How the AI bubble will burst

As often happens in the early buildout of a new technology, a fierce argument is raging about whether the AI spending boom is a rational investment or a speculative bubble.

The answer, very likely, is… both.

AI will radically change the economy and society — just as the internet, computers, cars, trains, canals, electricity, and other technologies did.

And…

Much (most?) of the money being invested right now will probably be lost — just as much of the money invested early in the prior booms was.

The early AI boom will likely end the way most prior eras have ended — with a gigantic bust that clobbers the first wave of companies and investors, followed by a long boom that helps build huge new companies and transforms the way we live.

Personally, I hope there’s never an AI bust. It would hurt millions of people, including me. AI investment accounts for a huge share of US economic growth (a third, by some estimates), so a bust will likely trigger a bear market and major recession. I don’t run an AI company or trade frequently, but I do own stocks and benefit from a healthy economy. So, if there’s an AI bust, I’ll get poleaxed along with everyone else.

But having had a front-row seat for two big booms-and-busts — the Internet (1995-2002) and the Great Financial Crisis (2002-2009) — I’m fascinated by the parallels (and differences) with this one. So, while I hope the AI boom will never turn to bust, I will also keep analyzing out loud.

Growth and leverage — we’ve got ’em!

Most speculative bubbles share two key elements:

1. An exciting innovation or product that leads to enormous demand (“Growth”)

2. Debt, credit, circular financing, or other leverage that amplifies this demand (“Leverage”)

In the internet bubble, the growth was real. Millions of new people connected every month, and usage went through the roof. For five years, insatiable demand drove astonishing fundamental growth at dozens of companies. (Amazon, AOL, Yahoo, Exodus, Level 3, Worldcom, Cisco, Juniper, etc.)

The “leverage” was also clear — especially in hindsight. Companies borrowed hundreds of billions of dollars to build out telecom networks and buy new gear. Internet service providers raised tens of billions in equity investments. For five years, these factors combined to produce one of the biggest tech booms the world had ever seen. But by 2000, the supply of internet products and services finally caught up with demand, rising interest rates raised the cost of borrowing, and hundreds of IPOs and follow-on offerings sated demand for internet investments…

The music stopped. Growth slowed. And leverage reversed.

Similarly, in the years leading up to the Great Financial Crisis, low interest rates and mortgage innovation fueled an extraordinary housing boom. But by 2008, there were no more lending standards to relax, and anyone who wanted a house had one. Returns for lenders and investors plunged, and stretched borrowers could no longer make payments. Demand and leverage dropped. Once again, it was “look out below.”

These same dynamics are now repeating themselves with the AI boom.

AI companies are growing faster than any in history.

Look around, and you can see clear evidence of both amazing growth (demand) and leverage in our current tech-fueled frenzy.

Take Anthropic’s new coding tool, Claude Code, which ignited spectacular demand and helped the company’s revenue explode to $4.8 billion in the first quarter of this year. In the second quarter, revenue more than doubled, to $11.6 billion. Based on July’s results, Anthropic is now generating an astonishing $65 billion in annualized revenue. That’s up a breathtaking 7x from last year. No company in history that I know of has ever grown like that. The Wall Street Journal also reports that Anthropic is now profitable. This nukes the previous (widespread) naysaying that AI economics don’t work and the leading model companies “can’t make money” and will go bankrupt. Anthropic may not stay profitable. But it is now.

Someday, AI supply will catch up with demand, and the financing leverage will max out.

Even OpenAI, which lost focus and fumbled its early industry lead, is now reportedly doing $40 billion of annualized revenue. That may not be apples-to-apples accounting with Anthropic, but it’s still a colossal number. Despite its revenue growth, OpenAI, in my view, is on track to become the Netscape of the AI era: The company that kicked off the boom and, for a brief moment, became synonymous with it… but then missed a turn and failed. OpenAI’s sequential growth slowed radically in the second quarter. Revenue was up “only” 18%, to $6.7 billion from $5.7 billion in the quarter before, the Journal reports. This is the equivalent of a racecar swerving off the track while another car blows past. But the growth of both companies shows that the underlying user demand for AI is astonishingly strong.

