Microsoft unveiled Maia 300, its third-generation custom AI accelerator, on 12 October 2026; production deployments begin in 2027 across Azure
It is built for AI inference (running trained models), with compute chiplets on TSMC's 2-nm-class N2P process and eight HBM4 stacks giving 288 GB of memory at 18 TB/s
Microsoft claims 40% more tokens per megawatt than leading GPU systems and, per 96-chip rack, up to 7x the token throughput of Maia 200
Maia is to become the primary inference system for Microsoft's products, including Copilot and OpenAI models hosted on Azure
Training builds a model; inference uses the trained model to answer new inputs. A chiplet design joins several smaller dies in one package instead of one giant die, while HBM stacks memory dies beside the processor for very high bandwidth.
Simple Analogy: Training writes the textbook; inference is answering questions from it.
Designs the Maia chips for its Azure cloud; will use them for Copilot, GitHub Copilot and OpenAI models on Azure
Contract chipmaker fabricating the Maia 300 compute chiplets on its N2P node
First-generation Maia (Maia 100) introduced
Maia 200 launched on TSMC 3 nm
Maia 300 unveiled
Maia 300 production deployments begin
General Awareness > Science and Technology > Products and launches
Running a trained AI model to produce outputs from new inputs
Stacked memory placed next to the processor to move data at very high rates