By Lynn Räbsamen, CFA | Advisory Board Member, CFA Institute | Author, Artificial Stupelligence

It is the most expensive question on the planet right now. Is artificial intelligence the revolution that rewrites the economy, or the most lavishly funded mirage in history?

The bears have made their case loudly, and some of them are famous for being right. But here is the strange part. The bubble panic is almost entirely a stock-market event. The people who lend AI companies their actual money are not behaving like a crash is coming. Not yet, at least.

So here is the honest version of the fight. The bull case, the bear case, and the one clever little signal that most of the noise ignores.

First, a simple trick for reading the room

Before the arguments, one idea that makes the whole thing easier to follow. Every big company is watched by two very different crowds.

The first crowd – the shareholders – owns the stock. They get rich if the company soars. These are the dreamers. When they are excited, they pay high prices today for profits they hope will arrive tomorrow.

The second crowd – the lenders – lends the company money. They get none of the glory. They just want to be paid back, with interest. These are the worriers. They are paid to imagine everything that could go wrong, and they charge more when they get nervous. When a worrier offers a low rate, it is not saying the company is unremarkable. It is saying one thing only: I am confident I will get my money back.

And here is the part that matters. Right now the dreamers are pricing AI companies like a miracle. That is where the bubble worry lives. But the worriers are lending to those same companies at the interest rate of an ordinary, unremarkable business. About what you would charge a furniture chain.

In other words, the worriers have not signed on to the bubble. Yet.

Hold that thought. It is the spine of everything below.

The referee: what a central bank found

The cleanest account came from the Bank for International Settlements (BIS), the bank that central banks themselves use. In January it published the rare report that neither cheerleads nor panics.

The calming part first. The macro risk is smaller than the headlines suggest.

By historical standards, the AI spending boom is not actually that large.

It runs at around 1% of US economic output, similar in size to the mid-2010s shale boom, roughly half the size of the late-1990s tech frenzy, and less than a fifth of Japan’s and Australia’s property and mining booms. So BIS judges the financial-stability risk as moderate, not acute.

Then the catch. All of it only works if these companies eventually earn the enormous profits their share prices assume. And the BIS spotted exactly the split we just described.

Stock markets dream up the valuations that make people cry bubble. The people professionally obligated to cry it first, are not crying.

Are the lenders the level-headed grown-ups in the room, or the last to know?

What the central bank found	The shareholders (the dreamers)	The lenders (the worriers)
How they price AI	Like a once-in-a-century winner	Like a normal business, around 6% interest
What that quietly assumes	Huge profits are on the way	Just pay us back, same as anyone
So if they are wrong	The profits never show up	They under-charged for a real danger

The bull case

The optimists are not naive, and they have receipts.

The scale is real and still climbing. Morgan Stanley expects roughly $3 trillion to be spent building AI infrastructure by 2028, with most of it still to come. Every few months the big technology companies raise their spending plans rather than cut them.

And real money is coming in. The bulls argue this is what separates 2026 from the dot-com crash, when companies burned cash and sold dreams. The divisions renting out AI computing power are growing fast, and global chip sales jumped more than 25% in a single year.

Anyone who actually uses these tools has felt the other half of it. The best models now command a premium, and people pay it without blinking. Claude’s top tier, Opus, costs real money, and the users who want to stay at the frontier hand it over rather than drop down to something cheaper and duller.

When the most capable model, Fable 5, was pulled from public access, the reaction was not a shrug. It was a queue of people waiting for its return, wallets open. That is what pricing power looks like, and it is the one thing a mirage cannot fake. Nobody lines up to overpay for something that does not work. The Apple store has been making that point for twenty years.

You can argue with a valuation. It is harder to argue with a customer who keeps paying more.

The investment firm KKR sums up the rest of the bull view: the roads and power plants of the AI age will keep paying off long after the excitement fades. In a gold rush, the people selling shovels get paid whether or not anyone strikes gold.

And the credit market is on their side. Lenders are still happy to fund all this on ordinary terms, which for a cash-rich giant like Microsoft is perfectly rational. A frothy stock price is not the same thing as a doomed company.

