FINANCE & RISK DESK · HONG KONG · WEEKLY

AI Selloffs Route Through Prime Broker Margin Calls

Regulators are watching AI cyberattacks and AI valuations as separate risks, when the real transmission channel from an AI equity correction into bank balance sheets runs through prime brokerage leverage neither register was built to catch.
MH

Two warnings, one week

Start with what the Bank of England actually said. Its Financial Policy Committee flagged three separate things to worry about: government borrowing, loans made outside regular banks, and the possibility that AI stocks are priced for a future that does not arrive. Three warnings, filed the same week. The European Systemic Risk Board, the EU's financial-stability watchdog, called frontier AI models 'a paradigm shift for cybersecurity.' Translated: the attackers get the upgrade first, and the defenders spend the next few years catching up. The European Central Bank did not wait around for a committee report to gather dust. Supervisory Board chair Claudia Buch wrote directly to Deutsche Bank, BNP Paribas, Santander and other major eurozone lenders, demanding AI-cybersecurity action plans by 31 October 2026. That is a real deadline, on a real letter, to real banks. And here is the gap: nothing on either regulator's desk requires a bank to file a plan for its prime brokerage book. Two warnings landed in the same week, pointed at two different risks, and both missed the pipe connecting them.

Wrong ledger, right number

Here is the exposure nobody put in a press release. UK and EU banks mostly do not lend directly to AI companies, so if an AI company goes bust, that is not actually the danger. Their real AI risk sits somewhere less obvious: prime brokerage, the desk that lends hedge funds cash and stock so those funds can run leveraged bets. Those loans are secured by collateral, meaning if a fund cannot pay, the bank seizes what the fund put up as security. And a lot of what's been put up as security lately is concentrated, momentum-driven AI stock. Regulators have modelled a 2-percentage-point shock to that kind of collateral. If it hits, a bank does not see it land as a bad AI loan on its books, because it never made one. It shows up instead as a margin call. Picture the Bank of England's proposed 'kill switch' for runaway trading algorithms: it would sit on the trading agents themselves, watching for erratic behavior. But the money doesn't move through the trading agent. The margin call still lands on the prime broker, the exact desk that deadline and that warning both forgot to mention.

The Bank for International Settlements puts global data-centre capex needs at $3.21 trillion through 2028, $1.75 trillion of it borrowed. AI-linked private credit, lending done outside ordinary banks, has gone from near zero to over $200 billion in just a few years. Gartner adds another figure: $234 billion in software spending it expects AI to displace by 2030. Notice something about all three numbers: every one has a sponsor selling something, whether that's a credit rating, a research subscription, or a seat at the table when the debt gets refinanced. Not one of them sits on a CFO's cyber-vendor exposure sheet. The bank statement doesn't have a line for it either. Ask anyway.

PREVIOUS COLUMNS, FINANCE & RISK DESK
The Wang Report's columns are produced by AI under human editorial oversight. See our Editorial Standards.