We asked 26 AI models — with knowledge cutoffs from September 2021 to February 2026, and no internet access — what OpenAI, Anthropic, NVIDIA, Alphabet and Meta would be worth on each January 1 from 2024 to 2030. The story in one line: the models forecast like a finance textbook — sensible growth rates, the right distribution shapes — and the AI boom landed in the tail of their own distributions (on dates that have since happened, the real valuation landed above their stated 90th percentile 69% of the time — though since all the misses share one correlated draw, one AI boom, that's an outlier by their own power-law beliefs, not proof of miscalibration).
github.com/TrelisResearch/llm-valuation-forecasts · by @ronankmcgovern
Each blue line is a “vintage” of models grouped by the year their knowledge ends (darker = knows more recent history). Red is reality. Every vintage revised upward as its knowledge advanced — and reality still escaped above the lot. Log scale, USD billions; dates a model could already know are excluded.
Convert each vintage’s forecast curve into an annual growth rate and a clean pattern appears. For the public companies: ~8–10%/yr — the required rate of return, since a current market price already contains everything predictable. For the startups the models project growth that fades with size: ~30%/yr in the early forecast years, ~20%/yr further out, and — when handed a ~$1T current valuation in a follow-up probe — the same ~6–10%/yr they give the mega-caps. In other words, they price startups on a decaying convergence path toward big-company returns, venture-like rates while “small.” Reality (red) grew several times faster than any point on that path.
We asked each model: what’s the chance the company is worth 1.5×, 2×, 5×… 100× today’s value in 2030? Plotted on log-log axes, a straight line means a power law. For Anthropic the models draw near-straight lines with tail exponent α ≈ 1.4–2.0 — the same range measured for real venture returns. For NVIDIA the curves bend downward: thinner, lognormal-like tails, the standard model for public equities. Hover any point for exact values.
Every model’s own median trajectory, colored by family. Hover for details.
Fix a target date; each dot is one model’s median forecast for it, placed at that model’s knowledge cutoff. The line joins cutoff-cohort medians — belief revision in slow motion.