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

Microsoft plans to transition its internal AI inference infrastructure to the Maia accelerator family starting in 2027, with deployments expected to reach the multi-gigawatt scale. The upcoming Maia 300 architecture aims to streamline data movement by unifying scale-up and scale-out processes on a single Ethernet fabric. This shift reflects a strategic move by Microsoft to reduce its reliance on third-party hardware for running its proprietary AI services. However, industry experts are currently questioning the specific efficiency metrics being used to evaluate these custom accelerators.