The ramp vs. reality: scoring the decade's most aggressive AI forecast

Most AI punditry is unfalsifiable. Leopold Aschenbrenner's Situational Awareness (June 2024) is the exception: it made dated, quantified forecasts — which means it can be scored. We scored the infrastructure half against our own primary-source data.

~$500B of annual AI investment by 2026. Our measured floor: the five hyperscalers' total capex alone reached $482B in the trailing twelve months through 2026Q1, with the latest quarter annualizing to $594B — and that counts none of the frontier labs, neoclouds, or sovereign projects his all-in figure includes. On trajectory, the narrow measure alone is crossing his number on schedule. Tracking–beaten.

Power becomes the binding constraint. Qualitative, but broadly corroborated by two years of reporting — grid interconnection queues, gas-turbine backlogs, datacenter siting fights. Our own power data (a utility output index) is still too coarse to test it properly; the planned gigawatt tracker is the right instrument. Tracking, on secondary evidence.

Gigawatt-scale single training clusters by 2026. Multiple campuses of roughly that class are reported under construction or energizing. Not yet in our measured data — same gap, same fix. Tracking, on secondary evidence.

The open bets. ~10 GW clusters by 2028; ~$8T of annual AI investment by 2030; AI at tens of percent of US electricity; AGI — "a drop-in AI researcher" — called strikingly plausible by ~2027. These are where the essay lives or dies, and none have resolved. Two honest cautions from scoring the closed half: the calls he's winning are the ones a capex boom makes near-tautological, and a forecast beaten on infrastructure says nothing yet about the capability timeline that justifies the infrastructure.

Why keep scoring him at all? Because a dated forecast that keeps landing is information — it tells you which scenario's priors deserve weight. And because the discipline cuts both ways: when the curve and reality diverge, a scored record is how you notice early. The chart on the Capital board overlays his investment curve on our measured capex line — one definition caveat applies (his figures are all-in AI investment; ours is five companies' total capex), so read trajectory, not identity.