Two of the most cautious voices in global finance spent the last week of June 2026 saying roughly the same thing about the same subject. AI market risk has moved from a contrarian talking point to a formal warning issued by the institutions whose job is to watch for the next crash. J.P. Morgan published an analysis full of concentration red flags, and the Bank for International Settlements used its annual report to warn that debt-fueled spending on AI data centers could turn a boom into a bust with consequences reaching well beyond tech stocks. Neither is predicting a collapse on a date. Both are saying the conditions that precede one are visibly building.
- J.P. Morgan says 42 AI-linked companies in the S&P 500 account for 65 to 80 percent of the index's profits and investment since late 2022.
- The top ten U.S. stocks now make up about 40 percent of the S&P 500's market value, up from 17 percent in 2015.
- The BIS warns that AI financing increasingly runs through debt and lightly regulated non-bank intermediaries, raising the odds of fire sales if sentiment turns.
- Both warnings stop short of calling a top, but they frame a correction as potentially more damaging to the wider economy than past ones.
The reports landed within a day of each other and reinforce one another. The Decoder summarized J.P. Morgan's findings on market concentration, while the BIS annual report focused on how the AI build-out is being financed. A separate Telegraph report framed the central-bank concern in starker terms, warning that the AI boom risks a global financial crash.
J.P. Morgan's concentration red flags
The bank's core worry is concentration, and the numbers it cites are striking. Since ChatGPT launched in late 2022, about 42 AI-linked companies in the S&P 500 have driven somewhere between 65 and 80 percent of the index's total profits and investment. That means a handful of firms are carrying the headline performance of the broadest measure of American corporate health. When that few names matter that much, a stumble at any of them stops being a company story and becomes an index story.
The market-cap picture is just as lopsided. J.P. Morgan notes that the ten largest U.S. stocks now represent roughly 40 percent of the S&P 500's total value, more than double the 17 percent they made up in 2015. An index that is supposed to spread risk across 500 companies increasingly behaves like a bet on ten. Passive investors who think they own a diversified slice of the economy are, in practice, heavily exposed to the fortunes of a small cluster of AI and chip giants.
The technical signals around semiconductors add another layer. According to the analysis, the chip rally is flashing patterns last seen during the dotcom bubble, and leveraged semiconductor ETFs have quintupled their influence on global stock markets since early 2024. Speculative behavior shows up in the plumbing too: margin lending on the Korean stock exchange has tripled since 2020, and options trading in semiconductor names runs at roughly five times its 2020 level. Leverage and options activity at that scale tend to accelerate moves in both directions, which makes any reversal sharper than the underlying fundamentals alone would justify.
J.P. Morgan frames these as distinct layers of concentration risk rather than a single worry. There is concentration in the markets, where a few stocks dominate the index. There is concentration in the infrastructure, where the AI build-out depends on a narrow set of chip suppliers and cloud providers. And there is concentration in the broader economy, where AI-linked capital spending has become a meaningful driver of growth. A shock to any one layer can propagate into the others, which is what makes the overall picture more fragile than any single statistic suggests. The same names that lead the stock market also build the data centers and place the chip orders, so the layers are not independent. They are the same companies viewed from different angles.
The cracks beneath the chip leader
Even the dominant player is not as untouchable as it was. J.P. Morgan flags that Nvidia's share of the AI accelerator market has slipped from about 85 percent in 2023 to an estimated 75 percent in 2026, as custom silicon from the hyperscalers eats into its position. In-house chips such as Google's TPUs and Amazon's Trainium reportedly cut operating costs by 30 to 40 percent for the buyers that deploy them, which gives the largest cloud providers a real incentive to design around the market leader rather than keep paying its margins.
That erosion matters for the concentration story because so much of the AI trade rests on a single supplier's pricing power. If the biggest customers successfully shift volume to their own chips, the revenue assumptions baked into the most valuable stock in the rally come under pressure. The bank also points to the labs themselves, where revenue is growing fast but compute costs are massive and future profitability remains unclear. The economics of the whole stack, from chips to models, lean on growth continuing to outrun cost, and that is not guaranteed.
Falling token prices, a theme covered in our piece on Chinese AI models and the enterprise cost shift, cut both ways here. Cheaper inference is good for buyers but compresses the revenue projections that justify the AI labs' valuations.
The BIS warning on how AI is financed
Where J.P. Morgan looks at stock prices, the Bank for International Settlements looks at the debt underneath them, and that is arguably the more serious concern. In its annual report, published in Basel on June 28, 2026, the BIS warned that the financing of the AI build-out increasingly relies on debt and on complex funding structures running through non-bank intermediaries that operate with less oversight than traditional lenders.
The BIS was direct about the failure mode. "Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust, with potential knock-on effects on financial conditions," the report said. The mechanism it describes is familiar from past crises. Money pours in while optimism holds, the spending is funded with leverage, and a shift in sentiment forces a rapid unwind that feeds on itself.
The role of hedge funds and private credit vehicles drew specific concern. "These hedge funds employ highly leveraged strategies that rely on short-term financing on favorable terms, creating risks of fire sales and de-leveraging feedback loops," the BIS wrote. When leveraged players funded by short-term money have to sell into a falling market, their selling pushes prices down further, which forces more selling. That loop is how a contained disappointment becomes a broad rout.
The opaque plumbing of AI deals
The BIS reserved some of its sharpest language for the structure of the deals financing data centers. These arrangements often mix equity and debt alongside long-term supplier contracts in ways that are hard for outsiders to assess. Chipmakers and hyperscalers take equity stakes in AI labs, and those same labs commit to buying chips for years, creating circular dependencies where each party's health props up the others.
