The money pouring into the AI build-out, set against a century of humanity's largest bets — measured as a share of the economy that had to pay for each one.
~5% of US GDP by 2027
~$3T poured in since ChatGPT
No income yet in sight to cover it
Peak annual spend · share of US GDP
One year of AI capex outweighs a century of moonshots
0.5%1.0%1.5%2.0%
AI data-center capex 2025
~1.9% of GDP
~1.9% in 2025 (Kobeissi; stricter 5-company tally ≈ 1.3%) · heading to ~5% of US GDP by 2027 (RBA)
Nationwide broadband early 2000s
1.2% of GDP
Apollo program 1961–72
0.6% of GDP
Interstate Highway System 1956–70s
0.6% of GDP
Rural electrification 1949
0.6% of GDP
Manhattan Project 1942–45
0.4% of GDP
Off the chart
The one bet on record that was bigger: the 1880s railroad build, around 6% of GDP — but of an economy a small fraction of today's size. It's the honest ceiling, and the last time capital rushed this far ahead of the returns. AI's 2025 spend is ~1.9% of a vastly larger economy, but on the RBA's read it races toward ~5% by 2027 — closing on that ceiling. Kedrosky
The value of that money, in today's terms
Apollo cost $309 billion in today's money. At the 2026 pace, the AI build-out spends one entire Apollo program every eleven weeks.
The roughly $3 trillion poured in since ChatGPT launched is about ten Apollo programs, compressed into three years.
Every bet on this chart paid for something we still use — the highways, the grid, the moon landing. This is the first one where the income to cover it isn't yet in sight.