The obvious question is how debt can exist without being counted as debt. The
short answer: this is not a trick, it is an accounting rule.
These are signed contracts — data centers, GPUs, servers, decades-long leases —
whose obligations are not yet recognized on the balance sheet because the
facilities are not operational. It is entirely legal. It also means that anyone
reading only the financial statements is not seeing the full picture.
Where these commitments come from
All of it stems from the race to build the infrastructure artificial
intelligence needs to run. That infrastructure is not bought on a single
invoice: it is contracted for years, sometimes decades.
- Data centers leased before they are built.
- GPU contracts committed in advance.
- Servers and related equipment.
- Long-term leases signed for decades.
Until those facilities go live, much of that commitment sits off the balance
sheet. It is disclosed — nothing is hidden — but in the footnotes to the
financial statements, not on the debt line that headlines and most data screens
pick up.
The two cases that stand out
| Company | What the investigation found |
|---|
| Meta |
Around $420 billion in obligations of this kind |
| Oracle |
Multiplied these commitments more than 30 times in four years |
Meta, Microsoft, Amazon,
Alphabet and Oracle |
Combined, about $1.65 trillion, above their reported debt |
Why this does not stay off the books forever
This is the part almost nobody is connecting: these contracts do not
stay off the balance sheet indefinitely. Once those data centers go
into operation, many of those obligations will start showing up in the financial
statements.
Put another way, this is not debt that disappears. It is debt with a
scheduled arrival date. What is a footnote today will be a balance sheet line in
some future
quarterly earnings report.
The question that actually matters
What happens if AI demand ends up growing less than markets are pricing in
today?
Because the data centers will still exist. The contracts will still be in
force. And so will the payment obligations.
That is the point: the commitment is fixed and the revenue meant to
pay for it is a projection. If the projection holds, the investment
justifies itself. If it does not, the obligation does not shrink to match.
Who is already watching
The Bank for International Settlements — the institution
that brings together the world's central banks — had already flagged this,
describing it as a form of shadow borrowing. Rating agencies such as
Moody's also began tracking these commitments closely.
It is worth being precise about what this does and does not mean.
It does not mean these companies are in trouble. They are five
of the most profitable companies on earth, with more than enough
free cash flow to carry what they have
committed to. What it means is that the market will have to verify, with results
in hand, whether future earnings justify an unprecedented infrastructure
investment.
What to watch in the next results
- The footnotes, not the headline. Headlines will talk about
record revenue and profit. Purchase commitments and future leases sit
further down, in the notes.
- How much moves onto the balance sheet each quarter. As
data centers come online, the transfer becomes visible. The pace of that
transfer is the data point.
- The gap between AI spending and AI revenue. If committed
spending grows faster than the revenue it generates, that gap is what to
watch.
- Whether the language about timelines changes. When a
company starts stretching the horizon on which it expects something to pay
off, that is usually the first signal — and it arrives before the numbers
do.
The analysis, then, is no longer about how much Big Tech is investing. We
already know that. It is about how long it will take to find out whether the bet
was worth the effort.
None of this is a buy or sell recommendation. It lays out an accounting
structure and the questions it leaves open. See our
legal notice.
Frequently asked questions
What is Big Tech off-balance-sheet AI debt?
Signed contracts — leased data centers, committed GPUs, servers, decades-long leases — whose obligations are not yet recognized on the debt line because the facilities are not operational. Nothing is hidden: they are disclosed in the footnotes to the financial statements. It is an accounting rule, not a trick.
How much do those commitments add up to?
A Nikkei Asia investigation puts Meta, Microsoft, Amazon, Alphabet and Oracle at a combined $1.65 trillion, more than the debt those same companies do report. Meta accounts for around $420 billion, and Oracle multiplied its own commitments more than thirtyfold in four years.
Will that debt eventually show up on the balance sheet?
Yes. It is not debt that disappears, it is debt with a scheduled arrival date: as those data centers go live, much of the obligation starts being recognized in the financial statements. What is a footnote today becomes a debt line in some future quarterly report.
