The AI Spending Paradox: One Side Pays $20 a Month, the Other Owes Trillions

Source: Patrick Boyle | Published: 2026-08-04T11:15:06Z

The Economist estimates that recouping this round of AI infrastructure investment would require $2.5 trillion in annual revenue—more than the entire global tech industry earns today.

On financial YouTube, one word kept coming up: Enron.

What the Enron Accusation Actually Means

Enron was the American energy giant that collapsed in 2001. It used a network of secret off-balance-sheet entities to hide massive debts and losses, leaving investors looking at accounts that were nearly fictional. The company imploded within weeks of exposure, taking down Arthur Andersen — then one of the world's five largest accounting firms — and wiping out thousands of employees' retirement savings.

So when someone invokes Enron alongside big tech, they're not complaining about messy books. They're accusing criminal fraud.

That accusation doesn't hold up.

When you actually dig into the composition of this "hidden debt," most of it is bulk GPU purchase contracts for hardware not yet delivered, and data center leases for buildings not yet built. Accounting rules are clear on this: no asset, no building — it stays off the balance sheet and gets disclosed in the footnotes.

If you sign a two-year phone plan at $50 a month, you've taken on $1,200 in real obligations. But you don't record a $1,200 liability on your personal balance sheet the day you sign — you pay monthly, as you go. Tech companies are doing exactly the same thing, just with more zeros. A 15-year data center lease in Ohio and your phone contract follow identical accounting logic.

This isn't fraud. It's a timing difference.

The Real Aggression Is Up Front

The place actually worth examining isn't the footnotes — it's the headline "adjusted earnings" on the cover page of the earnings release.

Tech companies love talking about EBITDA — earnings before interest, taxes, depreciation, and amortization. Charlie Munger suggested that every time you read the word, you should mentally substitute "BS earnings." His logic: depreciation is a kind of reverse float. You pay cash upfront to buy equipment; the expense appears gradually as the asset wears down. Strip out depreciation and you're assuming physical assets never deteriorate. That's a convenient assumption. It's almost never true.

Stock-based compensation gets even more aggressive treatment. Tech companies pay employees heavily in equity, then exclude that cost from the earnings they report to investors on the grounds that it's "non-cash." NYU's Aswath Damodaran has called this "one of the most egregious abuses in modern financial reporting." Buffett asked the same question: if options aren't compensation, what are they? If compensation isn't an expense, what is?

To prevent this equity from diluting share counts, companies spend real cash buying back shares — and announce this to the world as "returning value to shareholders." In practice, it's an expensive treadmill: the faster you run, the more you stay in place. And buybacks offer no timing flexibility — every quarter, you keep buying regardless of the price.

The real cash pressure is already showing. In the latest quarter, the four major hyperscalers combined generated their lowest free cash flow in a decade — just $7 billion total — and Alphabet posted its first net cash outflow since going public.

Whose Debt Is Nvidia Backing?

A harder-to-explain dynamic is taking shape.

Nvidia is currently engaged in more than $750 billion worth of AI deals. This includes: guaranteeing $250 billion in compute leases for OpenAI, then funding $350 billion in chip purchases for OpenAI; putting $5 billion into a new company founded by Ilya Sutskever; Google backing $35 billion in lease loans for Anthropic; SoftBank committing $65 billion to OpenAI and raising $40 billion in bridge financing specifically for this purpose.

Map out the relationships, and the companies at the core of the AI boom are largely investing in each other.

Nvidia CEO Jensen Huang dismissed the concern as "absurd" — a word he used while guaranteeing purchases of a quarter-trillion dollars' worth of his own products. His defense has some merit: this is old-fashioned vendor financing, something aircraft manufacturers and telecom equipment companies have done for decades. For Nvidia, it locks in customers, ensures chips actually get deployed, and if the bet pays off, yields a stake that could be worth a fortune.

The problem is the cascade if the bet fails. Equity going to zero is painful but usually survivable. Guaranteeing someone else's debt is different — if the customer defaults, a valuation problem becomes a solvency crisis. Nvidia generates roughly $200 billion in cash flow annually; one or two failing startups would be manageable. But the guarantees are climbing toward hundreds of billions, and the company that once carried no debt is now standing behind everyone else's.

The market noticed. On the same day these deals were announced, the cost of insurance against Nvidia debt default posted its single largest daily jump on record. The people whose job it is to price the risk of Nvidia's obligations looked at all of this and were, visibly, unsettled.

The $2.5 Trillion Gap

The Economist estimates global AI infrastructure spending this year at around $900 billion, with more than $400 billion of that coming from borrowing. The math that follows: to service these investments, the AI industry needs to generate roughly $2.5 trillion in annual revenue — a figure that exceeds what the entire global tech industry earns today from all of its businesses combined.

The current reality is nowhere near that number.

About one in five U.S. companies say they use AI in some capacity, but many are on free tiers. A joint Bank of England and Federal Reserve survey found executives spend an average of 1.5 hours per week using AI. The largest capital deployment in human history is supporting roughly ninety minutes of executive attention per week — somewhere between lunch and the commute home.

