The Foam and the Fortune

In January 2025, a Chinese artificial intelligence company named DeepSeek wiped nearly $600 billion off Nvidia’s market capitalization in a single day of trading.¹ Wall Street, it seemed, had decided that intelligence could be made more cheaply than anyone had thought. A week later, investors were entertaining the opposite conclusion: cheaper intelligence might mean the need for more chips, not fewer.

Both arguments seemed reasonable, yet neither was entirely satisfying.

The second argument rested on an idea from William Stanley Jevons, the nineteenth-century economist who observed that using coal more efficiently might expand the demand for coal rather than reduce it.² Applied to artificial intelligence, the idea looked convincing: the more efficiently computation is used, the more reasons there are to compute.

Maybe. But the argument described a possibility, not a law of nature. Efficiency gains could increase consumption without preserving suppliers’ profits. The demand for intelligence could be endless; the demand for a particular corporation’s chips would not be. A growing market could make its participants poorer.

This distinction was easily lost in the excitement that followed.

By October 2026, the question had acquired an expensive context. Investors were no longer speculating only about how much intelligence the world would consume, but also about whether anyone could earn enough money supplying it. Billions in investment, data centers, specialized chips and ever more sophisticated financing arrangements had turned a technological proposition into an economic challenge.

So the familiar question returned: Is AI a bubble?

It is the question people ask whenever prices rise faster than anyone can work out what they measure. But I suspect it is the wrong question.

Before deciding whether there is a bubble, we should first decide what we mean by one.

Imagine three beverages.

I. Coffee: The Virtue of Excess

Suppose we meet on a leisurely afternoon. Empty cups are set in front of us. I ask you what you would like to drink.

You choose coffee, saying you need something to energize you. I pour you a cappuccino with plenty of milk foam on top. You taste it and wipe a little froth from your upper lip. You are perfectly satisfied.

No one is likely to accuse the barista of deception. The foam is part of the drink; it adds to the pleasure. Remove the foam and you have not improved the coffee. You have simply made a worse cappuccino.

Some economic bubbles work the same way.

Speculation has a poor reputation, mostly because its failures are so visible. But an economy without speculation would struggle to finance projects it does not yet understand.

Consider Britain’s railway mania in the nineteenth century. Investors poured extraordinary amounts of capital into proposed railways, many of which never looked promising at all. Demand was exaggerated. Shares traded at prices that future profits could hardly justify. When the mania ended, investors suffered heavy losses.

But the railways did not vanish along with the money.

The track was laid. Cities were connected. Transport costs fell. Commerce reorganized itself around the new infrastructure, built by the overenthusiasm of many private investors.

Something similar happened after the dot-com bust. Some internet companies were wildly overvalued; some were silly businesses. Yet the collapse in their share prices did not deprive them of commercial viability.

There is a difference between capital that has been wasted and capital that simply failed to return a profit to the people who put it up. Excess construction may bankrupt its builders, but it will enrich everyone who comes afterwards.

This is not a plea for unlimited waste. Waste is waste. But the history of technological progress shows that the cost of discovering an economic frontier often falls on the people who discover it.

The artificial intelligence industry is one of the most vivid examples.

According to Stanford’s 2025 AI Index, the price of querying a model that performs at the level of GPT-3.5 fell from $20 per million tokens in November 2022 to seven cents in October 2024, a drop of more than 280-fold.³

The transformation is not only economical. It is economic.

Tasks that were too costly to automate become affordable. Services that once required specialized labor become available to small businesses. Whole categories of products emerge once the marginal cost of intelligence falls below the value it can add to a decision.

The economic promise does not lie in every company in the field making a profit. It lies in creating capabilities that become abundant enough to change economic activity completely.

Why does the distinction matter? Because investors tend to confuse technological inevitability with commercial inevitability. The first is inevitable; the second is not.

The internet was inevitable. The survival of Pets.com was not.