The question now is how long Anthropic’s growth can continue like this. If the company’s run-rate is $65 billion now, it will likely be $100 billion by early next year. This makes the company’s recent valuation of nearly $1 trillion look downright reasonable. The global market for enterprise software spending — the primary source of this revenue — is estimated to be $1.4 trillion, so there would seem to be plenty of room for growth in this segment alone. Beyond that, Anthropic will have to penetrate new segments to sustain its boom.

Leverage, meanwhile, is everywhere

Many have noted the jaw-dropping amounts that investors and companies are shoveling into AI-related projects. The internet boom looks like penny poker by comparison. The tech giants — Alphabet, Meta, Amazon, et al — have gone from generating tens of billions of dollars of free cash flow every year to investing so much in AI infrastructure that they’re burning cash. Moreover, as this excellent Journal article highlights, these companies (and others) are making far larger future spending and investment commitments than their current financial statements suggest.

As just one example, Alphabet’s “off balance sheet” commitments increased by almost $500 billion in the past three months alone. $500 billion! Last year, Alphabet generated about $25 billion of free cash flow per quarter, or about $100 billion a year. Even at that rate, it would take Alphabet 5 years to pay off the commitments it has made in the past 3 months. But Alphabet is no longer generating any cash. In the second quarter, for the first time in its history as a public company, it burned cash. Alphabet does, at least, have the capacity to generate cash through its own operations. Investors who have to borrow cash to make AI investments don’t.

Then there are the much-publicized “circular” financing deals. For example, chipmaker Nvidia just announced an agreement to invest $1.5 billion and provide $105 billion in credit to build out a giant Ohio data center. The data center owner will presumably use some of that money to buy Nvidia chips. Separately, Nvidia has agreed to invest $30 billion in OpenAI, which will use some of that cash to lease data center space. This vendor financing and investment isn’t sleazy or illegal. But it takes the cash and borrowing capacity of Nvidia and turns it into financing for future purchases of Nvidia chips. And it allows OpenAI and other AI providers to buy far more chips and compute than they would without the financing.

In short, it’s leverage. And it amplifies and supports demand for the whole AI ecosystem.

Someday…

Unless the AI boom truly is different — unless the end-user demand for AI services is so insatiable that demand remains ahead of supply forever — we know how it will end. Someday, AI supply will catch up with demand, and the financing leverage will max out. There could be another trigger, like rising global interest rates, that makes investors and executives more cautious about making such enormous long-term commitments. That day may be many years in the future. Or it may already be here.

Internet veterans like to debate where we are in the AI cycle. If the current moment is akin to, say, 1997, the AI investment boom is just getting rolling. If it’s more like late 1999, we’re Wile E. Coyote sprinting in midair after running off a cliff.

One of the lessons of both the Internet boom-and-bust and the Great Financial Crisis is that you can be almost certain that a phenomenon is a bubble and still miss the top. This is true not just for casual observers (and young analysts, like I was) but for the best investors and executives on the planet. Just ask Stan Druckenmiller and dozens of other brilliant and experienced investors who held on too long in 2000. Or the bank CEOs who levered up and bought too many risky mortgages in the early 2000s and steered their ships into disaster.

These folks weren’t blind or stupid. They knew everything that could be known, including that the internet and housing booms were probably bubbles. And they still missed it.

So, what to do?

If AI is one of the biggest opportunities in the history of the world… and also probably a bubble that will burst and lay waste to most early investments, how should investors and executives (and others) approach it?

The same way we approach all decisions that depend on an uncertain future: By acknowledging the uncertainty and being ready for anything. Specifically, by investing, learning, and taking advantage of the amazing new technology… while not betting more than we can afford to lose.

A version of this piece originally ran in Regenerator; it is reprinted here with permission.

Henry Blodget is the cofounder and former CEO of Business Insider. He now writes the Regenerator newsletter and hosts the Solutions podcast. He has a range of investments, including shares of Amazon.