB^R (reported)

TDP

750 W^R (reported)

FP4 peak performance

10+ petaFLOPS^R (reported)

FP8 peak performance

5+ petaFLOPS^R (reported)

FP4 vs Trainium3

3x higher^R (reported)