The $820 Billion That Isn't on the Balance Sheet

Look up the leverage ratios of the five companies building the AI economy and you will conclude they are barely borrowing at all. Four of the five sit at total debt to EBITDA of around one times or less. By any conventional credit screen, these are among the safest borrowers alive.

Now look at what isn't in that ratio.

FactSet's July analysis of Alphabet, Amazon, Meta, Microsoft and Oracle put aggregate FY26 capital expenditure above $690 billion — more than 80% growth in a single year. Free cash flow for the group is expected to land near zero or negative for everyone except Alphabet and Microsoft. Incremental annual debt has gone from 9% of capex in FY24 to 32% by mid-2026. Alphabet priced an $84.75 billion equity raise in June. Oracle was cut by S&P on 9 July to BBB−, one notch above speculative grade.

And then the number that actually stopped me: roughly $820 billion of lease-related commitments across the five that are not recognised as balance-sheet liabilities. Oracle alone has signed data-centre leases of fifteen to nineteen years totalling nearly three times its FY27 capex guidance. Meta's Hyperion campus in Louisiana sits inside a joint venture where Blue Owl's funds hold 80% and Meta 20%, with Meta leasing the finished asset back. The obligation is real, contractual and long-dated. It just doesn't land where debt/EBITDA can see it.

This is not fraud, and I want to be precise about that. Every one of these structures is disclosed, GAAP-compliant, and reviewed by auditors who are better at this than I am. Nobody is hiding anything. The problem is subtler and, in my experience, far more dangerous: disclosed is not the same as visible. The commitment is in the footnotes. The ratio is in the model. The model is what gets priced.

In fifteen years of building commercial credit software in the US, the most common failure I saw was never a missing document. It was a covenant, a ratio, a screen — all working exactly as designed, all pointed at the part of the obligation that had been moved somewhere else. Give a treasury team a metric and enough time, and the metric stops describing the company and starts describing the structure built to satisfy it.

The bull case deserves a fair hearing, and it is a good one. These are extraordinary businesses with real demand for the capacity they are building; contracted GPU leasing rates and second-hand chip prices are holding up, which argues the assets are worth what's being paid. The SPV genuinely transfers risk to investors who are paid to hold it. And unlike the 2001 telecom build-out, there is revenue arriving today, not a slide deck promising it.

But risk that has been transferred has not been destroyed. It has moved into private credit — the corner of the market with the least mark-to-market discipline and the longest lag between deterioration and disclosure. Meanwhile the capex mix keeps shifting toward compute, now around 60% of the total against 43% in 2022. Those are three-to-five-year assets being financed with fifteen-year obligations. That mismatch is the oldest failure mode in lending, and no amount of investment-grade paper repeals it.

I see the same mechanism at the other end of the market. In Indian MSME lending, the borrower whose ratios look immaculate is very often the borrower whose real obligation sits with a related party, a personal guarantee or an unrecorded supplier credit. Different zeros. Identical instinct.

The work this quarter isn't forecasting whether AI capex is too big. It's auditing your own instrument — finding the obligations your model is structurally unable to see, and asking who decided they didn't count.

Because if the safest borrowers in the world are financing the largest build-out in history through vehicles your leverage ratio ignores, what exactly is that ratio still telling you?