Howard Marks: Is AI in a Bubble?

On December 9, Howard Marks published a memo at Oaktree Capital titled "Is It a Bubble?" It lays out, clearly and systematically, that the real risk of an AI bubble lies not in the technology failing, but in this: we are betting on an "incalculable future."

Good technology does not equal a good investment.

This is a great, great, great piece, and I strongly recommend reading the whole thing. That said, I've quickly summarised the five insights I took away from it.

As for whether there really is a bubble… that's for each of you to judge.

Highlight

1. The biggest risk of AI is not "the bubble bursting" but "the bubble being impossible to price."

This is not 1999, and it is not 2007.

So far, when it comes to AI, we…

don't know its final form (tool / platform / infrastructure / general intelligence)

don't know where the profits will go (model companies / cloud providers / chipmakers / the application layer)

don't know the competitive structure (oligopoly / commoditisation / open-source disruption)

don't know the time horizon (5 years? 10 years? or an exponential leap?)

What this means: the market isn't "mispricing" so much as forcibly modelling something that cannot be modelled. That makes it different from traditional bubbles.

2. "Inflection bubbles" are dangerous precisely because they make wrong decisions look like right ones.

Railways, electricity, the internet, AI… are essential to civilisation, but not necessarily to investors. What history shows again and again is this: the more real the technological progress, the easier it is for investors to justify "the wrong price, the wrong company, the wrong structure." In the AI context, it sounds like: "The direction is right, never mind whether it's worth it; this is infrastructure, you can't judge it by traditional ROI; as long as you back the winner, nothing else matters"…

The most dangerous thing about inflection bubbles: they systematically dull investors' sensitivity to mistakes.

3. This may be the first technological revolution with deep debt involvement. This is the most underrated — and most lethal — insight in the whole piece.

Past technological revolutions — railways, electricity, the internet… — were mostly trial-and-errored with equity capital. This time, though, we have…

30-year bonds

off-balance-sheet SPV structures

data centres as collateral

long-term fixed interest

We are moving at scale into a highly uncertain technology race. Debt only works when "the outcome is uncertain but can be assessed." When "the outcome is pure guesswork," debt is dangerous. In the past, technological failure ≠ financial failure; this time, technological failure will mean stress on the financial system.

The question isn't "will the bubble burst?" but "if something goes wrong, will the financial structure amplify it into a systemic event?"

4. The AI industry is already showing a "perpetual-motion-machine"-style financial illusion. This is a very contemporary, very concrete risk. It doesn't say OpenAI or Nvidia have a problem; it points to a structural illusion —

A gives B money

B uses that money to buy A's services

Both sides get to "talk growth, talk investment, talk future"

This genuinely happened in the 1990s telecom bubble, and ultimately produced

a boom that "looked profitable, while the system created no real external demand."

When most of an industry's demand comes from "players buying from one another," the market overestimates how mature real societal demand is. This badly misleads:

the pace of capital allocation

the scale of capacity build-out

judgements about debt sustainability

5. AI may, for the first time, simultaneously destroy the "work–income–meaning" trinity.

This isn't an emotional postscript; it's the ethical foundation of the whole piece.

If what AI replaces is

entry-level jobs + cognitive work + apprenticeship paths,

then how does society go on producing "experienced people"?

This isn't Luddite (anti-technology) anxiety but a structural paradox —

AI raises productivity → fewer jobs

Fewer jobs → lower purchasing power

Experience-building jobs disappear → professional pipelines break

Loss of meaning → less social stability

This means AI's risk isn't just an economic cycle but a reshaping of the social structure.

These conclusions deserve real vigilance and serious attention, and even more precious is Howard's brilliant line of reasoning. Here goes — enjoy:

Is It a Bubble?

We are living through an extraordinary moment in world history. A transformative technology is emerging, and its backers claim it will change the world forever. To build it, companies must invest sums far beyond anything in living memory. News coverage is full of widespread worry that America's largest companies are propping up a bubble that is about to burst.

When I visited clients in Asia and the Middle East last month, I was often asked: is there a bubble around artificial intelligence? Those discussions are what prompted this memo. As usual, I'll start with some disclaimers: I don't take part in the stock market; I simply regard it as the best barometer of investor psychology. Nor am I a technology expert — I know no more about AI than most generalist investors. But I'll do my best.

The most interesting thing about bubbles is their "regularity" — not regularity in timing, but consistency in how they unfold. Something new and seemingly revolutionary appears and quietly works its way into people's minds. It fires the imagination, and the excitement quickly gets out of hand. Early participants reap enormous gains; onlookers are filled with intense envy and regret, and, driven by the fear of continuing to miss out, they rush in. They don't know what the future holds, nor do they care whether the price they're paying can deliver a reasonable return at an acceptable level of risk. The result, almost always, is pain for investors in the short to medium term, even if, given a long enough time, they may eventually come out ahead.

I've lived through several bubbles and read many accounts of others, and every one of them fits this description. You might think the losses from past bubbles bursting would be enough to prevent the next one from forming. But they aren't, and I'm convinced they never will be. Memories are short, and prudence and natural risk aversion are no match for the dream of getting rich on a revolutionary technology that "everyone knows" will change the world.

The passage I quoted at the start comes from Derek Thompson's November 4 newsletter, "AI Could Be the Railroad of the 21st Century. Brace Yourself," which explores the parallels between today's AI build-out and the railroad boom of the 1860s. The fact that the passage applies word for word to both illustrates the saying often attributed to Mark Twain: "History doesn't repeat itself, but it does rhyme."

