The AI build-out, read from the filings

They report $712B of debt. They have promised $1.48T.

"Reported debt" is what sits on the balance sheet as borrowings. But filings also commit money through leases and signed purchase deals. Add those and the promised total is 108% larger than the debt line alone. Committed obligations were 2.1x reported debt, showing that the broader promise pile is more than twice the debt line. Reported debt was $712.4 billion, while total committed obligations reached $1.48 trillion.

Largest 14 by total promised, latest quarter (9 smaller companies not shown)

Industry total over time

Money earned, money collected early, money merely signed

Deferred revenue is cash collected for work not yet delivered — it is not earned yet. Signed backlog is orders promised but not yet delivered or paid; it overlaps deferred, so we never add the two. Backlog now stands at $2.43T against $2.80T of trailing-year revenue. Backlog alone stood at 0.9x one year of revenue. Deferred revenue added $260.0 billion, putting the revenue already earned or promised in the filings against trailing-year revenue of $2.80 trillion.

Deferred revenue as a share of yearly revenue — the median company

What they spend on, what they earn from

Four lines carry almost all reported spending: making the product, inventing the next one (R&D), selling and running the company, and tax. Stock-based pay — $117B in 2025 — hides inside those lines, so we show it beside them, never as a fifth slice. Cloud & platforms generated $1.86 trillion in trailing-year revenue, far ahead of chips at $613.5 billion and hardware & networking at $233.9 billion. In 2025, cost of revenue was $974.4 billion, the largest listed expense.

Share of reported operating spend, by year

The cash story the expense lines miss: building vs inventing

Not every company tags every line; each sum covers only the companies that tag it (always more than half the industry).

Trailing-year revenue by group

One honest gap: filings publish per-segment revenue in a tagged data layer this pipeline does not parse yet, so this paper groups revenue by what each company builds, not by each company's own segment names.

$1.53T comes due in five years

Every dated obligation in the filings, stacked by when it falls due. Dark blue money has a filed payment schedule. Light blue money has a due date but no schedule, so its timing is a labeled straight-line estimate. Gray money is promised for "after year five" with no date at all — we refuse to invent one. The next five years carry $1.53 trillion of obligations against $266.6 billion of future revenue promised for that period. Filings show the due amount was 5.8x the promised revenue.

Against it: revenue already promised to them (backlog with dates)

The contrast is the finding: companies schedule spending decades out, but dated revenue promises barely reach past five years. Filings also show $155B of financing flowing back — money lent out that should return. We keep it out of the wall; a promise to receive is not a promise to pay.

Promises are growing faster than revenue

Each line starts at 100 in 2021-Q1 so the slopes compare fairly. Total committed obligations grew 23% per year, faster than revenue at 14% per year. Reported debt and deferred revenue each grew 14% per year.

How often the filed numbers moved

Our reconciler re-states each quarter using only evidence cited from the company's own filings: revenue that is really conditional moves out, spending that is really committed moves over. It changed 217 of 806 company-quarters. Reconciliation changed the as-filed numbers in 217 of 806 quarters. Extraction disagreements led the listed flag types with 177 cases.

Industry expenses: as filed vs after cited reconciliation, by year

554 flags across 34 companies

A flag is a pattern worth a second look, not an accusation: a claim that appears once and vanishes, a metric that quietly changes definition, financing that circles back as revenue. Every flag cites the exact filing text.

How this was made

Code owns every number: exact figures come from SEC's machine-readable XBRL data, and all sums, growth rates, and date math are computed, never estimated by a model. AI reads the narrative text — commitments buried in notes, related-party deals, guarantees — and every claim it catalogs cites its source filing.

Limits worth knowing: industry stats cover the 31 US-dollar filers only; ASML (euros), TSMC (Taiwan dollars) and Nebius's earlier ruble filings stay out of the averages. Money moved through third parties without disclosure is invisible to filings, so the ratio flags are the proxy. Dollar charts sum the industry; ratio charts use the median company so no giant can hide the typical story.

The same questions, one company at a time

AI software

Chip equipment

Chips

Cloud & platforms

Hardware & networking

Data centers

Aerospace