AI Borrowed $450 Billion. The Bank of England Is Watching the Exit.

The AI boom has crossed a financial line.

It is no longer funded mainly by venture capital, retained earnings and rising share prices. Debt is becoming a central piece of the infrastructure buildout — and one of the world’s most important central banks is now asking what happens if the promised returns arrive late.

The Bank of England’s Financial Policy Committee says global AI-related debt issuance reached an estimated $450 billion by early September 2026. That figure, produced by Morgan Stanley and cited by the committee, is more than double the total issued during all of 2025.

The number does not prove an AI bubble. It does prove that disappointment would no longer stay neatly inside technology stocks.

IN 30 SECONDS

The Bank of England says rapid growth in AI-related borrowing is widening the number of investors and funding markets exposed to the technology.

Its September financial-stability record cites an estimate of roughly $450 billion in global AI-related debt issuance this year. JPMorgan analysts estimate that $4.1 trillion of AI capital expenditure could be debt-financed between 2026 and 2030. Private credit may finance another $700 billion of data-centre spending between 2026 and 2028, according to a separate Morgan Stanley estimate cited by the Bank.

These are analyst estimates, not guaranteed outcomes.

The committee’s concern is the connection between leverage, opaque structures, high valuations and assumptions about future AI demand. If those assumptions weaken, losses could travel from AI companies and data-centre projects into private credit, corporate bonds, leveraged investors and, potentially, sovereign markets.

THE NUMBER THAT MATTERS

$450 BILLION

That is the estimated volume of global AI-related debt issuance in 2026 as of early September.

For perspective, the Bank said issuance this year was expected to exceed that of countries including the United Kingdom. In sterling corporate-bond markets, AI hyperscalers accounted for 47% of issuance so far in 2026.

This does not mean every dollar is dangerous. Debt can be an efficient way to finance assets with long useful lives and predictable cash flows. The issue is whether today’s AI infrastructure will become that kind of asset.

A power contract, data centre or chip cluster can last for years. The technology it supports may change in months. That mismatch is where financial risk begins.

From equity story to credit story

Equity investors can absorb large swings. A share price can fall without forcing the company to repay shareholders.

Debt behaves differently. Interest must be paid. Principal eventually matures. Covenants may tighten when performance weakens. Refinancing can become expensive exactly when a company needs more time.

That makes the source of funding important.

When AI expansion is financed by operating cash flow, the company is spending money the business has already generated. When it is financed with debt, the project is making a claim on future cash flow. The larger the claim, the more precisely the future must cooperate.

AI companies are not borrowing only for software teams. They are financing chips, data centres, land, power capacity, networking, cooling and long-term cloud commitments. The physical layer is costly before it is productive.

The Bank’s warning is not that this investment has no value. It is that the financial system is becoming more dependent on the assumptions behind it.

The three assumptions holding up the structure

The first assumption is demand.

Companies are building capacity for a world in which businesses and consumers use much more AI. If adoption grows more slowly, expensive infrastructure may sit below planned utilization.

The second is monetization.

High usage does not automatically produce high margins. Providers still need prices that cover compute, energy, depreciation, engineering and financing. A popular product can remain economically weak if every query is costly and competition prevents price increases.

The third is asset life.

Debt-financed infrastructure makes more sense when the asset produces cash for longer than the debt costs money. AI chips can lose commercial relevance quickly as newer generations improve performance and energy efficiency. Data centres may last, but their internal equipment may need replacement much sooner.

The Bank explicitly highlighted uncertainty around compute demand, power availability and depreciation rates. None of those variables is a side detail. Together, they determine whether the collateral behind the boom ages gracefully or brutally.

MNS LENS

AI debt is not one trade. It is a chain.

A hyperscaler issues bonds. A private-credit fund finances a data-centre developer. An insurer holds the fund. A utility builds generation for the campus. A chip supplier expands production. A landlord values property around expected power demand. A government forecasts tax revenue from the investment.

Each participant may look diversified. The underlying economic bet can still be the same: AI demand will rise fast enough, stay high enough and generate enough cash to service the capital committed today.

