Big Tech's off-balance-sheet debt isn't fraud — nobody reads the footnotes
The supposed $1.65 trillion of hidden debt is not Enron-style fraud but long-term leases on data centres that aren't running yet, plus purchase commitments — legal, but buried in the footnotes. The real danger is that the AI giants are financing each other and telling their story through adjusted earnings.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The $1.65 trillion isn't hidden debt, it's goods not yet delivered
The $1.65 trillion of off-balance-sheet debt reported by Nikkei is mostly long-term GPU purchase agreements and leases on data centres that are not yet in service. Under standard accounting rules, goods that have not been delivered and buildings that are not yet operating do not have to be booked as balance-sheet liabilities; disclosure in the footnotes is enough — much as signing a two-year phone contract does not put a $1,200 liability on your books the same day. Meta added $233 billion of commitments last quarter, of which $96 billion in leases will move onto the balance sheet once the data centres go into use. So this is not Enron-style fraud; it is a timing difference.
— Patrick BoyleBorrowing rather than selling equity is how management shows it believes
If management believes it can turn $1 into $5, it will prefer to borrow rather than sell equity, so as to keep the excess profit for itself; only when the outlook is uncertain is it willing to bring in outside shareholders. Alphabet closed the largest equity raise in history in June, at nearly $85 billion, and Berkshire put in $10 billion, buying at a discount of about 6% — this is not a firm that usually pays a premium for a story. The signal is not in the choice of financing but in the scale of the debt, equity and convertibles being drawn on at record levels all at once: the company thinks there is something behind this enormous outlay that is worth fighting over.
— Patrick BoyleIgnoring depreciation means assuming the equipment never wears out
Big Tech would rather talk about EBITDA, but Munger's advice was to read EBITDA as BS earnings. Depreciation floats in reverse: the cash goes out first, the expense shows up later; ignoring depreciation means assuming the equipment never wears out. In the latest earnings season, the four big hyperscalers generated only $7 billion of free cash flow between them, and Alphabet's cash flow turned negative for the first time since it went public. Stock compensation should not simply be added back either — it is barter: handing out shares directly is as much a cost as selling shares to pay wages. To offset the dilution, companies are forced to buy back on a fixed rhythm, and cannot wait for a lower price.
— Patrick BoyleIn a default you lose the customer and the money you lent them
Nvidia is pushing through more than $750 billion of AI deals: $250 billion of support for OpenAI's leases, then financing for $350 billion of chip purchases, plus $5 billion invested in Ilya Sutskever's startup. Google is providing $35 billion of lease support for Anthropic. SoftBank has committed $65 billion to OpenAI and lent it a $40 billion bridge loan. The giants are investing in each other and buying from each other. It resembles a carmaker lending to customers so they can buy cars, but it is also old-fashioned vendor financing. The risk is that in a default you lose both the customer and the money you lent the customer.
— Patrick BoyleEvery firm is priced as a winner, and the sum is mathematically impossible
Aswath Damodaran and Bradford Cornell call this the big market delusion: every investor treats the company they have bet on as the future winner, but add up the money each of them is expected to earn and the number exceeds the market itself — mathematically impossible. The Economist estimates roughly $900 billion of AI build-out spending this year, of which more than $400 billion comes from borrowing; earning that back would require around $2.5 trillion of AI revenue a year, more than the revenue of the entire global tech industry today. Actual adoption is modest: American executives use AI for about 100 minutes a week on average, and the median corporate spend is just $10.66 per person per month.
— Patrick Boyle93% of SpaceX's market comes from AI, not from rockets
In its listing prospectus, SpaceX attributed $26.5 trillion of a $28.5 trillion TAM — 93% — to AI/Grok, leaving about $2 trillion for the actual businesses of rockets, satellites and the rest. Analysts then put out an $800 price target, implying a market capitalisation of more than $10 trillion, against company revenue of less than $19 billion last year. The reason is that SpaceX disclosed about $235 billion of spending commitments through 2030, of which the IPO money covers only a small part; the roughly $170 billion gap has to be filled by issuing stock and debt repeatedly over years to come — precisely a Wall Street underwriting feast. Analysts at the 18 underwriting banks published bullish notes almost simultaneously; only one independent shop put out a sell rating.