Understanding Bubbles

Before getting into whether there's an AI bubble — I've read a great deal for this — I'd like to clarify a concept. People keep asking, "Is there an AI bubble?" But the question itself is ambiguous. I think we should distinguish between two different but related bubble possibilities: one in the behaviour of companies within the industry, and one in the behaviour of investors toward the industry. I'm in no position to judge whether AI companies' current aggressive expansion is justified, so I'll focus mainly on the question of a bubble in financial markets around AI.

The main job of investment analysts — especially those of us from the "value investing" school — is to:

(a) study companies and other assets and assess their intrinsic value and prospects;

(b) make investment decisions based on that value. Most of the change analysts encounter in the short to medium term shows up in asset prices and their relationship to intrinsic value. And that relationship is, at bottom, the product of investor psychology.

Market bubbles aren't caused directly by technological or financial developments, but by excessive optimism about those developments. As I wrote in my January memo "On Bubble Watch," bubbles are temporary manias in which the developments in question become the object of what former Fed Chairman Alan Greenspan called "irrational exuberance."

Bubbles usually form around new financial developments (for example, the South Sea Company in the early 18th century, or subprime residential mortgage-backed securities in 2005–06) or technological advances (fibre optics in the late 1990s, the internet in 1998–2000). "Newness" plays a huge role. Without historical constraints, imagination can run wild, and the future of the new thing seems to have no ceiling. And when the future is believed to be limitless, people justify valuations far beyond historical norms, causing asset prices to become detached from their predictable earning power.

My favourite description of the role of "newness" comes from a book that influenced me enormously — John Kenneth Galbraith's A Short History of Financial Euphoria. Galbraith talks about what he calls "the extreme brevity of the financial memory," and notes that in financial markets, "past experience, to the extent that it is part of memory at all, is dismissed as the primitive refuge of those who do not have the insight to appreciate the incredible wonders of the present." In other words, history can limit our awe at the present and our imagination about the future; in its absence, anything is possible.

The key point: new things naturally inspire great enthusiasm, and bubbles are what happens when that enthusiasm swells to irrational levels. Who can draw the line between rational and irrational? Who can say when an optimistic market has turned into a bubble? In the end, it's simply a matter of judgement.

It occurred to me last month that my two most successful "calls" came in 2000 — when I warned about the market for tech and internet stocks — and in 2005–07, when I pointed to the dearth of risk aversion and the prevalence of crazy deals ahead of the Global Financial Crisis.

First, in neither case did I have any expertise in the subject of the bubble itself: the internet, or subprime mortgage-backed securities. All I did was observe the behaviour going on around me.

Second, the value of my calls lay mainly in describing the foolishness of that behaviour, not in insisting on labelling it a "bubble."

Getting hung up on whether to use the word "bubble" often lands people in trouble and gets in the way of sound judgement; in fact, simply by assessing what's going on around us and working out the appropriate behaviour, we can accomplish a lot.

The "Benefits" of Bubbles

Before going on to whether AI is in a bubble, I'd like to spend a moment on a topic that may seem rather academic for investors: the positive role of bubbles. You may feel I devote too much space to it, but I do so because I find it highly illuminating.

The November 5 Stratechery newsletter was titled "The Benefits of Bubbles." In it, Ben Thompson (no relation to Derek) cites a book called Boom: Bubbles and the End of Stagnation. Its authors, Byrne Hobart and Tobias Huber, propose two kinds of bubbles:

"Inflection bubbles" — the good kind, as opposed to the more destructive "mean-reversion bubbles," such as the subprime bubble of the 2000s.

I think this is a very useful distinction.

The financial manias I've read about or lived through — the South Sea Company, portfolio insurance, subprime mortgage-backed securities — all fired the imagination with fantasies of "high returns without risk," but no one thought they would advance humanity as a whole. No one expected the subprime movement to revolutionise housing, for example; people just thought they could make money backing new buyers. Hobart and Huber call these "mean-reversion bubbles" because they aren't expected to move the world forward. The fad simply rises and recedes.

By contrast, bubbles based on technological progress — like railroads and the internet — are what they call "inflection bubbles." After such a bubble, the world doesn't go back to where it was. In these bubbles, "investors decide that the future will be meaningfully different from the past and trade accordingly." As Thompson puts it, the definitive work on bubbles has long been Carlota Perez's Technological Revolutions and Financial Capital. When Perez wrote the book in 2002, bubbles were seen as negative, something to be avoided; the world was then in the downturn following the bursting of the dot-com bubble.

Perez didn't deny the pain bubbles cause. In fact, she pointed out that similar crashes marked every previous revolution, including the Industrial Revolution, railways, electricity and the automobile. In each case, the bubble was not a regret but a necessity: the speculative mania enabled what she calls the "installation period," during which large amounts of necessary but not necessarily wise investment laid the groundwork for the "deployment period" that followed. The bursting of the bubble marks the transition to the deployment period — and what underpins that period are precisely the investments that lost money.

This distinction matters a great deal to Hobart and Huber, and I agree. They note: "Not all bubbles destroy wealth and value. Some can be understood as important catalysts for technological progress." But I'd restate it this way: "mean-reversion bubbles" — in which markets soar on some new financial miracle and then collapse — destroy wealth; whereas "inflection bubbles" based on revolutionary technologies accelerate technological progress and lay the foundation for a more prosperous future, while also destroying wealth. The key is not to be the investor whose wealth is destroyed in the process of driving progress.

Hobart and Huber go on to describe how bubbles finance the infrastructure new technologies need, thereby accelerating their adoption:

Most new technologies don't appear out of nowhere, fully formed. They build on earlier failures, trial and error, iteration and historical path dependence. Bubbles create opportunities for the capital needed for deployment, making large-scale, parallel experimentation possible and thereby speeding up potentially disruptive technologies and breakthroughs.