The artificial intelligence industry can revolutionize the economy while destroying enormous amounts of shareholder value. The eventual winners need not look like today’s market leaders. The companies that build the infrastructure making the revolution possible may not be the companies that benefit from it.

A healthy speculative boom finances possibilities before they are realized. It gives engineers room to build things the accountants do not yet understand.

Sometimes excess is part of the achievement.

At times, the foam is part of the coffee.

II. Wine: When Defects Become Features

We meet again, this time at a formal dinner party. A round table stands between us. Polite conversation circles the table. You decide the occasion calls for red wine.

I pour you a glass.

A few bubbles rise through the wine. They gather briefly at the surface and vanish.

You lift the glass, admire its legs and say something about a hint of oak.

But the bubbles deserve a closer look.

Bubbles in a still wine can indicate residual carbon dioxide or unintended fermentation. They do not necessarily mean the wine is bad; they are not a mark of superior craftsmanship either. They may simply be a side effect of a process that was never properly finished.

The danger comes when the side effect is mistaken for the product.

This is a different kind of economic bubble: not excessive investment in something valuable, but the promotion of something economically insignificant to the status of something valuable.

The confusion is possible because technological progress and commercial progress are different things.

An AI system can be extremely advanced and still have no economic value. It can improve on benchmarks while making no difference to a business. It can delight users without making them willing to pay for that delight. It can automate a task and still make the task more expensive.

There is a difference between demonstrating a technology and deploying it. There is a difference between deployment and adoption. There is a difference between adoption and revenue. And there is a difference between revenue and profit, as generations of failed businesses have clearly demonstrated.

Yet today’s AI industry has become remarkably skilled at treating all these steps as if they were the same.

A company announces that its agents can now do a job that humans used to do. Investors conclude that labor costs will fall. The announcement becomes a valuation. Somewhere between the demo and the spreadsheet, the inconvenient details disappear.

Who checks the output? Who is responsible when it is wrong? How much human supervision does it need? How often does the system have to be corrected? What happens when a case calls for an exception rather than the routine?

An AI agent that can automate 90 percent of a workflow may be revolutionary. Or it may hand the remaining 10 percent back to people, and that 10 percent accounts for 90 percent of the difficulty. The distinction matters.

A customer support system, for example, may handle easy requests at negligible cost while badly mishandling complicated ones. Its cost savings may hide the fact that escalations become more frequent, customer relationships deteriorate, or the company runs into costly compliance problems.

Or consider a company that replaces a simple software process with an assembly of AI agents, an orchestration layer, monitoring tools and inference calls. The system may be more intelligent, but it consumes more resources to produce exactly the same result.

We have made the system smarter. We have also made the process less efficient.

This kind of failure takes two especially telling forms. The first offers the customer no new value, only technological innovation dressed up as economic necessity. The second does create value, but undermines the economics of providing it. Users get a better experience; meanwhile, the company is left with a problem that has become more expensive.

Both kinds of innovation can attract investment and produce compelling demonstrations. Neither guarantees a viable business.

The second is the more dangerous, because its product can be genuinely outstanding. A business can have delighted customers, fast-growing adoption and a completely broken economic model. Subsidies can make almost anything popular. What matters is whether the popularity survives once someone else has to pay the cost.

That is why falling inference costs are both good news and bad news. Lower costs widen the range of useful applications while eroding the scarcity on which many business models were built. When a capability becomes cheap and widely available, its economic value moves away from the producers and toward distribution, customer relationships, proprietary workflows, or the ability to turn intelligence into action.

It becomes indispensable to society and interchangeable to its customers. There is no contradiction here. These are two recurring features of capitalism: technological importance and economic importance are not the same thing.

No one can build a durable business simply by being fashionable during an economic transformation. A company has to explain why customers will stay when competitors offer exactly the same service, when the subsidies stop and when returns are finally expected. Otherwise, progress may turn out to be nothing more than an unfinished fermentation, in a market so intoxicated by novelty that it will accept even the flaws.

