Microsoft's debt fell 24% during the borrowing binge it keeps getting named in

The five names, and the one that does not fit

On 10 July a widely syndicated analysis put the arithmetic plainly: Alphabet, Amazon, Meta, Microsoft and Oracle, the largest builders of American datacentres, added roughly $350 billion of debt over five years.

Four of those companies borrowed. Microsoft repaid.

Its gross debt was $53.3 billion at the end of fiscal 2021 and $40.3 billion at the end of fiscal 2026 — a fall of $13 billion, or 24%, across exactly the window the analysis covers. Over the same period its annual capital spending went from $20.6 billion to $115.9 billion.

Everyone else moved the other way. Measuring each company from fiscal 2021 to its last completed fiscal year, so that no filer is compared against a partial period:

A note on the arithmetic before anyone reaches for it: this measure totals about $243 billion of additions across the five, against the roughly $350 billion in the published analysis. The definitions differ, and the difference is the subject of this piece. Nobody has miscounted.

The obligation changed instrument

Building $116 billion of datacentre in a year without issuing debt requires the money to arrive some other way. It did, and the record is dated.

In October 2024, Microsoft’s finance leases — largely datacentres, scheduled to commence between fiscal 2025 and fiscal 2030 — stood at $108.4 billion, having risen by nearly $100 billion in two years. A finance lease is an enforceable obligation to pay for an asset over its life. It is not a bond, and a count of bonds will not find it.

In October 2025 Microsoft had committed more than $33 billion to neocloud providers, including a $19.4 billion agreement with Nebius. That is capacity bought from companies who raise the capital, take the construction risk, and carry the borrowing on their own balance sheets.

Neither route is improper. Both are ordinary, disclosed and audited. What they are not is comparable to bond issuance, which is the thing the headline number counts.

The proof that the number tracks classification

The strongest evidence that these figures measure accounting treatment rather than commitment came from Microsoft itself. On 30 July it reported quarterly capital spending up more than 70% year on year, to $41 billion — and in the same release cut its 2026 capital spending forecast from $190 billion to $175 billion because of an accounting change, while stating that its spending plans were unchanged.

Fifteen billion dollars left the forecast. Nothing left the construction schedule.

That is the whole argument in one disclosure. A number that moves by $15 billion without anything being built differently is measuring how obligations are classified.

Whose balance sheet holds it

Ray Dalio makes the point that the aggregate is the wrong instrument for this question: debt bubbles “typically emerge in one or a couple of markets” and “are often hidden beneath the averages.” A bloc figure covering five companies will hide any name that diverges — including one that diverges downward.

But the more useful half is that the obligation still exists, and someone holds it. CoreWeave had no debt at all before 2024 and closed fiscal 2025 with $21.4 billion, of which $6.7 billion falls due within the year against $3.1 billion of cash and $3.1 billion of annual operating cash flow. Cash plus a full year of operations covers those maturities 0.92 times. Among the 23 companies in our filings set carrying a billion dollars or more of near-term debt, it is the only one below 1.0, and the next lowest sits at 1.63.

Graham’s discipline in Security Analysis is to read the obligation rather than the label: a price, a term, an enforceable claim, and a payer who can meet it. Read that way, Microsoft’s five years look less like restraint than like a decision about where the borrowing sits — and the answer, increasingly, is on the balance sheets of its suppliers.

What would show this reading is wrong

Microsoft may simply be the strongest balance sheet in the group, funding from operations because it can: $182.9 billion of operating cash flow in fiscal 2026 covers a great deal without borrowing. That is a complete and unglamorous explanation, and it may be the right one.

Finance leases are also disclosed, not hidden. Every figure here comes from a filing, and any analyst comparing bond issuance across five companies knows what a lease is.

And nothing here shows distress. CoreWeave’s maturities are a concentration, not a default.

Three things in filings would settle it within two quarters: Microsoft’s finance-lease balance flattening while capex holds; its neocloud commitments not growing; or CoreWeave refinancing its 2027 maturities on terms no worse than the ones it has.

Sources

Balance-sheet and cash-flow figures are SEC EDGAR XBRL company facts, annual periods only. Fiscal years differ across these filers — Microsoft’s ends 30 June, Oracle’s 31 May, the rest 31 December — so each company is measured from its own fiscal 2021 to its own last completed fiscal year, and no filer is compared against a partial period. Gross debt is long-term plus current debt, which excludes finance-lease obligations; that exclusion is the piece’s subject rather than an oversight.

Dated events come from the coverage record and are dated by first coverage.