Microsoft plans multi-gigawatt deployments of its custom Maia 300 AI chip racks starting next year. Each rack houses 96 chips and draws approximately 167 kilowatts of power.
At that scale, electricity becomes a line-item competitor. The economics are straightforward: custom silicon reduces reliance on external GPU suppliers, lowers per-unit hardware costs, and optimizes performance-per-watt within Azure. That translates directly to gross margin on AI services.
The bet requires enormous capital allocation. Beyond chip design and manufacturing, Microsoft must expand data centers, build advanced cooling infrastructure, and upgrade power grids to handle gigawatt-scale loads. It is expensive. It is also defensible—a moat that off-the-shelf accelerator buyers cannot easily replicate.
AWS and Google have similar programs: Trainium and Inferentia chips for Amazon, TPUs for Google. Each player is betting that controlling silicon means controlling the economics of the AI arms race. Microsoft's multi-gigawatt commitment signals it is playing for long-term share in cloud AI inference and training, where margin compression from commoditized GPUs is a real risk.
The 167-kilowatt power envelope per rack highlights the real constraint: not chip performance but power and cooling at hyperscale. Companies that solve that equation efficiently win the margin game.
