Wall Street Journal analysis finds $3 trillion in off-balance-sheet AI commitments

A Wall Street Journal analysis of footnotes in securities filings found that nine of the largest technology and AI-infrastructure companies together carry roughly $3 trillion in commitments tied to the AI buildout that do not yet appear on their balance sheets — about $1.9 trillion in purchase obligations for chips, servers and other equipment, and $1.2 trillion in leases for data-centre capacity that has been signed but not yet begun. The Journal compared that to the roughly $600 billion the same companies reported in capital expenditure over the trailing twelve months, putting the unbooked commitments at roughly five times reported spending.

Under current accounting rules, a lease that has not yet commenced and a purchase order not yet fulfilled do not count as balance-sheet liabilities, so conventional measures such as debt-to-equity understate what companies have actually promised to pay. Among the nine — which the Journal reported also included Alphabet, Broadcom and others alongside Amazon, Nvidia, Meta, Microsoft and Oracle — Alphabet disclosed the largest single figure, $811 billion in purchase and contractual obligations as of 30 June, up sharply from three months earlier. Meta’s disclosed obligations were reported at $347 billion in one account and $420 billion in another, depending on which commitments were aggregated; Oracle’s were put at about $273 billion, which the Journal said had grown more than thirtyfold in four years. Nvidia’s exposure was described differently again, as a contingent liability arising from the compute-financing partnerships it had announced the previous week rather than a direct purchase or lease commitment. Individual figures for Microsoft and Amazon were not detailed in the coverage available.

The analysis landed amid broader investor scrutiny of whether AI infrastructure spending is outrunning revenue growth across the sector, and added a specific, if partial, answer to a question raised repeatedly through 2026: how much of the AI buildout’s true cost is visible in the numbers companies actually report.