Microsoft Commits Maia to Multi-Gigawatt Inference, and Its Efficiency Metric Draws Questions

▤ Summary

Microsoft is expanding its Maia AI accelerator family for multi-gigawatt inference deployments starting in 2027. The company says Maia 200 is already in production and powers internal services such as Copilot, with more Microsoft products scheduled to follow. The announcement focuses on the Maia 300 architecture, its memory-heavy design, and its custom Ethernet-based AI fabric. Microsoft also claims large efficiency gains, but those claims have drawn skepticism from practitioners who question the token-per-watt metric and the workload assumptions behind it. The posts do not say whether external customers will get direct access to Maia capacity.

▦ Linked Companies

Microsoft

$1B+