Paying customers aren't spending much more. Fintech company Ramp analyzed real corporate spending data and found the median enterprise spends $10.66 per employee per month on AI. The most striking number from the Bank of England research: nine in ten executives said AI had produced no measurable change in their company's productivity over the past three years.

Damodaran and his collaborator Bradford Cornell call this the "big market delusion": a new technology appears, tied to a vast potential market; a wave of companies piles in; investors price every company as a future winner. The problem is there's only one winner. Sum up the market expectations across all these companies and the total exceeds the market itself — everyone gets priced as first place in a race with one champion.

The 93% in the SpaceX Prospectus

This logic of narrative-driven pricing finds its clearest specimen in the SpaceX IPO.

In June, SpaceX went public. Its prospectus claimed a total addressable market of $28.5 trillion — of which $26.5 trillion, or 93%, was attributed to AI or Grok. The remaining $2 trillion covered everything else: rockets, launches, satellites, global broadband, Twitter... The actual space company was reduced to a rounding error. This was the basis for a $135-per-share price.

You might think that was aggressive enough. Then, weeks later, a Wall Street analyst issued an $800 price target, implying a valuation north of $10 trillion — for a company that earned less than $19 billion in revenue last year and is still losing money.

One analyst totaled up the company's committed expenditures from the prospectus and concluded SpaceX faces roughly $235 billion in obligations through 2030. IPO proceeds covered only a fraction of that, leaving approximately $170 billion to be filled through continuous stock and debt issuance over the coming years. That means a lot of underwriting business.

The 18 banks that underwrote the IPO all published research at almost exactly the same moment — the instant the lockup rules allowed it, roughly 25 days after the listing. Fortune described the reports as "almost uniformly bullish." Morgan Stanley called SpaceX "the ultimate AI frontier." Bank of America said it was "laying the superhighway to the stars." Raymond James compared it to the invention of electricity, railways, and the internet. Among 30-plus analysts, exactly one issued a sell rating — from an independent research firm with no underwriting business.

This dynamic was once constrained by something called the Global Research Analyst Settlement. The rule emerged from the rubble of the dot-com bust in 2003, erecting a firewall between investment bankers and research analysts to prevent analysts from publicly recommending a stock while privately emailing that it was garbage — not a hypothetical; the cases that triggered the settlement involved analysts who had literally done exactly that. The SEC repealed the rule in December, citing a desire to "reduce compliance friction."

Why Smart Money Doesn't Short This

So is anyone noticing these problems and profiting from them?

In theory, the logic is airtight: read the footnotes, identify the overvaluation, short the stock, wait for the market to come to its senses. In practice, three things make this fail.

In 1996, accounting professor Richard Sloan published one of the most cited papers in the field. He decomposed corporate earnings into two parts: real cash income, and accruals that depend on management judgment. He found something the market was clearly ignoring: accruals are far lower quality than cash. Companies propped up by accruals consistently disappointed later; companies with cash-backed earnings consistently outperformed. Yet the market treated both types of earnings as equally credible. Simply betting that the market would eventually notice this difference generated excess returns for years on end.

The second barrier comes from professor Robert Bloomfield's "incomplete revelation hypothesis": public and accessible are not the same thing. The number you need is technically disclosed — buried on page 83 of a 200-page filing, scattered across four footnotes, requiring you to piece it back together yourself. Extracting it costs time, effort, and attention, none of which are free. The harder a fact is to surface, the less it gets reflected in the price. The debt wasn't hidden — it was just filed somewhere boring enough that you wouldn't go looking.

The third barrier, documented extensively by economists Mitchell and Pulvino, is the limits of arbitrage. Finding a problem in the footnotes doesn't mean you can profit from it — you also have to stay solvent long enough for everyone else to catch on. If you short a narrative stock beloved by millions because you understood a contractual commitment clause, that stock can keep rising long after your short position has been wiped out, before the market slowly begins to care about what you found. Smart money knows this. So it often doesn't touch high-conviction narrative stocks at all. Mispricings survive not because no one sees them, but because the people who do can't outlast the people who don't.

Who Is Actually Benefiting from AI

There's a genuine reversal buried in this conversation.

The giants burning capital to build AI infrastructure are waiting for $2.5 trillion in annual revenue to arrive. The people most clearly benefiting from AI right now are on the other end of the transaction.

A survey by payroll company Gusto found the share of new founders using AI to launch their businesses doubled in two years, reaching 60%. They're not using AI to cure diseases or replace departments. They're building websites, handling local registration paperwork, doing things that once required hiring someone. They're paying about $20 a month.

This is currently AI's clearest real-world win: not the person who spent hundreds of billions waiting to monetize it, but the person who spent $20 to open a sole proprietorship.

That's a very good deal for the one renting the compute. For the one who spent hundreds of billions building it and is still waiting for $2.5 trillion in annual revenue to slowly accumulate — that's another story. Until that number arrives, these companies will keep borrowing, keep burying obligations in footnotes, and markets will keep glossing over them — until one day, suddenly, they decide they want to look.

Jamie Dimon put it recently: Will AI deliver returns? Probably, the way the internet did. Will it happen on the timeline you expect, in the way you expect? Definitely not.