Warnings that AI-related stocks are too expensive have been a constant refrain since 2025. But the letter that Bank of England (BOE) Governor Andrew Bailey sent to G20 finance ministers and central bank governors, in his capacity as chair of the Financial Stability Board (FSB), looks less at the level of share prices than at who is buying those shares and with what money. Investor borrowing is rising, capital is concentrating in a handful of big tech companies, and high valuations sit on top of both. When all three hold at once, the letter warns, a decline that would normally be absorbed can be amplified. Tracing the numbers that have piled up, and where today's AI investment money comes from, shows why the warning reaches beyond the stock market.
A letter arrives as AI-linked firms approach half of the S&P 500
The letter is dated August 28, 2026, and was published on the 31st. It was addressed to the G20 finance ministers and central bank governors meeting in Asheville, North Carolina. Bailey is both BOE governor and FSB chair, and he sent the letter in the latter role. According to CNBC, the core is this passage: leverage, high valuations and market concentration (notably cross-investment between AI companies and hyperscalers) can overlap and amplify a future market correction. Each of the three elements is a known issue on its own. What is new is the judgment that they now hold at the same time.
The FSB brings together national regulators and the bodies that set international standards, so issues its chair singles out tend to feed into supervisory plans across jurisdictions. This letter was issued in the FSB chair's name. A warning from the BOE governor aimed at UK markets and a letter delivered to the G20 differ in how far they reach.
Market capitalization is concentrated in a small number of stocks, and on top of that, AI companies and hyperscalers (giant cloud operators that run massive data centers) invest in one another, creating a loop in which revenue and investment feed each other. "Concentration" refers to this dual structure. The Bank for International Settlements (BIS) has flagged this circular flow of funds, and Bloomberg reported that it projects AI-related capital spending by the top five companies will exceed $1 trillion across 2025 and 2026. When the same money appears on both the investment and revenue sides, one company's spending cut hits several companies' revenue at once.
Borrowing changes how a decline propagates. Positions bought on credit trigger margin calls once prices fall, and investors sell holdings to raise cash. If the stocks being sold are scattered, the effect on the overall market is limited; if everyone has pledged the same few names as collateral, the selling converges on the same place. The loop in which falling collateral values and margin calls feed each other is the substance of the amplification Bailey describes.
The figures the BOE cited in its July 2026 Financial Stability Report show how large those positions are. Prime brokerage credit, the credit securities firms extend to hedge funds, has grown about 40% over the past year, and US leveraged ETF assets have reached about $200 billion. The letter points to the actual positions these two figures represent.
Both are ways to build up equity exposure with leverage, but the mechanisms differ. Prime brokerage credit is direct borrowing from a securities firm, while leveraged ETFs synthesize leverage through derivatives such as swaps and futures. Hedge fund positions financed by the former can be forcibly unwound when markets turn violent, and the latter generate mechanical trading through daily rebalancing.
Why the BOE's warning sharpened in ten months
The BOE's article that put AI-linked asset valuations front and center was the October 2025 Bank Overground piece "All chips in!" (The BOE's Monetary Policy Committee had already touched on the correction risk from stretched valuations and market concentration among AI-related tech firms in the minutes of its meeting published October 8 of that year.) The piece showed the median price-to-earnings ratio of AI-linked stocks at 31x, well above 19x for the market as a whole. AI-linked companies made up about 44% of the S&P 500, up from about 26% in late 2022. That article did mention leverage and external financing, but what stood out was the high stock multiples.
| Date | Document | Figures cited | Focus of warning |
|---|---|---|---|
| October 2025 | Bank Overground "All chips in!" | Median P/E of AI-linked stocks 31x (19x for the overall market); share of S&P 500 about 44% (about 26% in late 2022) | High stock multiples |
| July 2026 | Financial Stability Report | AI companies' share of the S&P 500 up from about a quarter of the 2022 level to about half. Excess CAPE Yield at its lowest since the dot-com bubble | Valuation of the market as a whole |
| August 2026 | Letter to the G20 as FSB chair | Focus on the interaction of borrowing, valuations and concentration rather than on presenting figures | Sources of funding and channels of amplification |
What strengthened over ten months was less a new argument than the clarity with which the elements are linked. In autumn 2025, high stock multiples were in front. The July 2026 report tied the high valuations of AI-linked stocks to rising leverage in the equity market, including hedge fund prime brokerage credit and leveraged ETFs, and noted that multiple vulnerabilities could materialize at once. The August letter put that link at the center of its warning. It took a little over three and a half years, from late 2022, for AI companies to swell from a quarter to half of the S&P 500, but the BOE's clarification of the link between valuations and leverage as the heart of its warning came in just these ten months.
