Bailey tells the G20 that leveraged AI exposure can amplify the next shock

The FSB's stated job is coordination among G20 regulators ^[2], so naming a risk in the chair's letter is an agenda-setting act, not a rule. The output is surveillance. An item lands in monitoring work, then in supervisory questions, then in the arguments banks have with their own national regulators about what an exposure is worth. That route is slow. It is also the route that eventually sets what a data centre loan costs.

The transmission into compute pricing runs through debt, not equity. The report notes Aswath Damodaran's argument that this build cycle requires heavy physical infrastructure spending, much of it debt-financed, so a correction would not stop at shareholders ^[11]. The way I would model a committed GPU-hour price is amortised hardware and shell, plus the cost of the money that bought them, plus power, plus margin. A share price drawdown leaves the second term alone. A repricing of credit does not.

So here is what would have to be true for Bailey's letter to show up on your invoice. First, lenders to data centre vehicles and neoclouds have to actually widen spreads or tighten covenants because supervisors are now asking about this exposure. Second, that cost has to land in new capacity contracts rather than being absorbed by an operator with cheap balance sheet capacity. Third, demand has to stay strong enough that suppliers can pass it through instead of eating it. Fail any one of the three and the letter changes nothing you can measure.

The concentration point is the one that maps cleanly onto engineering. Bailey writes that banks worldwide depend on a handful of large tech providers, that a successful attack on one could hit institutions in several countries at once, and that uneven national cyber defences are themselves a vulnerability ^[8]. That is the same dependency graph most inference stacks have, drawn from the other side. He pairs it with the claim that frontier AI could "materially alter the speed, scale and economics of cyber risk" ^[7], and the FSB is studying how financial firms can safely use frontier models for cyber defence ^[10]. Meanwhile many countries have no rules at all for developing, releasing or deploying advanced models, which he says makes safe-release work a priority ^[9].

The quotable line is about timing. Rising leverage is a feature of a maturing financial cycle, Bailey writes, reinforcing markets on the way up and intensifying declines when sentiment turns, "as recent weeks have demonstrated" ^[5]. A chair citing recent weeks as evidence is describing something already observed, not a scenario. He also records that the system absorbed the Middle East shock while conditions stayed tight, with higher interest rates and high valuations on risky assets ^[12]. His specific channel is the web of cross-investments between AI companies and hyperscalers, where one large AI company stumbling could pull down other tech firms and then the broader market ^[6]. That is a correlation claim about counterparties, and it applies to whoever signed your capacity contract.

One caveat on sourcing: both supplied blocks are the same the-decoder.com report of the letter ^[13], so every quote here is as reported by that publisher, and the primary document is a letter to G20 finance ministers that I have not read. The letter is a lagging indicator with an unusually good mailing list.