By creating a positive "enthusiasm–investment" feedback loop, bubbles can produce net benefits. Optimism can become a self-fulfilling prophecy. Speculation provides the enormous funding high-risk exploratory projects need; enthusiasm that looks excessive in the short run, or investments that look bad, end up being the necessary starting point for social and technological innovation… A bubble may be a collective delusion, but it can also be the expression of a collective vision. That vision becomes a platform for coordinating people and capital, allowing innovation to proceed in parallel. Progress doesn't happen linearly but in bursts, simultaneously across many fields. As enthusiasm runs high, risk tolerance increases and network effects become ever more pronounced. FOMO draws in more participants, entrepreneurs and speculators, further reinforcing the positive feedback loop. Like bubbles, FOMO has a bad reputation, but sometimes it's a healthy instinct. After all, no one wants to miss a once-in-a-lifetime chance to "build the future."

In other words, bubbles based on technological progress are "beneficial" because they inspire investors to pour in vast sums — a good share of which will be wasted — "carpet-bombing" a new area of opportunity and thereby kick-starting its development.

The key point: if people always remained patient, cautious and rational, and stuck to value principles, building out a new technology might take many years, even decades; the frenzy of a bubble compresses that process into a very short time — a few winners earn life-changing investment returns, while vast sums are incinerated.

Bubbles have both a technological and a financial dimension, but the passages above are written largely from the standpoint of those eager for technological progress, who are perfectly willing to have investors lose money for it. "We," on the other hand, want technological progress but don't want to burn money for nothing in the course of driving it.

Ben Thompson ends his discussion: "This is why I'm excited about new technologies whose prospects are unknown." I love that line — it expresses both excitement about future possibilities and an acknowledgement that the shape of the future is unknowable (in our world, we'd say: the risk is extremely high).

Assessing the Current Landscape

Now let's get back to "seeking truth from facts." What do we know?

First, I've never met anyone who doesn't believe AI has the potential to be one of the most important technologies in history. It could reshape daily life and the global economy. We also know that, in recent years, the economy and markets have come to depend more and more on AI:

AI accounts for a very large share of corporate capital expenditure;

capex on building AI capacity has been a major source of US GDP growth;

AI stocks have contributed the great majority of the S&P 500's gains.

Fortune ran this headline on October 7:

75% of gains, 80% of profits, 90% of capex — AI's grip on the S&P is total, and Morgan Stanley's top analyst is "very concerned"

Beyond that, I think it's important to note that while AI-related stocks account for a disproportionate share of the gains across all stocks, the excitement AI has injected into the market has surely pushed up the prices of non-AI stocks as well.

The performance of AI stocks has been astronomical, led by the AI chip leader Nvidia. Founded in 1993, the company was valued at about US$626 million when it went public in 1999, and later briefly became the world's first company with a US$5 trillion market cap. That's an increase of roughly 8,000 times, equivalent to a compound annual return of about 40% for more than 26 years. No wonder imaginations have been set on fire.

Where Are the Uncertainties?

It's fair to say that while we know AI will bring enormous change, almost none of us know what exactly it will be able to do, how it will be commercialised, or on what timetable.

Who will the winners be? What will they be worth?

When a technology is believed to be world-changing, people also tend to assume the leading companies that own it must be enormously valuable. But how accurate is that assumption? Warren Buffett observed in 1999:

"The automobile was perhaps the most important invention of the first half of the 20th century… If you had foreseen in the days of the first cars how the car would shape this country, you would have said, 'This is where I must be.'

But of the roughly 2,000 car companies, only three survived a few years later. So the car had an enormous impact on America, but the opposite effect on investors."

— Time, January 23, 2012

In AI there are currently some extremely powerful leaders, including some of the richest and most dominant companies in the world. But new technologies have always been highly disruptive: will today's leaders win, or be displaced by latecomers? How costly will the arms race be? Who will win?

What are shares in start-ups worth?

Compared with leaders worth trillions of dollars, some potential challengers have enterprise values of just a few billion — or even a few million — dollars. CNBC reported on June 25, 2024:

A team founded by college dropouts has raised US$120 million from investors led by Primary Venture Partners, with plans to build a new AI chip to take on Nvidia.

Etched CEO Gavin Uberti said the company is betting that, as AI develops, most of the most energy-intensive computing needs will be met by customised, hard-wired ASIC chips.

"If transformers go away, we're dead," Uberti told CNBC. "But if they stick around, we're the biggest company of all time."

Even if Etched doesn't become "the biggest company of all time," suppose success would bring it a valuation of one-fifth of Nvidia's peak — US$1 trillion. What probability of success would you need to justify a US$120 million investment? For simplicity, assume that investment buys 100% of the equity; then as long as you believe there's a 0.1% chance of reaching a trillion-dollar valuation, the expected return is more than 8x. Who can say Etched doesn't have that chance? And if so, who wouldn't take part? This is what I call "lottery-ticket thinking": the fantasy of a huge payoff is enough — indeed, bound — to drive people into a venture with a very high probability of failure.

There's nothing wrong with this way of calculating expected value in itself; top venture capitalists use it every day. But the premise is that the assumptions about the potential payoff and the probability of success must be reasonable. Once you're thinking about "trillion-dollar payoffs," any calculation can easily be overwhelmed by that fantasy and lose its reasonableness.

Will AI produce profits? Who will they go to?

We know almost nothing about two things:

How much profit will AI bring its providers?

What will its impact be on non-AI companies (the companies that use AI)?

