The $1.65tn 'hidden debt' behind Big Tech's AI buildout
A July 2026 Nikkei investigation set off a heated debate about how Big Tech is really paying for the AI boom. It estimates that five of the largest US tech companies — Alphabet, Microsoft, Amazon, Meta and OracleDealroom has a profile for this one. Try Dealroom → — have racked up roughly $1.65 trillion of future obligations that do not appear as debt on their balance sheets. That figure is larger than the ~$1.35 trillion of borrowing they report openly, and it has grown around eightfold since 2022. The number went viral and drew sharp pushback — including on whether it should be called "debt" at all (see below).
What does "off-balance-sheet" actually mean? When a company borrows money the normal way, the loan shows up as debt on its books, so investors can see it. But there are other ways to commit to paying large sums in future that accounting rules do not require to be listed as debt today — chiefly long-term rental (lease) contracts for data centres, multi-year commitments to buy chips and computing capacity, and joint ventures set up with outside investors. The obligation to pay is real; it just sits in the footnotes rather than the headline debt figure.
Why it is happening now. Building the data centres and buying the chips (GPUs) needed for AI is enormously expensive. Rather than put all of that borrowing directly on their own balance sheets, the companies increasingly sign long leases, lock in supply contracts, and partner with private-credit funds — arrangements that spread the cost off the books. Meta is the clearest example, with an estimated $420 billion off-balance-sheet (nearly triple its reported debt), partly through a $27 billion joint venture with Blue Owl to fund its Hyperion campus in Louisiana. Oracle's hidden obligations have grown fastest — an estimated $273 billion — and its overall leverage prompted S&P Global to cut its rating to the lowest investment-grade tier. Even Alphabet, long seen as having the cleanest balance sheet, now discloses tens of billions in such commitments, including a joint venture with Blackstone.
The concern. Critics — including the Bank for International Settlements, which calls it "shadow borrowing" — argue these structures make a company's true leverage hard for investors, rating agencies and regulators to see in real time. Moody's has flagged that many of the leases are "pre-operational": the company is committed to paying for capacity before it earns any revenue from it. Nikkei draws a parallel with the early-2000s telecoms bust, where equipment makers financed the very customers buying their gear — a circular structure that held up only while demand kept outrunning the debt taken on to build ahead of it. Morgan Stanley puts industry-wide off-balance-sheet AI exposure even higher, at around $1.8 trillion.
The other side — "a purchase commitment is not debt." A vocal camp argues the framing is overblown. SemiAnalysis founder Dylan Patel put it bluntly: "A contract to buy something is not debt. It's a contract to buy something — they have not received the good or service yet." On this view a multi-year commitment to buy chips or rent capacity is an ordinary operating obligation, not borrowed money: no principal has been lent, no interest accrues, and much of it can be tied to future revenue rather than funding a present-day cash hole. Critics of the study also note the accounting is entirely legal and disclosed — the commitments sit in the filings' footnotes exactly as the rules require — and that leases and supply contracts are a normal, long-established way to finance capital-heavy expansion. The companies involved are among the most cash-generative and profitable in the world, with large cash balances and investment-grade credit, so servicing these obligations looks very different from a stretched startup taking on debt. On this reading the $1.65tn reflects the sheer scale of the AI buildout — and some loose labelling — rather than a looming solvency problem.
The bottom line. Both sides agree the obligations are real and disclosed; they disagree on what to call them and how much they matter. The substantive questions are whether standard debt metrics still capture the true risk, and whether monetised AI demand will keep pace with the capacity being financed ahead of it.
Read more: Nikkei Asia · Seeking Alpha · Dylan Patel (X)