The bear case

The pessimists are not impressed, and their headline witness has done this before.

Michael Burry, the investor who famously bet against the 2008 housing bubble, says this one feels like the final months of the dot-com mania, and he has put more than a billion dollars behind that view. His favorite statistic, from Apollo economist Torsten Slok: almost 9 of every 10 venture capital dollars now chase AI. In 1999, the internet never came close. He enjoys reminding people that much of the “safe” debt issued in that earlier boom was worthless within two years.

He has company. Ray Dalio and the hedge fund Elliott lean bearish too, on the blunt logic that most AI companies will simply not survive long enough to matter.

Then comes the awkward number. Sequoia, one of Silicon Valley’s most respected investors, estimates the industry needs to find about $600 billion in new revenue every year just to justify what it is spending. Today it earns a small fraction of that. The gap is growing, not shrinking.

A bubble is not a big spending number. It is the gap between what you paid and what eventually shows up.

And when the executives building the technology start using the word “irrational” in public, that is not modesty. That is them quietly buying insurance.

The BIS added one warning supporting the bear case worth repeating. A lot of the borrowing is being tucked into corners that are hard to see. And, as the BIS put it, debt does not get safer just because it is harder to find.

So where is the debt hiding? Mostly inside special-purpose vehicles (SPVs), a polite name for shell companies built to hold debt at arm’s length. The tech firm sets one up, the shell borrows the money and owns the data center and the chips, then rents it all back. The debt sits on the shell; the parent keeps a clean balance sheet and just pays lease fees.

And the examples are not small. Meta raised about $30 billion this way for one Louisiana site, kept it off its books, then borrowed another $30 billion in bonds weeks later. Oracle has done the same across several projects, including a $38 billion package for two of them. xAI built a vehicle to buy Nvidia chips and lease them back to itself.

By one Financial Times count, over $120 billion of AI financing, most of it debt, now sits off the books.

Here’s where the bulls and the bears draw their lines:

The question	Bull case	Bear case
The spending	~$3 trillion to be invested by 2028, most still to come (Morgan Stanley)	Big Tech now spends nearly all the cash it makes, and borrows the rest (Bank of America, JPMorgan)
The revenue	Cloud and chip revenue is real and growing fast, and customers pay a premium for the best models (chip sales up 25% in a year)	Real AI revenue is still a fraction of the spend, a ~$600 billion yearly gap that keeps widening (Sequoia)
The financing	The debt is backed by cash-rich giants and real assets, and is modest next to their cash flows	Over $120 billion of it is hidden in shell companies, off the books and out of sight (Financial Times)
The history lesson	The infrastructure keeps paying off long after the hype (KKR)	“The last months of the dot-com bubble,” and that era’s safe debt went bad fast (Michael Burry, citing Apollo’s Torsten Slok)
The verdict	A market picking real winners, not a mania	Most AI companies will not survive (Ray Dalio, Elliott)

Here’s the signal that headlines miss

The useful question is no longer “is AI a bubble.” The bubble argument is almost entirely a stock-market conversation. The bears are shouting, the bulls are shouting back. The lenders, however, are quietly getting on with their day.

That is the signal worth watching, and it is refreshingly specific. The day lenders start charging AI companies more to borrow is the day the bond market finally agrees with the bubble-callers. Credit markets are slow, but they are not sentimental. When the worriers start worrying, that tends to happen before the magazine covers do, and well before the stock price does.

Watch the spread. It will move before anyone rings a bell.

For now, the entire shouting match is happening on one side of the room. The people who lent the money have not looked up.

The last word

Bubbles never send a save-the-date. They leave a price on something, wait for the lenders to get nervous, and by the time everyone notices, the window has closed.

Right now the lenders are not nervous. That is either the most reassuring thing about AI, or the most interesting thing to keep watching.

Either they know something the bears do not. Or they are about to get a very expensive lesson.

---

For more insights about what AI can or cannot do, check out my book “Artificial Stupelligence: The Hilarious Truth About AI”.

Subscribe here to be the first to receive my insights.