— Patrick BoyleBurying the debt on page 83 is a rational bet
Public is not the same as usable. Robert Bloomfield's incomplete revelation hypothesis holds that the harder information is to extract, the less fully it gets priced; putting the debt on page 83 of a 200-page document, spread across four footnotes, amounts to choosing that most people will not look. Richard Sloan found in 1996 that the market treats cash profits and accrual profits as equivalent, even though accrual profits are less sustainable. The limits of arbitrage then explain why someone who does read the footnotes can short but may be wiped out before the market corrects, which is why smart money often stays away from hyped stocks. Burying inconvenient numbers in the footnotes is a rational bet for a company: it has disclosed, it is legal, and it often works.
— Patrick BoyleThe winners from AI are the people paying $20 a month
This round of «hidden debt» panic is not the discovery of the next Enron; it is people finally reading the footnotes and realising the true scale of the spending. AI is not a fraud — it is obviously useful — but the best returns are not going to the giants burning hundreds of billions; they are going to the founders running a one-person business on AI, for about $20 a month. For the renter it is a great deal; for the side that spent enormous sums building it and is still waiting for $20 monthly fees to add up to $2.5 trillion of annual revenue, it is terrible. As Jamie Dimon put it: AI may deliver returns like the internet did, but not on the timetable you expect.
— Patrick BoyleIn their own words · checked verbatim
So it isn't Enron-like fraud, it's camouflage which only really works on people who aren't paying much attention.
Patrick Boyle7:21
every time you read the word EBITDA you should replace it in your head with BS earnings.
Patrick Boyle8:22
If options aren't a form of compensation, what are they? If compensation isn't an expense, what is it? And if expenses shouldn't go into the calculation of earnings, where in the world should they go?
Patrick Boyle9:29
The story is doing all the work and nobody's minding the numbers.
Patrick Boyle15:46
The largest capital investment in the history of the species is being justified by an hour and a half per executive per week, so somewhere between lunch and a drive home.
Patrick Boyle16:51
Will AI pay off? Probably the way the internet did. Will it pay off the way you expect on the timeline you expect? Definitely not.
Patrick Boyle31:37
Figures
| Off-balance-sheet debt reported at the five big tech companies | $1.65 trillion | 0:00 |
| Size of Alphabet's equity raise (largest in history) | nearly $85 billion | 5:12 |
| Meta's new AI commitments in one quarter, and the leases within them | $233 billion, of which $96 billion in leases | 6:17 |
| Total AI-related deals Nvidia is pushing through | more than $750 billion | 11:29 |
| The Economist's estimate of total AI build-out spending this year | about $900 billion, of which more than $400 billion is borrowed | 16:51 |
| The Economist's calculated annual AI revenue threshold | about $2.5 trillion | 16:51 |
| Median corporate AI spend per person per month, per Ramp | $10.66 per person per month | 16:51 |
| Bank of England survey: executives saying AI has not raised company productivity in three years | 9 out of 10 | 17:57 |
| Market capitalisation implied by underwriting analysts' target price at the SpaceX listing | more than $10 trillion ($800 a share) | 20:06 |
| SpaceX's disclosed spending commitments through 2030, and the financing gap | about $235 billion, with a gap of about $170 billion | 21:09 |
Glossary
- off-balance-sheet vehicle
- A special-purpose entity that keeps debt off the company's own statements, though it must still be disclosed in the footnotes.
- vendor financing
- The seller lends to the buyer so the buyer can purchase the seller's own product, locking in the customer while taking on default risk.
- big market delusion
- Investors price every company in a new arena as the eventual winner, so the valuations add up to more than the addressable market truly is.
- incomplete revelation hypothesis
- The harder information is to extract, the less fully the market prices it; facts hidden in hard-to-read footnotes get underweighted.
- limits of arbitrage
- Even when mispricing is spotted, a short can be wiped out because irrationality persists, which keeps smart money out of the trade.
How to listen
Investors trying to judge whether AI capital spending has become a bubble, and founders and analysts who need to read a tech company's financials properly.
Listeners who already know the Enron case can skip the first three minutes of case recap.