Science & Tech · AI · 6 hrs ago
Microsoft plans to make Maia chips its main AI inference system from 2027
Microsoft plans to make its Maia chips the main system for AI inference across its products and services, with large-scale deployments starting in 2027.
The plan centers on Maia 300, a new accelerator designed for the part of AI processing that generates answers one token at a time.
Maia 300 is built to deliver 2.5 times Maia 200's memory bandwidth, with 288GB of memory and a stated bandwidth of 18 terabytes per second.
That focus reflects how inference can be limited by the speed of supplying data to computing hardware, rather than by raw computing power alone.
The move could reduce Microsoft's reliance on Nvidia chips for some of its AI workloads, putting pressure on a key source of Nvidia's data-center revenue and pricing power.
Google and Amazon have also developed their own chips, as major cloud companies seek more control over AI infrastructure.
Microsoft expects to begin scaling Maia deployments in 2027, but the plan's effect on Nvidia and the wider AI chip market remains uncertain.
Microsoft plans to make its Maia accelerators the primary inference system for its products and services, with deployments starting in 2027.
The plan would scale Azure Maia systems to multi-gigawatt deployments.
The upcoming Maia 300 is designed to increase memory bandwidth for AI inference, with eight HBM4 stacks providing 288GB of memory at 18TB/s.
Microsoft is shifting toward in-house chips as other large technology companies also develop or use their own AI silicon.
- Who
- Microsoft, through its Azure Maia chip program.
- What
- It plans to make Maia its primary inference system for its products and services. Deployments are planned at multi-gigawatt scale.
- When
- Starting in 2027.
- Where
- Microsoft Azure and its products and services. Specific locations are not stated.
- Why
- The Maia 300 is designed to improve memory bandwidth for inference workloads.
This story does not have two clearly opposing sides.
Microsoft told the European Commission that Google was the only company vertically integrated at every AI layer.
Microsoft plans to begin multi-gigawatt deployments of Azure Maia accelerators and make Maia its primary inference system.
- Deployment start
- 2027
- Scale
- Multi-gigawatt
- Maia 300 memory
- 288GB across eight HBM4 stacks
- Maia 300 memory bandwidth
- 18TB/s
- Maia 200 memory bandwidth
- 7.1TB/s
What people are saying
everybody else must rely on partnerships to innovate and compete.
seeing strong demand for TPUs as standalone systems from other customers.