Microsoft unveils Maia 300, its own AI chip, and plans multi-gigawatt rollouts from 2027
Microsoft's third-generation Maia chip is built to run AI models, not train them. Microsoft says it will become the main system behind its own AI products, starting in 2027.
By ScoviAI Team
Key points
- Maia 300 was announced on October 12, 2026 as the third generation of Microsoft's AI chip
- It's designed mainly for inference: running AI models for users
- Microsoft expects 40% more tokens per megawatt than leading GPU systems
- Production deployments start in 2027, at multi-gigawatt scale
Microsoft has announced Maia 300, the third generation of its own AI chip. It's designed for inference, which means running AI models every time someone asks Copilot a question, rather than training new models.
Microsoft says Maia will become the primary inference system for its own products and services, with multi-gigawatt deployments starting in 2027.
Why Microsoft is building its own chips
Running AI for hundreds of millions of people is expensive, and most of that cost is electricity and hardware. Today much of the industry depends on GPUs from Nvidia. A chip designed in-house lets Microsoft tune everything, from the chip to the rack and the cooling, for its own workloads.
"We expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems." — Microsoft
"Tokens per megawatt" is the useful measure here: how much AI output Microsoft gets for the power it pays for.
Maia 300 by the numbers
| Maia 300 (Microsoft's figures) | |
|---|---|
| Throughput | Up to 7x the rack-scale token throughput of Maia 200 |
| Rack | 96 accelerators in a 167-kilowatt power envelope |
| Memory per rack | 27.6 TB of HBM, 1.7 PB/s total bandwidth |
| Per chip | 8 HBM4 stacks, 18 TB/s peak memory bandwidth, 950-watt TDP |
| Manufacturing | Multi-chiplet design, compute chiplets on TSMC's N2P process |
These are Microsoft's own projections. Independent tests will only be possible once the systems are running.
What it means for the AI business
The big cloud companies are all trying to depend less on outside chip suppliers. With Maia 300, Microsoft joins Google and Amazon in betting that custom chips can make AI cheaper to run at huge scale. If the efficiency gains hold up, they matter most for services that run constantly, like Copilot in Microsoft 365.
What it means for you
Nothing changes for users today; deployments start in 2027. Over time, cheaper inference is one of the things that decides whether AI features stay included in the plans you already pay for. Our Microsoft 365 Copilot page lists the current plans and offers.
Sources
ScoviAI Team is the ScoviAI newsroom. We write every article ourselves from the original announcements and link our sources. Spotted a mistake? Tell us and we'll fix it.
Related Topics
ADVERTISEMENT
