Microsoft says its own Maia 300 chip will run most of its AI inference from 2027
Training stays on Nvidia and AMD for now. Answering Copilot queries is moving onto Microsoft's own silicon.
By The desk, News · · 1 min read
Tags: microsoft · chips · maia · inference
- The news
- Training stays on Nvidia and AMD for now.
- Why it matters
- Answering Copilot queries is moving onto Microsoft's own silicon.
- What's confirmed
- The photograph is generated for this story.
Tied to a named source. Not marked verified. Microsoft Azure blog
The news
Microsoft announced Maia 300, built on TSMC's N2P process with eight HBM4 memory stacks and a 950-watt power budget. Cloud chief Scott Guthrie said Microsoft will scale the Maia family to multi-gigawatt deployments starting in 2027 and make it the primary inference system for Microsoft's own products.
Why it matters
Inference is where most AI compute spending is heading. A hyperscaler saying it will serve most of it on its own chips is a direct threat to Nvidia's share of that budget.
The details
Microsoft claims 40 percent more tokens per megawatt than leading GPU systems and up to seven times the per-rack throughput of Maia 200. Both figures are vendor-reported, and the comparison GPU is not named. A rack holds 96 chips in a 167-kilowatt envelope.
CNBC reported that Microsoft will keep using Nvidia and AMD hardware to train models.
What we could not check
- The GPU used for Microsoft's efficiency comparison was not named.
- No independent benchmark of Maia 300 exists yet.
Sources
The photograph is generated for this story. It does not show the people named in it.