Will AI become a monopoly or duopoly, letting one or two leading companies charge high prices for its capabilities? Or will it turn into a fiercely competitive free-for-all in which many companies wage price wars for users' AI spending, turning it into a commodity? Or, more likely, will the industry take a hybrid "giants + specialists" form, with some companies competing on price and others on proprietary advantages? Some say the services that answer AI queries today (such as ChatGPT and Gemini) lose money on every query (of course, "loss leaders" are not unusual in new industries). Will the leading tech companies — used to winning in winner-take-all markets — be willing to lose money on AI for years in exchange for share? Hundreds of billions of dollars are being poured into the race for AI leadership. Who will win? What will the outcome be?

Likewise, how will AI affect the companies that use it? Clearly AI is a powerful tool for raising users' productivity, including by replacing employees with labour and intelligence supplied by computers. But will this ability to cut costs lift users' profit margins? Or will it set off price wars among those companies as they fight for customers? If the latter, the cost savings may be passed on to customers rather than kept by the companies as profit. In other words: is it possible that AI raises corporate efficiency without raising corporate profitability?

Should we worry about so-called "circular deals"?

In the telecom boom of the late 1990s, fibre was overbuilt, and the companies that owned it traded with one another so that they could report profits. If two companies each owned fibre, it was just an asset on their books; but if each bought capacity from the other, both could report profits… so that's what they did. In other cases, equipment makers lent money to network operators to buy their equipment, when the operators didn't yet even have the customers to justify the build-out. All of this led to illusory profits.

Today, some deals have come to light that give the impression of money "going round in a circle" among AI players. Those who believe there's an AI bubble can easily view these deals with suspicion: are they aimed at achieving genuine business objectives, or at exaggerating progress?

Adding to the concern, critics argue that some of the deals OpenAI has struck with chipmakers, cloud computing companies and others form a strange loop: OpenAI receives billions of dollars from tech companies, then pays billions back to those same companies for computing power and other services…

Nvidia, too, has done deals that raise the question of whether the company is paying itself. It announced it would invest US$100 billion in OpenAI. The start-up receives this money and, at the same time, uses it to buy or lease Nvidia's chips…

Goldman Sachs estimates that 15% of Nvidia's sales next year will come from what critics likewise call "circular deals." (The New York Times, November 20)

Notably, even though OpenAI isn't yet profitable, it has made investment commitments to industry partners totalling US$1.4 trillion. The company has said explicitly that these investments will come from revenue paid to OpenAI by those same partners, and that OpenAI also has ways out of these commitments. But all of this raises the question: has the AI industry developed a "perpetual motion machine"?

On this subject, I've always enjoyed articles questioning whether humans can really grasp the word "trillion." I think the idea is spot on: a million dollars is US$1 a second for 11.6 days; a billion dollars is US$1 a second for 31.7 years. We can grasp that. But a trillion dollars is US$1 a second for 31,700 years. Who can really fit the meaning of 31,700 years into their head?

How long will AI assets last?

We can't help wondering whether the question of "depreciation/obsolescence" is being handled correctly in the AI world. How long will AI chips last? In estimating P/E ratios for AI-related stocks, how many years of earnings growth should be assumed? Will chips and other AI infrastructure last long enough to repay the debt taken on to buy them? Will "artificial general intelligence" (a machine that can do anything the human brain can) be achieved? If it is, does that mark the end of progress, or will there be further revolutions? Which companies will win? Will businesses reach a stage where the technology stabilises and they can extract economic value from it? Or will new technologies keep threatening to replace old ones as the new path to success?

On this point, two items mentioned in a Financial Times newsletter hint at how fluid the competitive landscape is:

A study by MIT and the open-source AI start-up Hugging Face found that newly released Chinese open-source models rose to 17% of downloads over the past year. That exceeds the 15.8% share of US developers such as Google, Meta and OpenAI — the first time Chinese teams have overtaken their American counterparts…

Nvidia's shares fell sharply yesterday on fears about Google's progress in AI, wiping US$115 billion off its market value. (FirstFT Americas, November 26)

Dynamic change creates opportunities for breakthrough technologies, but the same dynamics can also threaten the dominance of leading companies. Amid all this uncertainty, investors must ask: are the assumptions of "continued success" implicit in the prices they pay fully tenable?

Is the frenzy leading to speculative behaviour?

To take an extreme example: in venture capital there's a trend toward US$1 billion "seed rounds" for start-ups. Here's a little anecdote:

Thinking Machines, the AI start-up led by former OpenAI executive Mira Murati, just completed the largest seed round in history: US$2 billion at a US$10 billion valuation. The company hasn't released any product and declined to tell investors what exactly it plans to do. "It was the most absurd pitch meeting I've ever seen," said one investor who met with Murati. "She was basically saying, 'We're going to build an AI company with the best AI talent, but we can't answer any questions.'" ("How the AI Bubble Will Pop," Derek Thompson's Substack, October 2)

But that's already "ancient history"… a mere two months ago. Here's the latest:

Thinking Machines Lab, the artificial intelligence start-up founded by former OpenAI executive Mira Murati, is in early talks on a new funding round at a valuation of about US$50 billion, Bloomberg reported on Thursday. The company was valued at US$12 billion when it raised about US$2 billion in July. (Reuters, November 13)

And Thinking Machines Lab isn't alone:

In one of the boldest bets of the AI arms race, the stealth start-up Safe Superintelligence (SSI), founded by former OpenAI chief scientist Ilya Sutskever, has raised US$2 billion at a US$32 billion valuation — even though it hasn't publicly released any product or service. (CTech by Calcalist, April 13)

Where will it all end up?

One problem with AI is the unusual nature of its "newness." It isn't like a traditional business, which designs and sells a product and makes money as long as the price exceeds the cost. It's more like a group of companies "building the plane while flying it": only when the plane is built will they know what it can actually do, and whether anyone will pay for its services.

