Microsoft disclosed the Maia 300 architecture and a 2027 production-deployment plan; performance and efficiency figures are vendor-reported.
Microsoft details Maia 300 before production deployments begin
2026-10-12Verified announcementWATCH
Microsoft disclosed Maia 300, a custom inference accelerator and rack system planned for production deployment from 2027. The company says a 96-accelerator rack should deliver 40% more tokens per megawatt than leading GPU systems and up to seven times Maia 200 rack throughput. These are vendor projections; production telemetry and independent comparisons are not yet available.
Official · AnnouncementMicrosoft to expand Azure Maia system to multi-gigawatt deploymentsOfficial · DocumentationDeep dive into Azure Maia 300 architecture
Availability: production deployments are planned to start in 2027; not currently deployed
What this means for you
For inference-platform teams, Maia 300 signals that hyperscalers are optimizing the whole rack around tokens per watt, not only chip peak performance. It is not deployable today, so procurement and capacity decisions should wait for workload-level throughput, latency, reliability and availability evidence.
Evidence & related material
Technical details & scores
| Availability | production deployments are planned to start in 2027; not currently deployed |
|---|---|
| Rack | 96 accelerators, 167 kW, 27.6 TB HBM and 1.7 PB/s aggregate memory bandwidth |
| Accelerator | TSMC N2P compute chiplets, eight HBM4 stacks, 288 GB HBM, 18 TB/s and 950 W |
| Vendor Claims | 40% more tokens per megawatt than leading GPU systems; up to 7x Maia 200 rack throughput and more than 2.5x performance per watt |
| Numerics | MXFP4 and formats below three bits |
| Time Basis | official pages state 2026-10-12 without a time of day; inclusion follows the next complete daily window and does not claim hourly precision |
Technical impact 4
Engineering value 3
Adoption signal 1
Maturity 1
Verification cost 5
Risk 4