Rory O’Driscoll: Software Not Dead but Far Harder to Win
SaaStr AI session outlines $688B AI capex against $110B revenue and a 2031-2032 crossover point for foundation model leaders.
AI Capex Outpaces Revenue by Hundreds of Billions
Rory O’Driscoll told attendees at SaaStr AI that software is not dead. He said it has become much harder to win after reviewing spending and revenue numbers compiled by his firm. In 2026 hyperscalers are projected to spend roughly $688B on AI capex while the same market is expected to generate about $110B in revenue, according to SaaStr. Of that revenue, $89B is attributed to the two leading foundation model companies on a GAAP basis, with the balance estimated at $20-30B.
Revenue Crossover Projected for 2031-2032
Running the same estimates forward shows the two model leaders reaching cumulative revenue parity with capex near $1T around 2031-2032. O’Driscoll noted this leaves five or six years of continued capital deployment ahead of returns. He also flagged the possibility of a market pullback before that point if spend remains far above revenue.
Knowledge Worker Wages as the Target Pool
O’Driscoll stated that the trillion-dollar revenue target requires AI to capture a material share of the knowledge worker wage bill. If all spend occurred in the United States, the figure would represent 15-17% of total knowledge worker compensation. For software developers the share would exceed 25%, equating to roughly $50,000 in token spend for every $200,000 developer salary.
Making AI versus Using AI
O’Driscoll described the AI stack in layers: energy, chips, infra, models, and apps. The bottom three layers account for the $688B capex and are roughly 80% not venture-backable, with half of that spend directed to chips, primarily Nvidia. Capex has risen from about $200B four years ago to about $600B today. The top layers, where models are turned into customer value, are the domain of application software companies.
The Harness Layer
O’Driscoll described a software layer above raw models that manages context, actions, output selection, logging, escalation, and model routing. He compared the emerging pattern to the LAMP stack that supported fifteen years of differentiated B2B applications, according to SaaStr.