Many companies justify their spending because they aren't building a product but creating something world-changing: artificial general intelligence (AGI)… The problem is that none of them really knows how to get there.

Still, University of Virginia economist Anton Korinek says that if Silicon Valley achieves its goal, all of this spending will be justified. He's optimistic that it can.

"This is a bet on A.G.I. or bust," Korinek said. (The New York Times, November 20)

This quality of "an industry being built whose nature hasn't been determined" is best captured by a paraphrase of OpenAI CEO Sam Altman: "We'll first build a system that is generally intelligent, and then have it figure out how to generate an investment return for us."

For investors who have always invested in businesses whose "models are fully understandable," this ought to be a warning. A technology equal to or surpassing the human brain should obviously be worth a great deal — but is it worth so much that it's beyond any calculation?

A Note on the Use of Debt

So far, investment in AI and its supporting infrastructure has come mainly from equity capital generated by operating cash flow. But now companies are making commitments that require debt financing, and for some of them, the investment and leverage are fairly described as aggressive.

The AI data centre boom can't be financed with cash alone. The projects are too big to pay for out of pocket. JPMorgan analysts did some rough estimates (presumably on the back of a napkin, or a tablecloth) and concluded the total bill for the infrastructure build-out will reach US$5 trillion (not including the tip). Who knows if that's accurate, but we have every reason to expect spending next year to approach US$500 billion. Meanwhile, the biggest spenders (Microsoft, Alphabet, Amazon, Meta, Oracle) had only about US$350 billion in cash between them at the end of the third quarter. (Unhedged, Financial Times, November 13)

All of these companies rely on their powerful non-AI businesses to generate abundant cash flow. But AI's winner-take-all arms race is forcing some of them to start borrowing. In fact, one could reasonably argue that one reason they're spending so heavily is to make it hard for smaller companies to keep up.

Oracle, Meta and Alphabet have issued 30-year bonds to fund AI investment. For the latter two, the bond yields are less than 100 basis points above US Treasuries of the same maturity. To earn a little more than a risk-free bond, would you be willing to bear 30 years of technological uncertainty? Will the chips and data centres bought with debt stay productive long enough to repay these 30-year obligations?

On November 14, Alex Kantrowitz's Big Technology Podcast aired a conversation with Gil Luria, head of technology research at D.A. Davidson, focused mainly on the use of debt in the AI industry. Here are some of Luria's points:

Healthy behaviour is being practised by "…rational, thoughtful business leaders at Microsoft, Amazon, Google," who are making sound investments in expanding their capacity to deliver AI. They can afford to be sound because they have customers… They're using cash from their balance sheets, backed by strong cash flow; they understand these are high-risk investments, and they've struck a balance.

Unhealthy behaviour: he describes "…a start-up borrowing money to build data centres for another start-up. Both are burning huge amounts of cash, yet they can still raise debt capital to drive the build-out, while they have neither customers nor visibility into a return on investment."

"There's a whole spectrum of behaviour between healthy and unhealthy, and we need to tell them apart so we don't repeat past mistakes."

"Some things are suited to equity financing (you own the growth), and some to debt financing (you take on an obligation to pay interest on schedule). For a long time, society has been clear about the division of labour between the two: when you have predictable cash flows and/or assets that can serve as collateral, debt makes sense… and when you're investing in something more speculative, pursuing growth with uncertain cash flows, equity is how a normal economy works. Once you mix the two up, things go wrong."

Luria also pointed to some potential causes for concern:

"A speculative asset… we don't know how much of it we'll actually need two to five years from now."

Lenders are incentivised to "get loans done" but don't bear the long-term consequences.

The possibility that AI capacity supply catches up with, or even exceeds, demand.

The risk that more powerful next-generation AI chips make existing chips obsolete or reduce their value as collateral.

Strong competitors cutting prices and operating at a loss to grab share, pushing down data centre rents.

Here are the key passages from Azeem Azhar's October 18 Exponential View:

When does an AI boom tip into a bubble? [Investor and engineer] Paul Kedrosky points to the "Minsky moment" — the inflection point at which credit expansion exhausts the good projects and starts chasing bad ones: funding marginal deals with vendor financing and questionable coverage ratios. For AI infrastructure, that shift may already be under way; the telltale signs include hyperscalers' capex outrunning their revenue momentum, and lenders sweetening terms to keep the party going.

Paul's argument is persuasive. We've already entered the realm of speculative finance — perhaps even moved past the testing stage — and recent deals will set dangerous precedents. As Paul warns, such financing will "provide a template for future similar deals," driving a surge in junk bond issuance and SPVs (special purpose vehicles) as hyperscalers fight for dominance at any cost…

For AI infrastructure, the warning lights are flashing: vendor financing is spreading, coverage ratios are thinning, and hyperscalers are using their balance sheets to keep up the pace of capex even as revenue momentum lags. We're seeing both sides: genuine infrastructure expansion, and financial gymnastics reminiscent of the telecom crash of 2000. The boom may still prove productive, but only if revenue catches up before credit tightens. When does healthy strain become systemic risk? That's the question we must answer before the market answers it for us.

Azhar mentions off-balance-sheet financing through SPVs, which was one of the key drivers of Enron's fragility and eventual collapse. A company and its partners set up an SPV for a specific purpose and provide equity capital. The parent may have operating control but, because it doesn't hold majority ownership, needn't consolidate the SPV. The SPV borrows, but the debt doesn't appear on the parent's books. The parent may be an investment-grade borrower, but the debt is neither its obligation nor guaranteed by it. Today, an SPV's debt may be backed by rent promised by a data centre's tenant — who is sometimes also an equity partner — but the debt isn't the tenant's direct obligation. In short, SPVs let a company look as if it isn't doing what the SPV is doing and doesn't carry the debt the SPV carries. (Private equity funds and private credit funds are very likely to appear in these structures as partners and lenders.)

