Microsoft wants Maia to deliver 40% more tokens per megawatt than leading GPU systems. But here's the catch: 40% more tokens doesn't mean 40% lower costs. If useful output rises 40% at the same power, power consumption per token falls roughly 28.6%. That's meaningful. But electricity is only one part of inference costs. Hardware, memory, software and utilization still matter. And if Microsoft passes some savings to Azure customers, it won't keep all the benefit. Maia 200 is already serving production workloads. Maia 300 deployment begins in 2027, with Microsoft planning to make its own chips the primary inference system for its products and services. @AryaTschand's architectural analysis raises an important point: reducing unnecessary computation and data movement can make specialized inference chips attractive, even while GPUs remain competitive for other workloads. The financial test comes later. Microsoft generated $19.6B in quarterly free cash flow, but its cloud gross margin fell to 65%. That isn't a Maia specific margin. It's a reminder that efficiency gains can coexist with rising infrastructure costs. The real question isn't how many more tokens Maia can produce. It's how much of the resulting savings Microsoft can actually keep.