The debate over whether the AI-driven stock market is a bubble has become a contest of analogies. One side points to the dot-com crash and warns of history rhyming. The other points to profits and demand that the dot-com companies never had. Both are marshaling real evidence, and the argument is unresolvable in that form, because "is it a bubble" can only be answered after the fact. A more useful question is available: not whether it will burst, but what specifically would have to break for it to, and how much of the market goes with it if it does.
The concentration is the fact everything else hangs on
Start with the number that is not in dispute. The largest handful of companies, most of them AI-related, now make up an extraordinary share of the entire US market. By various measures the top 10% of companies account for roughly 75% of total market capitalization, the highest concentration on record, higher than the late-1990s peak. The top 10 stocks alone are around 35% of the S&P 500, against roughly 25% at the dot-com peak.
This matters more than valuation, and it is worth being clear about why. A concentrated market means the index no longer represents what most people think it does. Someone who owns an S&P 500 fund believing they hold a diversified basket of 500 companies is, in practice, making a large bet on a handful of AI-exposed names, because those names drive the index's movement. The diversification is nominal. The exposure is concentrated.
That is the real risk, and it is structural rather than speculative. Even an investor who has never knowingly bought an AI stock is heavily exposed to the AI trade through ordinary index funds. If those names correct sharply, the correction reaches essentially every retirement account in the country, whether or not its owner ever formed a view on artificial intelligence. Concentration converts a sector risk into a market risk.
Why this is genuinely not the dot-com bubble
The bears' analogy weakens under examination, and intellectual honesty requires saying so clearly.
The dot-com leaders were priced on stories. Cisco traded at something like 472 times earnings at the 2000 peak; many of the era's darlings had no earnings at all, and some had no revenue. The companies leading today are among the most profitable enterprises in history. Nvidia at roughly 44 times earnings is expensive, but it is a different universe from 472 times, and it is expensive because it is selling enormous quantities of a product at high margins, not because a narrative says it might someday.
The infrastructure is also real in a way the telecom bubble's was not. The dark fiber laid in the late 1990s sat unused for years because demand did not exist. The data centers and chips being bought now are running at capacity, serving actual workloads that actual customers are paying for. And the spending is funded largely by the cash flow of profitable companies rather than by speculative debt, which means a downturn does not automatically trigger the forced selling and bankruptcies that deepened past crashes.
So the crude "it's 2000 all over again" case is weak. The leaders are real businesses, the demand is real, and the balance sheets are strong. Anyone dismissing the whole thing as obvious mania is ignoring the most important differences.
Where the real vulnerability actually sits
But dismissing the bubble case entirely requires ignoring something specific, and it is not valuation. It is a monetization gap.
The hyperscalers, Alphabet, Amazon, Meta, Microsoft, and Oracle, are on track to spend on the order of $755 billion on AI-related capital expenditure in 2026, with combined Magnificent Seven capex having climbed from roughly $160 billion to nearly half a trillion since ChatGPT's release. That spending is the revenue of the chipmakers and infrastructure providers whose stocks have soared. It is real money changing hands, which is exactly why the earnings look real.
The problem is what justifies continuing it. By several estimates, the software and services revenue that enterprise AI currently generates covers less than half of the infrastructure being built to deliver it. The buildout is a bet that demand and monetization will grow into the capacity. That bet may well pay off. But it means the entire structure rests on hyperscalers choosing to keep spending, and that choice depends on their belief that the return will eventually arrive.
Here is the vulnerability stated plainly. The chip and infrastructure companies' revenue depends on hyperscaler capex. Hyperscaler capex depends on confidence in future AI returns. If that confidence wavers, even without any company failing, the spending slows, and the revenue that has been validating the infrastructure stocks slows with it. The risk is not a fraud being exposed. It is a large, rational, cash-rich group of buyers deciding to spend less because the payback is taking longer than hoped, and that is a decision, not a catastrophe, which makes it both more likely and harder to predict than a classic bust.
The circularity that inflates the signal
There is a second structural feature worth understanding, because it makes the revenue look more solid than it may be.
A recurring pattern in this cycle is companies investing in each other and buying each other's products, so that one firm's capital becomes another firm's revenue in a loop. A chipmaker takes a stake in an AI developer that commits to buying the chipmaker's hardware. An infrastructure provider signs a supply deal with a cloud company it is also investing in. Each transaction is real, and in aggregate they can create the appearance of end-user demand that is partly the sector funding itself.
This does not mean the demand is fake. Much of it is genuine. It means some portion of the revenue growth reflects capital circulating within the AI complex rather than customers outside it paying for AI, and from the outside those two are hard to separate. When they are hard to separate, the market tends to assume the more favorable interpretation during good times, which is precisely when scrutiny is lowest.
What would actually have to be true
Reframing the bubble question as a conditional one makes it answerable in a useful way. The AI trade corrects meaningfully if, and largely only if, one of a few specific things happens.
Hyperscalers conclude the return on their spending is too slow and cut capex. This is the most probable trigger, because it requires no failure, only a recalculation, and the monetization gap gives them a reason to make it.
Monetization visibly stalls, meaning enterprise AI revenue plateaus well below what the infrastructure spending implied, which would undermine the case for continued buildout directly.
Or a genuine technological disappointment, a plateau in model capability or a ceiling on useful applications, that resets expectations for the whole category.
Notice what is not on the list: valuation alone. Expensive stocks can stay expensive or grow into their prices for years, and "the multiple is high" has been true throughout a period in which these stocks rose enormously. Valuation determines how far a correction falls once triggered. It rarely supplies the trigger.
How to hold both truths
The honest synthesis is that both camps are partly right and are answering different questions. The bulls are right that these are real, profitable companies serving real demand, which is why this is not 2000 and why betting on an imminent collapse has been a losing position. The bears are right that concentration is at historic extremes and that a monetization gap sits underneath the spending, which is why complacency is dangerous even though the businesses are sound.
The practical takeaway does not require predicting the unpredictable. It requires knowing your actual exposure, which for most index-fund holders is far more concentrated in this single theme than they realize, and deciding whether that concentration matches the risk they intend to take. Reducing concentration is not a bet that the AI trade dies. It is insurance against the possibility that a market this dependent on one theme corrects, whether or not the theme itself was right.
The companies can be genuinely transformative and the stocks can still be priced for a perfection that a single quarter of slowing capex would puncture. Both halves of that sentence are true, and holding them together is more useful than resolving the bubble debate that cannot be resolved until it is over.