KKR flags AI concentration risk: what the credit binge means for bank NPL ratios now

The three metrics at the centre of bank credit portfolio management, the NPL ratio, the coverage ratio, and the cost of risk, are individually straightforward. The confusion arises in how they interact, how supervisors read them together, and what happens when a single sector drives disproportionate deterioration across all three simultaneously. KKR's recent warning about concentration in AI-related lending is a useful stress test for that interaction.

Why bank CFOs should care about KKR's AI credit warning

KKR's concern is specific: lenders, including banks and private credit funds, have extended significant volumes of credit to AI infrastructure borrowers, data centre developers, GPU financing vehicles and the like. The investment firm flags overexposure and concentration as the twin vulnerabilities. If the sector corrects, the defaults will not be evenly distributed across a diversified portfolio. They will cluster in the books of whoever was most active in the space between 2023 and 2026.

For a bank CFO, this matters structurally. Banks turn deposits into profit by pricing credit risk at origination and managing it through the cycle. When a portfolio has sectoral concentration, the usual assumption that losses are idiosyncratic and partially offsetting breaks down. A wave of correlated defaults in AI-related exposures would hit NPL ratios across multiple counterparties at once, forcing simultaneous provisions and compressing net interest margin precisely when funding costs tend to spike.

The regulatory dimension is not hypothetical. Under Basel III and its successor frameworks, banks are required to hold capital against expected and unexpected losses. Supervisors, whether the ECB's Single Supervisory Mechanism for eurozone banks or the PRA in the UK, routinely examine sectoral concentration in credit portfolios and can require additional Pillar 2 capital buffers if they judge that the internal model underestimates correlation risk. A bank that has grown AI-related lending without a corresponding concentration framework is holding a regulatory conversation it does not want to have during a downturn.

How NPL ratio, coverage ratio, and cost of risk actually work together

The NPL ratio is the starting point: non-performing loans divided by total gross loans. A bank with 2bn in NPLs on a 100bn loan book reports a 2% ratio. The EBA defines non-performing as any exposure more than 90 days past due or assessed as unlikely to be repaid without collateral enforcement. The ratio tells you the stock of impaired credit.

Coverage ratio divides loan loss provisions (the accounting reserve) by the NPL stock. If that same bank holds 1.4bn in provisions against 2bn in NPLs, coverage is 70%. This ratio tells you how much of the bad debt the bank has already absorbed through the income statement. European banks averaged around 45% coverage in the early post-financial crisis period; by 2025 many had pushed above 60% under ECB pressure, though the range across institutions and geographies remains wide.

Cost of risk is the dynamic measure: new provisions recognised in a given period divided by average gross loans, typically expressed in basis points. A bank reporting 40 basis points cost of risk is setting aside 0.4% of its loan book annually to cover anticipated losses. This is the figure most sensitive to forward-looking assumptions under IFRSIFRSThe global accounting rulebook that governs how companies report financial results, used across the EU and 140+ jurisdictions.View full definition → 9, where Stage 2 migrations (loans where credit risk has increased significantly since origination) force earlier recognition of expected credit losses.

Here is where the AI concentration scenario becomes concrete. Suppose a bank has originated 8bn in loans to data centre operators and AI infrastructure funds over the past three years, representing 8% of total loans. The borrowers are current, so NPL ratio looks fine today. But if rising rates, delayed AI monetisation, or a correction in GPU valuations causes a cluster of those borrowers to migrate from Stage 1 to Stage 2, the bank must immediately recognise lifetime expected credit losses on those exposures. Cost of risk spikes, the coverage ratio on the Stage 2 pool looks thin relative to the probability of further migration, and analysts start asking about NPL ratio trajectory. All three metrics deteriorate in sequence, not because the bank made poor decisions in isolation, but because concentration collapsed the diversification assumption embedded in its ECL model.

Understandinghow banks model expected credit losses through the cycle is where the technical depth lies: the probability of default curves, loss given default assumptions, and the macroeconomic overlays that management applies on top of the model output.

When NPL ratio alone misleads and coverage ratio is the better diagnostic

The NPL ratio is a lagging indicator. A portfolio can look clean right up to the moment it does not, particularly when borrowers are refinancing distressed positions rather than defaulting. The Paramount $52bn financing package, which priced near 9% yield in 2026 to fund the Warner Bros deal, illustrates the point: the loans are performing, the ratio is zero, but the cost of capitalcost of capitalThe blended rate a company pays to finance itself through debt and equity. It sets the minimum return an investment must clear to create value.View full definition → embedded in that pricing reflects something between distressed and high yield. Coverage ratio is more instructive because it shows management's own estimate of loss embedded in the current book.

That said, coverage ratio can also be managed. Banks have discretion in staging decisions and in the macroeconomic scenarios they weight. A bank reporting stable coverage while aggressively keeping borrowers in Stage 1 is not necessarily safe; it is deferring recognition. This is precisely the kind of pattern thatstress testing against a coherent risk appetite framework is designed to expose, both internally and under supervisory review.

Cost of risk is the most operationally useful of the three for a CFO running a credit portfolio, because it directly affects earnings guidance and capital planning. A 10 basis point increase in cost of risk on a 100bn loan book is 100mn of additional provisions before tax. That is not an abstract number; it flows through to CET1 ratios, dividend capacity, and the bank's ability to grow its book in the next period.

The honest tradeoff is that managing cost of risk through the cycle requires accepting short-term earnings volatility to build coverage before defaults crystallise. Banks that smooth provisions to meet consensus forecasts tend to have thinner coverage when the cycle turns. The ECB has called this out in multiple SREP cycles and it remains a live supervisory concern heading into a potential AI credit correction.

KKR's warning is most useful not as a prediction but as a prompt to audit portfolio concentration now. A bank CFO who cannot quickly answer how much of the loan book is exposed to AI-related borrowers, what the Stage 2 migration would look like under a stress scenario, and whether current provisions reflect that risk is already behind the question supervisors will eventually ask.

The full course on this sector:Finance in Banking.