As Perez, whom I quoted earlier, put it (writing after the dot-com bubble burst): "What underpins the deployment period are the investments that lost money." Early investments get consumed in the "Minsky moment": unwise commitments made during a long upswing meet value destruction in the correction. On debt, there are three things we know for certain:

If losses occur, debt magnifies them (just as it magnifies the hoped-for gains);

Debt raises the probability that a project fails if it runs into difficulty;

Even with an equity cushion beneath the debt, if the difficulty is severe enough, lenders' principal will still be at risk.

One risk that deserves serious consideration: the data centre building boom could lead to oversupply. Some data centres may become uneconomic, and some owners may go bankrupt. At that point, a new generation of owners may buy the data centres from foreclosing lenders for "pennies on the dollar" and profit once the industry stabilises. This is one of the ways "creative destruction" pulls a market back into equilibrium, pushing costs down to levels that can be profitable in the future.

Debt in itself is neither good nor bad. Likewise, the use of leverage in the AI industry should be neither simply celebrated nor feared. Everything depends on: the proportion of debt in the capital structure; the quality of the assets or cash flows you're lending against; the borrower's alternative sources of liquidity; and whether lenders get an adequate margin of safety. We'll see which lenders can stay disciplined in today's dizzying environment.

It's worth mentioning that Oaktree has made a small number of investments in data centres; our parent company, Brookfield, is raising a US$10 billion fund to invest in AI infrastructure. Brookfield will put in its own money, has equity commitments from sovereign wealth funds and Nvidia, and plans to use "prudent" debt on top of that. Brookfield's investments seem likely to go mainly to regions less saturated with data centres, and toward building the infrastructure that supplies data centres' enormous power needs. Of course, we all believe we're doing these things on the basis of prudent decisions.

I know I don't know enough about AI to comment on it. But I do know a little about debt. My view is:

It's fine to provide debt financing for a venture whose outcome is uncertain;

But it's not fine if the outcome is pure speculation;

Those who understand the difference still have to get the judgement right in practice.

The Financial Times's Unhedged quotes Chong Sin, JPMorgan's head of CMBS research: "…in our conversations with investment-grade ABS and CMBS investors, one frequently raised concern is whether they're willing to take on the residual value risk of data centres when the bonds mature." I'm glad potential lenders are asking the questions they should be asking.

Here's how Bob O'Leary, Oaktree's co-CEO and co-portfolio manager of our Opportunities funds, sees "the intersection of debt and AI":

Most technological advances eventually evolve into winner-take-all or winner-take-most competition. Under those dynamics, the "right" way to participate is through equity, not debt. As long as you can diversify your equity exposure and include the eventual winner, the winner's huge gains will far more than make up for the capital lost on the losers. That's the classic formula for venture capital success.

Diversified debt exposure is just the opposite: even on the winners you only collect interest coupons, which are nowhere near enough to offset the losses you suffer on the losers' debt. Of course, if you can't even identify the group of companies from which the winner will emerge, the distinction between debt and equity is moot — you get "zero" on both. I mention this because it's exactly what happened in search and social media: early front-runners (Lycos in search, MySpace in social) were crushed by companies that came along later (Google in search, Facebook in social).

Trying to Reach a Conclusion

There's no doubt that today's behaviour is "speculative" — that is, based on guesses about the future. And there's no doubt that no one knows what the future will bring, yet investors are betting enormous sums on it.

Here I'd like to talk about AI's unique nature. The AI revolution differs from previous technological revolutions, and the ways it differs are both exhilarating and worrying. To me, it's like a genie let out of the bottle that won't go back in:

AI may be not just a tool for humans but more like a substitute. It may take over cognitive abilities that were once a human monopoly. So it may differ from previous developments not just "in degree" but "in kind." (I'll say more about this in the postscript.)

AI technology is advancing at astonishing speed, possibly leaving humans almost no time to adjust. I'll give two examples:

Coding (60 years ago we called it "computer programming") is like the canary in the coal mine when it comes to AI's impact. In many advanced software teams, developers no longer write code; they type in requirements and AI systems generate the code for them. AI coding has reached world-class level, which wasn't the case a year ago. As my guide here put it: "In that vertical, there's no longer anything speculative about whether human replacement is happening."

In digital advertising, when a user logs into an app, AI performs "ad matching," showing the user ads based on their earlier browsing preferences. Humans are no longer needed for this work.

Perhaps most important: growth in AI demand seems completely unpredictable. One of my young advisers explained: "The speed and scale of improvement mean it's hard to predict AI demand. Today's adoption may have nothing to do with tomorrow's, because in a year or two AI may be able to do 10 or 100 times what it can today. In that case, who can say how many data centres will be needed? And how would even a successful company know how much computing capacity to contract for?"

Given these differences, who can correctly judge what AI means for the future?