This is what “interconnected” means in practice. The risk is not simply that one model disappoints. It is that many balance sheets rely on similar expectations while appearing to own different assets.

The Bank also pointed to opacity and circular arrangements. In an AI ecosystem, suppliers can invest in customers, customers can commit to suppliers, and financing can be supported by contracts inside the same network. Such arrangements may accelerate construction. They can also make independent demand harder to measure.

Revenue looks strongest when every participant believes the next participant will keep spending.

Why bonds make the warning louder

Government-bond yields are already elevated. The Bank said sovereign yields in several advanced economies reached levels not seen since 2008 as energy costs, inflation and debt issuance tightened financial conditions.

That matters because corporate debt is priced relative to sovereign benchmarks. When the risk-free base rises, AI projects must earn more before they create value.

A project that looked attractive with cheap financing may fail the same calculation at higher rates. Refinancing risk also grows. A company can survive a weak year if debt is long-dated and fixed. The same company can be forced into painful choices if large maturities arrive during a market repricing.

The Bank did not report a current systemic failure. It said the financial system remained resilient and kept the UK countercyclical capital buffer at 2%. The warning is about the probability that several weaknesses crystallize together.

Energy inflation pushes yields higher. Higher yields tighten credit. Tighter credit challenges leveraged AI investment. A fall in AI earnings expectations hits equities and debt at the same time. Leveraged investors unwind positions. That is the sequence policymakers do not want to discover in real time.

The private-credit blind spot

Public bonds are visible. Private credit is less transparent.

The Bank cited an estimate that $700 billion of data-centre capital expenditure from 2026 through 2028 could be financed by private credit. These funds can provide patient capital and tailor terms to complex projects. They can also make exposure harder to map because loans are not traded and disclosed like public securities.

Private-market assets are often valued less frequently. That can reduce day-to-day volatility without reducing economic risk. A price that does not move is not necessarily a value that has not changed.

The question is who ultimately holds the risk: pension funds, insurers, banks, wealthy investors or financing vehicles using additional leverage. The answer may be several of them at once.

WHAT DOES THIS CHANGE?

For AI companies, the era of celebrating capital commitments without explaining returns is ending. Investors will increasingly ask how much infrastructure is debt-financed, when payments begin and what utilization is required to cover them.

For lenders, technical due diligence becomes financial due diligence. Chip depreciation, grid access, cooling efficiency and model economics are now credit variables.

For investors, “AI exposure” no longer means only Nvidia, hyperscalers or software companies. It can sit inside bond funds, private-credit portfolios, utilities, real estate and insurers.

For households, the connection is indirect but real. If AI borrowing crowds into capital markets or reprices sharply, borrowing conditions for companies can tighten. That can affect hiring, investment and portfolios long before an AI product disappears from a screen.

WHAT TO WATCH NEXT

• Actual AI-related bond issuance versus the $450 billion estimate.
• Debt maturities, interest expense and refinancing needs at major infrastructure builders.
• Utilization rates and contracted revenue for new data centres.
• The share of capital expenditure funded by operating cash flow versus borrowing.
• Private-credit exposure and whether lenders disclose concentration by project, customer and power market.
• Circular commercial arrangements that may overstate independent demand.
• Evidence that AI adoption produces measurable productivity and cash flow, not only usage growth.
• Whether a correction in AI assets spreads into core credit or sovereign markets.

The AI boom is still a technology story. It is now also a duration story, a refinancing story and a balance-sheet story.

The industry borrowed against a very profitable future. The Bank of England is not saying that future will fail.

It is asking who gets paid if it arrives late.

ASK NOVA

How can I tell when a booming investment theme is becoming a debt risk?

SOURCES

Bank of England, Financial Policy Committee Record — September 2026, published 30 Sep 2026.
https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026

Reuters, 30 Sep 2026, on the FPC record and Andrew Bailey’s accompanying comments.
https://www.reuters.com/business/finance/bank-england-sees-growing-risk-that-dangers-ai-debt-will-materialise-2026-09-30/