On the disclosure problem, the report was blunt. "The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times," the BIS said. An asset pledged as collateral to more than one lender is a classic hidden fragility, because it means the true amount of leverage in the system is larger than any single participant can see. Third-party data center leasing with embedded exit clauses adds further undisclosed risk that only becomes visible under stress.
The circularity is the part that should give investors pause. When a chipmaker takes an equity stake in an AI lab, and that lab then commits to multi-year chip purchases from the same chipmaker, each side is booking the other's spending as its own demand. Revenue at one company becomes a cost commitment at another that owns a piece of the first. As long as everyone keeps growing, the arrangement looks like a virtuous cycle. If growth slows, the same links that amplified the upside transmit the pain in every direction at once. The interdependence that makes the ecosystem feel robust in good times is precisely what makes it brittle in bad ones, because there are fewer truly independent players to absorb a shock.
Outsiders cannot easily measure this, which is the deeper issue. A lender looking at one data center deal sees collateral and a contract. It cannot necessarily see that the same asset backs another loan elsewhere, or that the tenant's exit clause shifts the risk back onto the builder. The BIS warning about assets pledged multiple times is a warning that the system's true leverage is unknowable from any single vantage point, and unknowable leverage is the raw material of every financial crisis. The 2008 parallel the bank draws rests on exactly this feature, not on the specific instruments involved.
The BIS summarized the moment with a phrase that captures the ambivalence: "The global economy remains caught in the crosscurrents of progress and peril." The technology may be real and useful, and the financing around it may still be dangerous. Those two things are not in tension. The history of bubbles is largely a history of genuine innovations financed recklessly.
Why a correction could hit harder this time
The BIS made a point that deserves emphasis: the scale of the AI trade means a sell-off would not stay contained to equities. "A major equity-market correction could have larger macroeconomic consequences today than in the past," the report said, because households and pension funds across the broader economy are more exposed to these concentrated gains than in previous cycles. The bank went further, warning that a repricing of risk, "whether triggered by higher interest rates or an AI bust, has the potential to be similarly disruptive" to the 2008 financial crisis.
That comparison is not thrown around lightly by an institution like the BIS. The reasoning is that the wealth effect runs in reverse during a crash. When a small set of stocks represents a large share of retirement accounts and index funds, a steep decline drains household balance sheets quickly, which pulls back consumer spending, which slows the real economy, which feeds back into corporate earnings. Concentration on the way up becomes concentration on the way down.
This is the feature that distinguishes the current setup from a broad-based expansion. In a market where gains are spread across many sectors, a slump in one corner can be cushioned by strength elsewhere. When the index is effectively a wager on a dozen AI-linked names, there is no such cushion, and a decline in those names pulls everything connected to them down together. The BIS is essentially warning that the diversification investors believe they hold is partly an illusion at todays level of concentration.
The historical rhymes regulators keep citing
Central bankers reached for history to make the risk legible. As reported by The Telegraph, the BIS drew parallels between the AI infrastructure surge and the dotcom boom and the British railway mania of the 1840s, with echoes of the speculative excess that preceded the Great Depression before the Great Depression. Each of those episodes paired a major new technology or expansion with a wave of debt-financed overbuilding, and each ended with a painful clearing of the excess even where the underlying innovation endured.
The railway comparison is the most instructive. Britain genuinely needed railways, and the network built during the mania served the country for a century. Investors who financed it at the peak still lost fortunes when the bubble burst. The lesson is not that AI is fake. It is that being right about the technology offers no protection against being wrong about the price, and the two questions are entirely separate. The Bank of England had already signaled this nervousness in December, warning that share prices were the most stretched they had been since the 2008 crisis.
That December warning from the Bank of England is worth weighing on its own. A central bank does not casually invoke 2008 in describing current valuations, and the comparison was about price levels rather than financing structure, a separate measure from what the BIS later emphasized. Taken together, the two institutions are pointing at different parts of the same animal. One sees prices that have run far ahead of history, the other sees the borrowing that funds the spending behind those prices. A risk that shows up in both the valuation data and the credit plumbing is harder to dismiss as the worry of a single perspective, which is part of why these particular warnings have carried more weight than the steady background hum of bubble talk that has accompanied the rally for two years.
The case for not panicking
It would be one-sided to present only the bearish read, and the reports themselves are careful. J.P. Morgan acknowledges that AI revenue growth remains strong, which is not something that was true of many dotcom-era companies that had no profits at all. The leading firms are generating real earnings, and the demand for compute is backed by actual usage rather than pure speculation. A market can be concentrated and richly valued without being a bubble that has to burst on schedule.
The honest position is that nobody, including the BIS and J.P. Morgan, knows the timing. Warnings about stretched valuations have been issued before and markets have climbed for years afterward. What the two reports establish is not a prediction but a risk profile. The exposure is concentrated, the financing is leveraged and opaque, and the macroeconomic stakes are higher than usual. Those are conditions, not catalysts. A catalyst, when it comes, is usually something nobody flagged in advance.
Reading the warnings without overreacting
The value in these reports is not a trading signal. It is a clearer map of where the fragility sits. The concentration that J.P. Morgan documents and the financing structure that worries the BIS are the parts of the system that would transmit a shock fastest, which is exactly why regulators are pointing at them now rather than after the fact. An investor or operator who understands that a few names carry the index, and that much of the build-out rests on leverage that is hard to see, is better positioned than one who assumes the breadth of the market still spreads the risk.
What both institutions are really arguing is that the AI story has outgrown the simple framing of technology versus skeptics. The technology can keep advancing while the financial architecture around it becomes more dangerous, and the second trend can hurt people who never placed a bet on the first. That is the uncomfortable synthesis these warnings offer. The boom is producing genuine capability and genuine systemic risk at the same time, and the prudent response is to take both seriously rather than picking the half that fits a prior conviction.