III. Beer: The Problem of Proportion

Our last meeting takes place in Guangzhou, on the bank of the Pearl River. The restaurant has no tablecloths. Plastic chairs crowd the sidewalk. Someone shouts an order to the kitchen. The air is thick with the smell of grilled seafood, tobacco and the damp warmth of the river.

I ask what you would like to drink. You answer without hesitation: Pearl River beer. I open a bottle. Foam runs over my fingers. I pour too fast, and the plastic cup fills with a mix of beer and foam. You laugh, raise the cup and drink. You insist this is the proper way to serve beer. Perhaps you are right.

Some people pour carefully to minimize the foam; others pour straight into the middle of the glass and let it explode. Neither approach is wrong. Foam adds aroma and texture; it is part of the experience of drinking beer. Tastes differ. But a glass that is 90 percent foam and 10 percent beer is not a masterpiece of brewing. It is deception disguised as abundance.

This is precisely the problem with the AI bubble question. People discuss the industry as if it were a single glass. It is not. It is a collection of different markets, each with its own economics, competitive landscape, technological limits and gap between expectation and value. A semiconductor maker selling scarce hardware into a market of intense demand is economically different from an AI infrastructure provider spending heavily to protect a rapidly commoditizing product. Neither resembles an enterprise software company embedded in its customers’ proprietary workflows, or a consumer app buying users with promotional subsidies, even though all of them belong to the same industry. One is serving beer; another is selling foam.

So to say that AI is a bubble is an incomplete thought. To say that it is not is equally incomplete. The real question is one of proportion. How much of the price reflects value already created? How much reflects value that could still be created? And how much is value waiting for someone else to pay for it tomorrow?

These questions become especially urgent when speculation starts to finance itself.

In the early stages of a technology boom, investors put in money hoping for high returns and accept losses along the way. Further down the road, capital commitments grow, infrastructure becomes more expensive and financing becomes ever more sophisticated.

By October 2026, worries about debt-fueled AI expansion had become hard to ignore. According to the Financial Times, CoreWeave, an AI infrastructure company, had debt and lease obligations of about $52 billion at the end of June.⁴ That growth reflected strong customer demand, but also the rising financial burden of meeting it.

The danger is not that demand slows, but that it grows more slowly than the obligations taken on in its name. A technology boom can predict the future correctly and still get the timing wrong. A business can forecast the size of a market correctly and still fail on the way to it.

This is especially problematic for capital-intensive technologies, where infrastructure often has to be built before the demand materializes. The obligations attached to that capital cannot wait for reality to prove the hypothesis right. And technological progress does not guarantee that the infrastructure will be profitable in the long run.

A data center built for today’s demand for computing capacity may meet tomorrow’s gains in inference efficiency. A hardware advantage may be undermined by advances in architecture. A capacity shortage can turn into oversupply when suppliers project today’s demand too far into the future.

The market may ultimately need far more intelligence, and far less of a particular kind of infrastructure, than its builders assumed.

The demand will be satisfied. But not every brewery will survive.

This is the point at which the question needs to be asked more carefully.

Instead of asking whether AI valuations are justified in general, we should look at where economic value is actually being created.

Does this company have a durable advantage, or only temporary access to something scarce? Does its product actually change economic outcomes? Is its growth driven by customers’ willingness to pay, or subsidized, for now, by the generosity of investors? Will lower costs bring better margins, or less differentiation?

More importantly, who ends up with the surplus?

The customer? The application provider? The model maker? The chipmaker? The energy producer? The financier?

A technology industry can generate trillions of dollars of economic value without handing that surplus to the investors who paid for its emergence. That is not an anomaly. It is a defining feature of economic revolutions.

IV. The Price of Being Early

We tend to judge technological booms in moral terms. The optimist is a visionary until the failures begin, and then a fool. The skeptic is dismissed as unimaginative until the failures begin, and then hailed as a prophet. Both are judged too much by their timing.

