What do AI models think these companies are worth?

Forecasts vs reality, by knowledge-cutoff vintage

Median forecast across all models sharing a cutoff year (3 samples per model per question, log scale, USD billions). Targets a model could already “know” are excluded.

The flaw in point estimates: the downside mass they hide

Asked for a single “best estimate” of a 2030 valuation, models draw a smooth success path. Asked for their full distribution (anchored on the same current valuation), the same models place large probability on decline. The band spans each model’s 10th–90th percentile with a dot at the median; the green marker is the same model’s point estimate from the main experiment. Note the green markers sit below the medians for a reason worth knowing: the main experiment gave no anchor, so models forecast from their (stale, lower) training-data knowledge of the valuation — handed the true current number, their median forecast implies only ~6–10%/yr growth, close to a required rate of return. Name-blinded twins test whether famous names get rosier odds (mostly: no).

P(worth less in 2030 than today)

Each model’s stated probability that the company’s value on Jan 1, 2030 is below its current valuation. A fairly-priced, venture-volatility asset sits around 35–45% (shaded). Point estimates carry none of this.

Hindsight calibration: where did reality land in the models’ own distributions?

For past targets (Jan 2024 → Jul 2026) we elicited full p10/p50/p90 distributions — unanchored, so each model forecasts from its own knowledge — and checked where the realized valuation fell. A calibrated forecaster puts ~10% of outcomes above its 90th percentile. The observed rate is the red segment.

Per-model forecasts

Every model’s own median trajectory, colored by family. Hover for details.

How the belief evolved with the cutoff

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.