The Excess CAPE Yield used in the July 2026 report (the inverse of the cyclically adjusted price-to-earnings ratio minus the real interest rate; the lower it is, the more expensive equities are) had fallen to its lowest level since the dot-com bubble. Excluding the top 30 AI-linked stocks, it is still the lowest since 2007. That means the overvaluation is not confined to AI names, and it is an indicator that non-AI stocks could move in the same direction when a correction comes.
Still, joining the three figures into a single line would be a misreading. The 44% in October 2025 and the "about half" in July 2026 both come from BOE materials, but the definition of the companies counted may not match exactly, so it is safer to treat them as snapshots at each point in time. A tally based on the seven Magnificent Seven stocks puts their share of S&P 500 market cap at about 34% in mid-2026 (about 7.5% for NVIDIA alone), which counts yet another population. Whichever number is cited, it cannot be compared without noting what it counts.
Debt heading toward 4x EBITDA, and lenders demanding 100bp more
Oracle's debt could expand to nearly four times EBITDA (earnings before interest, taxes and depreciation), Moody's said in an outlook as of February 2026. By some analysts' estimates, OpenAI's share of Oracle's revenue could reach 33% by 2028. Both are projections, not settled figures, but Oracle itself has said it will fund its 2026 raising plan of $4.5 billion to $5 billion roughly evenly between debt and equity, so the structure is visible: the weight of AI data center investment shows up on the debt side as well.
The thicker the debt relative to EBITDA, the narrower the room to absorb rising interest rates or weaker earnings, and the easier it becomes for a downgrade to be considered. Moody's cited the 4x level because it is looking at how much cushion Oracle has left. And on the earnings side, the assumption includes that 33% concentration in OpenAI. If one company's ability to pay changes, both the source of repayment and the terms of borrowing move at once.
In a syndicated loan (a joint loan extended by multiple financial institutions) that CoreWeave arranged in July 2026, investors demanded a spread 100 to 125bp higher (1bp is 0.01%) and maintenance covenants (an obligation to keep meeting financial metrics on an ongoing basis), Tech Times reported on August 3. The higher spread means the market's assessment of the borrower's creditworthiness fell by that much, and the covenants mean that lenders can take contractual remedial action once performance falls below a threshold (existing loans are not automatically called in immediately). The period in which AI lending was approved as soon as it was offered is coming to an end.
Whether the funding comes from equity or debt changes how far the damage spreads after a decline. If a bubble financed by equity bursts, losses fall first on the investors who held the shares. The BOE explains that even so, it can spill over into the real economy through reduced capital spending and weaker consumption via wealth effects, so equity-led does not mean the real economy is unaffected. If the funding was debt, then in addition, once the borrower's ability to repay deteriorates, losses pass in turn to banks, private credit funds (funds run by non-bank managers that lend directly) and bondholders. For loans secured by GPUs, the price of the collateral falls in the same direction as share prices, so the lender's recovery capacity thins just as the decline happens.
The difference from the dot-com bubble that is commonly pointed to is that the main source of funds around 2000 was initial public offerings and equity issuance. However, a comparison lining up debt ratios from that time against today's Oracle and CoreWeave on the same yardstick has not been presented in the series of warnings from the BOE, the International Monetary Fund (IMF) or the BIS. What is certain is that this time corporate bonds, private credit and loans secured by GPUs have come to the fore. The three institutions do not end the discussion with valuation figures because the source of funds changes the transmission channel itself.
The IMF got to "leverage over valuations" two months earlier
On June 30, 2026, the IMF's Tobias Adrian told Bloomberg that for AI, debt financing is a bigger financial stability concern than valuations themselves. The BOE's Financial Stability Report came a week later, and the letter in Bailey's capacity as FSB chair two months after that. The sequence in which the center of gravity of official discussion moves from "stocks are expensive" to "those stocks are being bought with borrowed money" fits inside those two months.
That said, the BOE did not take valuations lightly. The July report stated the AI companies' share of the S&P 500 and the Excess CAPE Yield as independent issues, and in the August letter high valuations remain one of the three elements. The difference between the IMF and the BOE is not a conflict over who is right. The IMF monitors financial systems across countries, while Bailey, as FSB chair, presses regulators in each jurisdiction toward specific preparations, and the latter has more reason to count who has piled up what.
