AI capex and required revenues (Goldman Sachs, September 2026): hyperscalers need $300B a year to break even

Goldman Sachs Research answers the three questions clients ask most about the AI trade. The largest US hyperscalers are on track for roughly $800 billion of capex in 2026, consensus sees $1.1 trillion in 2027, and Goldman's analysts go higher at $1.2 trillion for 2027 and $1.4 trillion for 2028, while growth slows from near 100% this year to 54% and then 12%. Capex in 2027 would be a bigger share of GDP than any technology cycle since the railroads. The note then works backward from that spending. Hyperscalers need about $300 billion of annual AI revenue to break even, and end users would need to spend roughly $1 trillion a year on AI applications for both the infrastructure and application layers to earn healthy margins, against about $1.5 trillion of total global software spending today. For founders, this is the clearest public math on how much application-layer revenue the buildout needs. The market is pricing skepticism, especially in chips and memory, so companies that can show real revenue against compute costs will stand out.