At this point, one thing many observers (including me) are doing is looking for analogies with past bubbles. Here's the historical perspective offered by a recent Wired article:

AI's closest historical analogue here may not be electric light but radio. When RCA began broadcasting in 1919, people immediately realised it held a powerful information technology. What was less clear was how that would translate into a business. "Would radio be a loss leader for department stores? A public service for broadcasting Sunday sermons? Or an advertising-supported entertainment medium?" wrote Brent Goldfarb and David A. Kirsch of the University of Maryland. "All of these were possible. They were all part of the technological narrative." As a result, radio turned into one of the biggest bubbles in history — peaking in 1929 and then losing 97% of its value in the crash. This wasn't a fringe industry; RCA, along with the Ford Motor Company, was among the most actively traded stocks on the market at the time. As The New Yorker recently put it, it was "the Nvidia of its day"…

In 1927, Charles Lindbergh made the first solo non-stop transatlantic flight from New York to Paris… It was the greatest technological demonstration of its time, and became a huge coordinating event on the scale of the ChatGPT launch — signalling to investors: pour money into this industry.

Goldfarb and Kirsch write that "professional investors did correctly understand the importance of aeroplanes and air travel," but "the narrative of inevitability largely drowned out their caution. Technological uncertainty was framed as opportunity rather than risk. The market overestimated how quickly the industry would reach maturity in technical feasibility and profitability." As a result, the bubble burst in 1929 — from their peak that May, aviation stocks fell 96% by May 1932…

It bears repeating: in the history of tech bubbles, AI's two closest analogues seem to be aviation and radio. Both were full of deep uncertainty, and both were wrapped in extraordinarily powerful coordinating narratives. In both, investors seized on "pure play companies" in hopes of profiting from a game-changing new technology, and both were accessible to the retail investors of the day. Together they helped inflate a huge bubble, so that when it burst in 1929, the Great Depression followed. ("AI Is the Bubble to Burst Them All," Brian Merchant, Wired, October 27 — emphasis in original. Note: the Great Depression had many causes; the bursting of the radio/aviation bubble was only one of them.)

Derek Thompson (the writer I quoted at the beginning) ends his newsletter with a wonderful historical perspective:

The railroads were a bubble, but they transformed America. Electricity was a bubble, but it transformed America. The broadband build-out of the late '90s was a bubble too, but it transformed America. I don't wish for a bubble; on the contrary, I hope the US economy goes many years without another recession. But given the amount of debt flowing into AI data centre construction today, I think it's unlikely that AI will be the first technology to be transformative without being overbuilt and without going through a brief, painful correction. ("AI Could Be the Railroad of the 21st Century. Brace Yourself," November 4)

Sceptics can easily list the similarities between today and the dot-com bubble:

A world-changing technology

Euphoric, speculative behaviour

The role of FOMO

Questionable circular deals

The use of SPVs

US$1 billion seed rounds

Supporters offer reasons why "this time is different":

There are already products, with strong demand

There are already 1 billion users (far more than internet users at the height of the bubble)

The main players are long-established and have revenue, profits and cash flow

There's no IPO frenzy (stocks doubling on their first day)

The P/E ratios of established players are reasonable

I'd like to say a bit more about the first "difference." Unlike in the dot-com era, AI products already exist at scale, demand for them is exploding, and they're generating rapidly growing revenue. For example, Anthropic (described as one of the two leaders in AI coding models; see page 12) has reportedly grown revenue 10x a year for each of the past two years (for those who didn't study advanced maths, that's 100x in two years). Claude Code, the coding program Anthropic launched earlier this year, is reportedly already at a US$1 billion annualised revenue run rate. Another leader, Cursor, had revenue of US$1 million in 2023 and US$100 million in 2024, and is also expected to reach US$1 billion this year.

As for the last point, "reasonable P/E ratios," see the table below (from Goldman Sachs, via Derek Thompson). You'll notice that during the 1998–2000 dot-com bubble, Microsoft, Cisco and Oracle had far higher P/E ratios than today's major AI players — Nvidia, Microsoft, Alphabet, Amazon, Meta (OpenAI has no earnings). In fact, compared with 26 years ago, Microsoft's P/E today looks like it's on sale at half price! And in the first bubble I lived through — the "Nifty Fifty" of 1969–72 — the leading companies' P/E ratios were even higher than in 1998–2000.

Closing

For my final quotation, I'll return to OpenAI's Sam Altman. His comment seems to me to capture the essence of the current situation:

"When bubbles happen, smart people get overexcited about a kernel of truth," Altman told reporters this year. "Are we in a phase where investors as a whole are overexcited about AI? I think so. Is AI the most important thing to happen in a very long time? I think so too." (The New York Times, November 20)

So do I have a "bottom line"? I do. Greenspan's phrase (mentioned earlier) is perfect for summing up stock market bubbles: "irrational exuberance." There's no doubt investors are applying "exuberance" to AI. The question is whether it's "irrational." Given AI's enormous potential and equally enormous unknowns, I think almost no one can answer with certainty. We can speculate about whether today's enthusiasm is excessive, but we'll only know years from now. Bubbles are easiest to identify in hindsight.

While the similarities to past bubbles are almost impossible to ignore, technology believers will say, "This time is different." Those four words show up in nearly every bubble, to explain why this time isn't like those similar past episodes that were bubbles. On the other hand, when Sir John Templeton reminded me of that phrase in 1987, he was quick to point out that 20% of the time, things really are different. But from a third angle, we must remember: it's precisely behaviour based on the belief that "this time is different" that causes "this time" to be no different!

The current situation reminds me of something the American economist Stuart Chase (reportedly) said about faith. I think it applies to AI as well (and to gold and crypto):

For those who believe, no proof is necessary.

For those who don't believe, no proof is possible.

Here is my real bottom line:

Transformative technologies consistently inspire excessive enthusiasm and overinvestment, leading to more infrastructure than is needed and to asset prices that prove too high. It's these excesses that accelerate the technology's adoption — without them, adoption wouldn't be nearly so fast. The usual word for these excesses is "bubble."

AI may become one of the greatest transformative technologies in history.

