Microsoft to Deploy Multi-Gigawatt Maia Chips Across Its Data Centers

Microsoft's decision to keep Maia compute off the rental market is the more consequential detail: it treats inference capacity as an internal cost advantage rather than a merchant cloud product, potentially tightening external supply of accelerator capacity even as hyperscaler demand grows. The efficiency claim, if it holds in production, shifts the competitive axis from raw chip availability toward tokens per megawatt, a metric that favors firms with both silicon and power procurement. Whether rivals respond with comparable in-house inference silicon, or whether Microsoft later opens Maia to customers, is the open question.