In the current AI build-out, ambition is free and capacity is rationed. Every hyperscaler has the same order-of-magnitude appetite for custom silicon, and the only thing separating a roadmap from a shipment is what TSMC, HBM, and advanced packaging can be made to deliver. Maia 300 is the cleanest illustration of the pattern.
Microsoft's stated ceiling, per the Reuters report on The Information's reporting, is more than 1 million Maia 300 chips. The 2027 plan is 300,000 units, and the report's own qualifier, that component supplies and ongoing TSMC capacity negotiations could constrain the ramp, does the actual explaining. Most readers will read "more than 1 million" as a confirmed order and treat Maia 300 as Microsoft's answer to Nvidia. The pattern is the opposite.
For Microsoft's custom-silicon program, the binding constraint is TSMC capacity and component supply — not how many chips the customer wants to buy. The headline scale lives in the cloud; the binding constraint lives on a fab floor.
The corporate shape follows. Microsoft is recruiting cloud customers like Anthropic because the only way to underwrite a custom-silicon program today is to spread a fixed manufacturing envelope across as much committed demand as possible. The 1-million figure is the aspiration. The 300,000 figure is the actual bet. Read the second one when sizing the company.
Reported by Sky for Type0, from Microsoft plans to unveil next-generation AI chip in September, The Information reports. Read the original: thestar.com.my