AI concentration risk: what it does to bank NPL ratios

0:05Host:You're listening to Leaders Insights. Today's subject: KKR flags AI concentration risk: what the credit binge means for bank NPL ratios now. KKR just told everyone the AI lending party might be overcrowded — so why should a bank CFO, who doesn't touch private credit, lose a minute of sleep over it?

0:23Expert:Because the borrowers overlap. KKR's co-CEO flagged AI concentration risk last week, and the names drawing on private credit lines are the same hyperscalers and data-centre developers your corporate loan book is quietly financing. The exposure isn't in one bucket. It's smeared across both.

0:41Host:Give me the artefact. What's the one number you'd put under the microscope right now?

0:46Expert:The cost of risk line. That's the charge a bank books each quarter for expected loan losses, measured in basis points of the loan book — a basis point being one hundredth of a percent. Most European banks are running it around 30 to 40 basis points right now. Historically calm. Suspiciously calm.

1:05Host:Suspiciously how?

1:07Expert:Because that number is backward-looking by construction. It's built on default data from 2023 to 2025, when AI capital spending was all upside and nobody missed a payment. You're pricing tomorrow's defaults with yesterday's good behaviour.

1:21Host:Walk me through the dashboard, then. What's actually in that cost-of-risk figure and where does it break?

1:27Expert:Three pieces. First, the NPL ratio — non-performing loans, the share of your book where the borrower has stopped paying, usually ninety days past due. Across the eurozone that's sitting near 1.9 percent, near record lows. Looks pristine.

1:43Host:And the pristine part is the problem.

1:45Expert:The pristine part is a lagging indicator wearing a costume. A data-centre loan doesn't go non-performing when the AI narrative cools. It goes non-performing eighteen months later, when the tenant renegotiates or the compute demand doesn't show. Your NPL ratio tells you about a recession you already had.

2:05Host:Second piece.

2:06Expert:Coverage ratio. That's how much you've provisioned against those bad loans — the cushion. Say you're at 45 percent coverage. Sounds responsible. But coverage is calculated against the loans you've already flagged as bad. If the bad loans are still disguised as good ones, your cushion is covering the wrong fire.

2:25Host:So the denominator's lying to you.

2:27Expert:The denominator's not lying. It's just answering a question nobody should be asking anymore. And here's where the vendor data gets interesting — Moody's and S&P both publish AI-exposure estimates, and I'd treat those with a raised eyebrow, because the rating agencies are selling you the model that measures the risk they also rate. Useful, conflicted, read accordingly.

2:50Host:Third piece.

2:51Expert:The staging logic under IFRS 9 — the accounting rule that forces banks to provision for expected losses before a loan actually sours. Loans sit in Stage 1 if they're healthy, Stage 2 if risk has climbed, Stage 3 if they're impaired. The trick is the trigger to move a loan from Stage 1 to Stage 2.

3:10Host:And let me guess — concentration doesn't trip that trigger.

3:14Expert:It doesn't. The models watch each borrower individually. They don't see that forty of your best credits are all betting on the same AI demand curve. So a correlated shock — everyone wobbling at once — arrives as a surprise the dashboard was never designed to catch.

3:30Host:What works, then? You've torn down three pieces. Anything in that stack you'd keep?

3:35Expert:The IFRS 9 architecture is sound — forward-looking provisioning was the right instinct after 2008. The failure is the input, not the engine. Feed it a scenario where AI capex drops thirty percent and watch which loans should migrate to Stage 2. Most banks haven't run that overlay because the base case still looks gorgeous.

3:57Host:So the cost of risk isn't wrong. It's just answering last year's exam.

4:01Expert:Exactly. Thirty-five basis points is the honest answer to the question "what did losses look like recently." It is a dishonest answer to "what will they look like if the AI story disappoints." Those are different questions, and the board keeps confusing them.

4:17Host:One thing a CFO does Monday morning.

4:20Expert:Pull every loan touching AI infrastructure — data centres, chip supply, the hyperscalers — and run a single correlated stress scenario across all of them at once, not name by name. If the combined Stage 2 migration scares you, your current cost of risk is fiction. Better to find out on a spreadsheet than in a filing.

4:40Host:This episode draws on Financial Times, Accounting Today, CFO Dive, Treasury Today. That's all. For an honest read on your level, the CFO assessment is at mba-training.com.