America is attempting the largest infrastructure buildout in its modern history — and the bill is starting to strain both the power grid and the balance sheet. A 2026 paper by Columbia Business School’s Stijn van Nieuwerburgh puts US AI infrastructure investment at $10.3 trillion between 2025 and 2032, or about 3.63 per cent of GDP annually. For comparison, the railroad buildout of the late 19th century peaked at roughly 2.2 per cent of GDP. The AI boom is already the biggest capital project the country has ever attempted, and it is running into hard limits.
The money is concentrated. The five largest hyperscalers — Microsoft, Amazon, Meta, Alphabet and Oracle — are projected to spend a combined $725 billion to $750 billion on capital expenditure in 2026 alone, up about 77 per cent year over year. That spending is on track to exceed their operating cash flow, meaning even the richest companies on Earth must fund the buildout with external capital. Combined free cash flow for the group has collapsed from roughly $61 billion quarterly in 2024 to under $5 billion by mid-2026, by one reconstruction from company filings. The hyperscalers now hold more debt than cash for the first time, having issued over $121 billion in bonds in 2025 alone.
The grid is the bottleneck
Then there is physics. Google has cited expected grid connection delays of five to ten years, driven by data-centre demand, with pressure acute in regional grids such as PJM and ERCOT. The research flags a potential power shortfall of up to 68 GW for data centres. Microsoft reportedly holds $80 billion in unfulfilled Azure orders — not from weak demand, but because it cannot find enough electricity to power GPUs already sitting in warehouses. Power transformer lead times have stretched to 128 weeks.
The revenue gap
The arithmetic that worries markets: the industry would need roughly $3.7 trillion in annual revenue by 2032 to support the projected investment, against current revenue estimates of around $100 billion. The IMF has warned of potential stock price corrections if profitability on AI investment trends downward. The buildout may still be justified — vacancy is near 1.4 per cent, GPUs have 12-to-18-month lead times — but the era of funding it painlessly from cash flow is over.