John Maynard Keynes saw this problem in financial markets, where investors must anticipate not only future value but also the future expectations of other investors. His famous beauty-contest analogy captures the nature of speculation: the winning entry is not the one that best judges who is most beautiful, but the one that best predicts how everyone else will judge.⁵ Modern technology markets have industrialized this activity.

Venture capital flows to companies whose valuations assume future adoption. Adoption, in turn, depends on infrastructure that investors fund. That infrastructure produces revenue, which seems to validate the original assumption about adoption. The process may be productive. It may also be circular.

Revenue earned during an investment boom is real. But it does not, by itself, prove that the end users of the technology create enough economic value to sustain the whole system. At some point, the money in the system has to come from customers paying for outcomes, not from investors paying for expectations.

The difference is hard to see when every participant points to another participant’s spending as proof of demand. It becomes more visible when capital gets more expensive, competition intensifies, or progress turns yesterday’s capabilities into tomorrow’s commodities. None of this means the AI revolution is imaginary. It means the financial system is pricing the revolution’s effects with far more confidence than it should.

We should be wary of anyone who tells us every bubble bursts before it produces anything useful. Technological development is under no obligation to move at the speed of cautious investment. Society can benefit from investments that ruin their financiers; competition can generate enormous consumer surplus precisely because companies cannot keep the value they create. The public can win even when shareholders lose.

But there is also a limit to how much capital an economy can devote to accelerating an uncertain future. Money spent on one technological bet is money that could have gone to another. Electricity, land, engineers, financial and political resources do not become free just because the project consuming them promises some form of intelligence.

The issue is not whether optimism should be allowed. It is whether optimism is being mistaken for evidence. A good bubble brings future possibilities into the present; a bad one brings future profits into present valuations without bringing the means to earn them. While prices are rising, the two look identical. Only the economics tells them apart.

V. What Is in the Glass?

An earlier version of this essay was written in February 2025, while the market was still absorbing DeepSeek and the main debate was whether cheaper models would strengthen or break the market for computing. Time has made that original question less urgent, and the underlying issue far more pressing.

We now have better evidence about what artificial intelligence can do. We also have better reasons to doubt how its economic benefits will be distributed, how fast they will arrive, and how much money must be spent to reach them. These are different questions, and they deserve different answers.

The industry contains productive speculation and technological theater, good businesses and bad ones, infrastructure that will become indispensable and infrastructure that will never pay for itself. It includes companies that create real value but are overvalued by investors, and companies whose success rests on value that has yet to be created. Some of the foam is worth drinking. Some of it is a warning. And some of it comes from too many people pouring at once.

We should stop demanding a single verdict on an industry whose economic logic is still taking shape. The analyst’s task is not to forecast the moment the enthusiasm fades. It is to work out which part of the enthusiasm is funding future progress and which part is funding nothing but itself. The first can produce something extraordinary even as investors lose money; the second produces only disappointed expectations and an expensive lesson in simple arithmetic.

The tragedy of speculative fever is not that people believe in the future. It is that they so often confuse belief in the future with a claim on its profits.

A hundred years from now, nobody will remember Nvidia’s share price or which AI company had the most outrageous valuation. What will be remembered, if the technology delivers on its promise, is what became possible once intelligence became cheap and ubiquitous.

But history’s verdict on a technology is no refund for those who paid too much to own it.

So perhaps the question was never whether there is an AI bubble.

There is always foam.

The question is whether we are paying for the drink itself – or only for the right to see it rise.

Footnotes

1. “Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history,” CNBC, January 27, 2025. ↩

2. W. Stanley Jevons, The Coal Question, 2nd ed. (1866), chapter VII, “Of the Economy of Fuel.” ↩

3. Stanford Institute for Human-Centered AI, The 2025 AI Index Report. ↩

4. Financial Times, on CoreWeave’s debt and lease obligations, October 2026. ↩

5. John Maynard Keynes, The General Theory of Employment, Interest and Money (1936), chapter 12, “The State of Long-Term Expectation.” ↩