Research note
Research note: Microsoft details Maia 300, its third-generation in-house AI inference chip
How we separated Microsoft’s published specs from its performance claims for Maia 300.
What we set out to check
What Microsoft published about Maia 300, which numbers are specifications and which are performance claims, and when it will be used.
What we found
- Microsoft’s blog gives the process node, memory, rack and superpod design, and comparisons with Maia 200.
- The 40% tokens-per-megawatt comparison with GPU systems appears in SiliconANGLE; we saw it in Microsoft’s own words only in a search snippet of an X post.
- Microsoft’s blog gives no deployment date.
Numbers that differ
SiliconANGLE says the chips will provide “multiple gigabytes” of computing capacity, which looks like a typo for gigawatts. We do not use that figure.
Labels we chose and why
- Confirmed for specifications in Microsoft’s post.
- Claim for throughput and efficiency comparisons.
- Unconfirmed for the 2027 deployment, seen only in SiliconANGLE.
What we could not verify
- Independent performance tests.
- Pricing or customer availability on Azure.
What we searched
- Microsoft Maia 300 announcement
- Maia 300 azure blog architecture
- Maia 300 tokens per megawatt
Every source we read
| Source | Used? | Why |
|---|---|---|
| Microsoft Azure Blog | USED | Primary source for specs, rack design and comparisons with Maia 200. |
| SiliconANGLE | USED | Reports the 40% claim, the 2027 deployment plan and the Rubin comparison. |
| UPI | NOT USED | Page failed to load. |
| Microsoft on X | NOT USED | Seen only as a search snippet repeating the 40% and 7x claims. |