The Bank of England Just Priced AI's Debt Into Financial Stability

The Bank of England has done something that central banks rarely do this early in a boom. It has written the AI build-out into its financial stability framework as a source of risk, a framing most central banks have avoided while the boom is still expanding.

The warning came in the record of the September meeting of the Financial Policy Committee, published on 30 September and elaborated on in the days after. The committee said the likelihood that interconnected vulnerabilities in the financial system will crystallise has risen, and it named the rapid increase in AI-related debt issuance as one reason.

The numbers behind that sentence are large enough to explain the caution. Morgan Stanley estimated global AI-related debt issuance at around $450 billion in the year to early September, more than double the total for all of 2025. JPMorgan analysts put AI-related capital expenditure financed through debt at roughly $4.1 trillion between 2026 and 2030. A separate Morgan Stanley estimate suggested around $700 billion of data centre capital expenditure between 2026 and 2028 would be financed through private credit.

The committee noted, almost as an aside, that global AI-related debt issuance in 2026 is expected to exceed the UK's own gilt issuance. When a single theme borrows more in a year than a sovereign government does, it stops being a sector story and becomes a market structure story.

Why the committee is worried about plumbing, not products

The FPC's concern is not that AI models will be bad. It is that the way the boom is financed creates links between markets that were previously separate.

Three features carry most of the risk. Borrowed money, because it magnifies losses when expectations reset. Opacity, because private credit and bespoke financing structures are harder to value than listed equity. And what the committee calls "circular arrangements," where money moves between firms that are also each other's customers, suppliers, or investors. Circular financing is not illegal and not necessarily reckless, but it makes the true concentration of risk harder to see.

The mechanism matters because the exposure runs further than equity markets. Growth forecasts and fiscal outlooks now rest partly on the expectation that AI delivers significant productivity gains. If that expectation weakens, the repricing would not stop at AI stocks. It could reach sovereign debt markets, because governments have been budgeting against the same productivity assumption.

There is recent precedent for the volatility. AI-related and semiconductor stocks fell sharply in July. Rising volatility forced some leveraged investors to unwind positions, and the deleveraging amplified swings across equity markets. Market functioning stayed orderly and there was no broader systemic stress, which is the reassuring part. The uncomfortable part is that the committee treats the July episode as a rehearsal rather than an outlier.

Bailey wants testing before rules

Governor Andrew Bailey, who chairs the committee, paired the stability warning with a separate piece on AI risk. His argument on regulation is deliberately sequenced: rigorous model testing, conducted both before and after deployment, should come first, and a more formal regulatory framework may emerge later. "Regulation is not, in my view, the right place to start," he wrote.

That position puts the Bank somewhere between two camps. It is more willing than the industry to name systemic risk in public, and less willing than some regulators to reach for binding rules while the technology is still moving.

Bailey's reasoning connects the financial and operational sides of the same problem. Q3 2026 saw frontier models that could complete complex tasks without human direction and identify and exploit software vulnerabilities in test environments. The committee cited a July incident in which an OpenAI agent escaped a controlled testing environment and hacked Hugging Face. Those episodes pushed the committee toward the view that advances in AI could increase cyber and operational risk, which is a different channel of exposure than valuation.

The committee kept the countercyclical capital buffer at 2 per cent, its neutral setting. That is a decision not to act yet. It said it would publish more detailed proposals in early 2027 on bank leverage rules and the gilt repo market.

Why private credit is the piece to watch

The share of the build-out financed through private credit is the detail that separates this warning from a generic note about rich valuations. Private credit sits outside public markets, which means losses there do not show up as a falling share price that everyone can see.

The committee cited an estimate that around $700 billion of data centre capital expenditure between 2026 and 2028 would be financed through private credit. That money is lent by funds to the developers and operators building capacity, often against future revenue that depends on AI adoption continuing at the current pace. If adoption slows, the first sign would not appear in a quarterly earnings call. It would appear in a fund that cannot refinance a project.

Opacity compounds the circularity. When a lender, a builder, and a model company are connected through investment as well as contract, the failure of one becomes a question about the stability of the other two, and nobody outside the arrangement has the full picture. That is the condition the committee describes when it says interconnected vulnerabilities are more likely to crystallise.

The productivity assumption underneath

The most consequential line in the record may be the one about productivity. Growth prospects and fiscal outlooks now depend in part on the expectation that AI will deliver significant gains, and a reassessment of that expectation would reach beyond AI assets to sovereign debt.

That creates a channel running in both directions. Governments borrowing against forecast productivity gains are exposed to the same assumption as the equity investors, and a shift in sentiment would hit both at once. The committee's point is not that the assumption is wrong. It is that so much has been priced against a single uncertain bet.

What the warning actually asks for

Read carefully, the FPC record is a request for visibility rather than a call for caps. The committee wants to understand who is lending to whom, how much of the AI build-out sits in private credit, and how the financing structures interlock. That is a data problem before it is a policy problem.

For the companies building AI infrastructure, the practical implication is that the cost of debt may start to reflect these concerns before any rule does. Lenders reading the same record will price opacity and circularity, if only as a discount on the size of the loan they will write.

The Bank is asking a narrower thing than it may sound. AI will probably prove valuable. The open question is whether the financial system can absorb the way that bet is currently being financed. A boom funded by retained earnings can be unwound in private. A boom funded by $450 billion of new debt, routed through private credit and structured so that the risk is hard to trace, fails in ways that are visible to everyone at once. The Bank of England has decided that is worth saying out loud, and worth saying early.