Microsoft plans million-chip Maia 300 deployment to cut AI costs

Reuters reported that Microsoft plans multi-gigawatt deployments of Maia 300 for its internal AI services. Each rack holds 96 processors and draws about 167 kilowatts; the central compute chiplets use TSMC’s 2-nanometre process, while other chiplets use 3-nanometre technology.

Microsoft says the system can deliver about 40% more useful computing per megawatt than leading GPU-based systems, but it did not identify the comparison hardware. The scale and efficiency claims therefore remain company benchmarks until independent workload results emerge.