September 29, 2026
AI Buildout Costs More Than Railroads: 3.63% of US GDP per Year
Brookings' Sep 23 scenario projects $10.3 trillion in investment in US AI infrastructure over 2025–2032: 3.63% of GDP on average per year. US railroad capital expenditures accounted for 2.24% of GDP in 1870–1890. The central scenario includes 182.8 GW coming online by 2032, another 117.2 GW afterward, and the cancellation of 226.9 GW of announced projects.
Ethan Mollick
@emollick
As a share of GDP, today’s AI infrastructure build-out is larger than railroads at their peak, even though AI is nowhere near as consequential to the economy as railroads were. In 1890, one in every twelve men in the US worked for railroads, and railroads themselves consumed 56% of all machine power!
· 6.6K views
The author estimates a typical 200 MW AI training campus at $8.216 billion. IT equipment accounts for $5.616 billion, while the data center and additional power infrastructure account for $2.6 billion. Capex at Oracle, Microsoft, Amazon, Meta and Alphabet rose from $97 billion in 2020 to more than $400 billion in 2025. For 2026, the scenario projects more than $800 billion, exceeding their combined operating cash flow for the first time.
The revenue still has to be earned. For a 10% return at a 50% operating cash flow margin, the central scenario requires $3.7 trillion in annual revenue by 2032. That is $5.5 per installed GB300 GPU-hour at full utilization. Meta is building the 2.0 GW Hyperion for about $30 billion: Blue Owl received 80% equity for $2.5 billion, while the Beignet structure raised $27 billion in debt.
In July 2026, the Federal Reserve found limited evidence of AI having a broad economic impact: its effect remains concentrated in specific sectors. The “1 in 12” figure from the available historical report refers to injuries among trainmen in the year ending June 30, 1889, not employment: 704,743 people worked in the US railroad system.
By 2032, it will be clear whether the 182.8 GW in the central scenario turns into revenue that pays for this build-out.
Source
