Bayes and Base Rates: How History Can Guide Our Assessment of the Future
Anchor AI revenue forecasts to historical base rates, then adjust only as far as hard evidence such as adoption data justifies. Big capacity announcements may also be preemptive deterrence rather than demand signals.
Summary
The report applies Bayes' theorem to AI growth forecasts, starting from base rates for companies of similar size. OpenAI's plan to go from $3.7 billion (2024) to $145 billion (2029), a 108% CAGR, has no precedent among U.S. public companies with $2-5 billion in sales since 1950 (base-rate mean 7.0%, SD 10.6%); Oracle's cloud plan ($10 billion to $166 billion by fiscal 2030, 75% CAGR) is likewise unprecedented. Offsetting evidence includes ChatGPT's fast adoption, but Flyvbjerg's database of 16,000 large projects shows fewer than 9% finish on time and on budget. The authors also read infrastructure announcements as a possible Porter-style preemptive capacity strategy.
Key concepts
Bayes' theorem · Base rates · Reference class forecasting · Superforecasting · Flyvbjerg's think slow, act fast · Preemptive capacity commitments · Total addressable market