The BIS, too, has repeatedly warned throughout 2026 about the circular flow in which investment and revenue loop among AI companies. Three international institutions are arriving at the same issue through different entrances. High valuations had already appeared in each institution's reports; what increased in 2026 was the description of where the money supporting those valuations comes from.
While valuations are the topic, regulators can do little beyond warnings and encouraging disclosure. Once leverage becomes the topic, areas that can be bound by numbers come into scope: margin ratios, hedge fund position reporting, and concentration of lending to non-bank financial institutions. That is the practical significance of the August letter going out in the FSB's name.
A named, unreleased model alongside the warning: a paradox
What Bailey cited in the letter as one of the biggest concerns for the financial system was not asset prices but cyber risk. "Frontier AI significantly changes the speed, scale and economics of cyber risk, and could undermine confidence across markets." What the governor is specifically wary of is a security breach that causes failures at multiple firms at once, and he urges authorities in each jurisdiction to prepare.
Euronews pointed to Anthropic's unreleased model "Claude Mythos," announced on April 7, 2026, as an example of the kind of breach the letter warns about. It is a model withheld from general release because it can carry out cyberattacks autonomously. I have not been able to confirm whether the name Mythos itself appears in the original letter, because text could not be extracted from the PDF. It is treated here strictly as an example offered by Euronews.
Mythos has not been released publicly, but according to TechCrunch, access to a preview has been extended under a framework called "Project Glasswing" to 12 launch partner companies, including Amazon and Microsoft, plus more than 40 additional organizations (the breakdown of organization counts varies by source). The sharing is limited to defensive security uses, so it is not the case that nothing has gone outside the developer. Since the parties Bailey urged to prepare are regulators in each jurisdiction, the supervisory question shifts from whether a specific model is released to the public to when comparable capabilities will become widely available.
The frontier models financial institutions adopt, and the infrastructure they run on, are concentrated in a small number of providers, so a breach in one model or one cloud has a path to reach multiple financial institutions simultaneously. Bailey said "at multiple firms at once" because he is looking at shared dependencies rather than the defenses of individual banks. Because the technology lineage boosting attackers' capabilities is the same one financial institutions are trying to bring into their own operations, this warning runs on a different timeline from the valuation debate.
Cyber risk lacks the years of accumulated balance statistics that valuations and leverage have. The BOE is exploring quantitative approaches, such as measuring the level of tasks models can perform and their performance in cyber exercises, but these have not yet matured into indicators that can be tracked continuously like prime brokerage credit. Bailey's decision to place this issue at the top of financial stability concerns can be read as a statement of intent to get ahead of a risk for which no established yardstick yet exists.
The next number for Japanese investors lies outside the BOJ
What the Bank of Japan heightened its vigilance about after its July 2026 monetary policy meeting was the demand and inflation that AI investment is pushing up. Nikkei reports that the BOJ cites AI-related investment as a source of price pressure once the decline in oil prices has run its course. While the BOE looks at asset prices and borrowing, the BOJ observes the same AI boom from the price side. This difference in perspective is also why the domestic discussion is not heading toward amplification of asset prices.
The BOJ's April 2026 Financial System Report judged that even including scenarios in which the AI outlook worsens, domestic financial institutions generally have ample capital and funding bases. It then placed more emphasis on the channel through which overseas hedge funds unwinding positions globally would spread via reduced liquidity in bond markets and the like, with relatively thin discussion of domestic financial institutions' equity holding risk. The prime brokerage credit and leveraged ETFs that the BOE is counting sit right at the entrance to that channel.
An example of amplification had occurred at an individual fund a month before the letter. As this site reported on July 31, 2026, Leopold Aschenbrenner's AI-focused fund was hit with a margin call on positions leveraged 4x, and its assets under management fell by 80%, from $45 billion to $10 billion. The unwinding of one fund did not shake the entire financial system. But this case has already shown how positions combining borrowing and concentration can break.
The next numbers to check are published by the BOE itself: whether prime brokerage credit keeps up its roughly 40% growth over the past year, whether the roughly $200 billion in leveraged ETFs swells further, and whether the spread premium on Oracle's and CoreWeave's refinancings widens beyond 100 to 125bp. If all three slow together, AI capital spending would return to the realm of equity and internal funds, and the channel of amplification Bailey warned about would narrow. The first two are figures from the July 2026 Financial Stability Report, and how they move in the next edition will be the first clue.