As described above, AI is currently the object of great enthusiasm. If that enthusiasm doesn't end up producing a bubble that fits the historical pattern, it will be the first time.

The bubbles that form in this process usually end in losses for those who fuel them.

The main reason for the losses: the technology's newness makes the scope and timing of its impact highly unpredictable. That makes it easy, amid the enthusiasm, to be overly positive about companies, and hard to identify the winners before the dust settles.

If you want to fully participate in a new technology's potential gains, you can't avoid exposure to the risk of losses brought on by "excessive enthusiasm and excessive behaviour."

The use of debt in this process — which in past technological revolutions was often inappropriate because of the high uncertainty — could amplify all of the above this time.

Since no one can say definitively whether there's a bubble, I'd suggest that no one should go all-in without acknowledging that they could face ruin if things go badly; but neither should anyone stay out entirely and risk missing one of the great leaps in the history of human technology. The best approach seems to be a moderate position, taken selectively and prudently.

Finally, we must remember: there are no "magic words" in investing. Today, people promoting real estate funds say, "Office buildings are yesterday's news, but we're investing in the future through data centres." And everyone nods in agreement. But data centres could be in short supply or in surplus; rents could rise more than expected or fall unexpectedly. So they may be profitable… or they may not. Wise investing in data centres — and therefore in AI — requires, like all investing, cool, deep judgement and a high level of execution.

December 9, 2025

P.S.

What follows has nothing to do with financial markets or with whether AI is a bubble. My topic is AI's impact on society through job losses and the loss of a sense of purpose. You don't have to read it — that's why it's a postscript — but it matters to me, and I've been looking for a suitable place to say a few words.

On November 18, a Barclays research report said Fed Governor Christopher Waller "emphasised that the stock market's recent enthusiasm for AI has yet to translate into job creation." Given that I believe one of AI's main effects will be to raise productivity and thereby eliminate jobs, that sentence struck me as rather contradictory. And that's the source of my worry.

I see AI mainly as an incredible labour-saving device. Joe Davis, Vanguard's global chief economist and head of its Investment Strategy Group, says: "For most jobs — perhaps four out of five — AI's impact will be a mix of innovation and automation, and could save about 43% of the time people currently spend on work tasks." (Exponential View, September 3)

I find the outlook for employment frightening. I'm extremely worried about what will happen to people displaced by AI, or who can't find work because of AI. Optimists say, "New jobs have appeared after every past technological advance." I hope that will be true of AI too, but hope isn't something to rely on, and I find it hard to imagine where those jobs will come from. Of course, I'm not a futurist, nor a financial optimist — which is one of the reasons I moved from stocks to bonds in 1978.

Optimists also say, "AI's positive effect on productivity will make GDP grow much faster." I have some specific doubts about this:

The change in GDP can be thought of as the change in hours worked × the change in output per hour (that is, "productivity"). AI raising productivity means it will take fewer hours of work — that is, fewer workers — to produce the goods we need.

Or, looked at from the other direction: a productivity boom means the same labour force can produce more goods. But if AI eliminates large numbers of jobs, how will people afford the extra goods AI enables us to produce?

I find it hard to picture a world in which AI works side by side with everyone who is employed today. How could employment not fall? AI may well replace large numbers of entry-level employees, people who process paperwork without exercising judgement, and junior lawyers who comb through statutes looking for precedents. It may even replace junior investment analysts who build spreadsheets and put together presentations. AI is said to read MRIs better than the average doctor. Driving is one of the most common occupations in America, and driverless vehicles are already arriving; where will the people who now drive taxis, limousines, buses and trucks go to work?

I imagine governments will respond with what's called "universal basic income": the government sends cheques to millions of people without jobs. But the worrier in me sees problems here too:

Where will the money for the cheques come from? The job losses I foresee mean less income tax revenue and more spending on benefits, which will further burden the part of the population still working and mean bigger fiscal deficits. In this new world, will governments be able to afford ever-larger deficits?

More important, people get far more from work than a paycheque: work gives them a reason to get up in the morning, gives structure to their day, gives them a productive role in society and self-respect, and provides challenges; overcoming challenges brings satisfaction. How will these things be replaced? I worry about large numbers of people living on basic cheques alone with nothing to do all day. I worry that there's a correlation between the loss of mining and manufacturing jobs in recent decades and opioid addiction and shortened lifespans.

By the way: if we eliminate large numbers of junior lawyers, analysts and doctors, where will the senior experts come from — the ones who need decades of training in judgement and pattern recognition to solve complex problems?

Which jobs won't be eliminated? What careers should our children and grandchildren prepare for? Think about the jobs machines can't do. My list starts with plumbers, electricians and massage therapists — tasks that require physical work. Perhaps nurses will earn more than doctors, because they provide hands-on care.

What do artists, athletes, doctors, lawyers and, I hope, the best investors have in common? I think it's "talent or insight," which AI may or may not be able to replicate. But how many people are needed at the top of these professions? A presidential candidate once said he'd give a laptop to everyone who lost their job to outsourcing. How many "laptop operators" do we actually need?

Finally, I worry that a small number of highly educated billionaires living on the coasts will be seen as having created the technology that put millions of people out of work. That could cause social and political divisions even worse than those we have today, making the world more fertile ground for populist demagogues.

I've witnessed astonishing progress in my lifetime, but in many ways I miss the simpler world I grew up in. I fear this will be yet another enormous change. I'm not happy writing this. Can the optimists explain why I'm wrong?

Interestingly, Vanguard's Joe Davis notes that more Americans will turn 65 in 2025 than in any previous year, and that about 16 million baby boomers will retire between now and 2035. Might AI simply fill that gap? That would be an optimistic interpretation.

HM