The AI build-out is being financed in ways that make reported profits look stronger than the cash behind them. Satyajit Das, who spent a career inside the complex debt products that blew up in 2008, argues that the giant technology firms and the AI labs are repeating familiar accounting tricks: owning stakes in one another, lending customers the money to buy their products, trading favors instead of cash, stretching the life of equipment on paper, and parking huge payment promises where the official debt number cannot see them. History, he says, does not repeat, but it rhymes—and the rhyme with earlier boom-busts is hard to miss.
When the Sale and the Financing Are the Same Trade
A normal sale is simple. You hand over a product, the buyer hands over money, and that money is revenue. Das says much of the AI trade is not working that way. Microsoft (MSFT) owns a piece of customers that own a piece of Microsoft. Chipmakers and cloud companies sell equipment to buyers they are also financing. Firms buy from one another so that, as Das put it, “no cash changes hands.”
The accounting split is what matters. When Nvidia (NVDA) sells chips, the sale counts as revenue, the top-line number investors watch. When Nvidia helps pay for that purchase by taking a stake in the buyer, that outlay is not subtracted from profit. It is recorded as an asset, something the company owns. On paper the sale happened. In the bank account, the cash never arrived. It is a little like a car dealer booking a sale, then lending the customer the purchase price and calling the loan an investment.
$160 Billion of Income That Never Arrived
Das says the gap widened last quarter. Close to half the reported income of the biggest AI spenders, about $160 billion, was not money received. It was a write-up: the companies raised the paper value of AI stakes they already held, using valuation models rather than an actual sale. None of it was cash. Think of marking your house up by $100,000 because a website says the neighborhood is hot, then treating that $100,000 as income even though you have not sold the house and no buyer has paid you.
Depreciation stretches the picture further. When a company buys chips, it does not count the whole cost as an expense on day one. It spreads the cost over the years it claims the chips will last. Das thinks those schedules are too long. The chips become outdated faster than the accounts assume, so the expense shows up later than the equipment actually wears out. Profit looks better now because part of the bill has been pushed into the future.
$1.4 Trillion on the Books, $1.8 Trillion Off Them
A balance sheet is the official list of what a company owns and what it owes. Das says a large share of the AI bill never makes that list. Nvidia may show less than $50 billion of debt on its balance sheet. Add the payment promises kept off the books and the figure approaches $200 billion. A Bloomberg tally of Alphabet (GOOGL), Amazon (AMZN), Meta (META), Microsoft, Nvidia, and Oracle (ORCL) found about $1.4 trillion of debt on the books against $1.8 trillion of future lease and purchase commitments that do not appear there. The second number is money these companies are already on the hook to spend. It just is not labeled “debt” in the headline figure.
The Cash Flow Is Already Spoken For
OpenAI and Anthropic sit in the middle of this web of mutual purchases and financing. OpenAI alone, Das says, has to fund something like $1.5 trillion of commitments over time. That is another way of saying it must eventually bring in that much revenue. Until it does, the risk is concentrated: if the lab cannot pay, the companies that sold to it and financed it feel it too. The growth is real on the income statement. The ability to pay is not yet proven.
The answer Das hears most often is that this is fine because it is funded by equity, meaning shareholders’ money rather than loans. He does not buy it. Equity lost is still savings destroyed. The dot-com boom was largely equity-financed, and investors still lost fortunes. This boom is not equity alone anyway. The big technology firms are using up the spare cash their existing businesses throw off, and they are borrowing in markets Das says he has not seen firms like Amazon tap before: Australia, Switzerland, Europe, Britain. If that cash keeps going into projects he expects will lose money, the dividends and share buybacks that have supported stock prices are likely to shrink. Longer term, he thinks they will have to sell new shares, diluting existing owners.
Gods in the Texas Desert, or a Delusion in Real Time
Scale sharpens rather than softens the exposure. The outlay is roughly 17 times the dot-com investment and four times the subprime housing bubble. Investors who say they own only the “picks and shovels” are not outside the history. Most of the fiber-optic capacity laid in the 1990s is still only about half lit, with average utilization near 26 percent. Cisco (CSCO) took 25 years to regain its 2000 share price; survivors such as Microsoft, Apple (AAPL), Amazon, and Oracle fell 60 to 90 percent and needed decades to recover. A commentator Das quotes put the present wager cleanly: “Either the gods are being built in the Texas desert… or the greatest financial delusion in human history is unfolding in real time.” The momentum, he allows, is not finished. The accounts, in his reading, already show where the cash is missing.
Comments
Log in or sign up to join the conversation.