Here are two numbers that shouldn’t both be true at the same time.
According to a 2025 LendingTree survey, 41% of people who use buy now, pay later services made at least one late payment in the past year. Buy now, pay later refers to the checkout installment plans offered by companies like Klarna, Affirm, and Afterpay. That late-payment rate has been climbing. A year earlier it was 34%.
According to the Consumer Financial Protection Bureau, the charge-off rate for buy now, pay later loans was 1.8% in 2023, down from 2.6% the year before. A charge-off is the share of outstanding debt that lenders have officially written off as unrecoverable. That rate has been falling steadily.
Same borrowers. Same behavior. Opposite signals.
One of these numbers is generated by borrowers describing their own experience. The other is generated by lenders reporting on their own books. They are measuring the same thing. They are not close to agreeing.
That gap is what this piece is about.
The working hypothesis: Buy now, pay later’s clean balance sheet is a measurement artifact — produced by a reporting architecture that keeps this debt invisible to every other lender, sustained by borrower behavior that prioritizes these payments above everything else, and confirmed by the delinquency signals already leaking through on the instruments that do get measured. The stress isn’t missing. It’s been displaced.
The consensus case, fairly stated
Before departing from the mainstream view, it deserves its due.
Buy now, pay later, often shortened to BNPL, lets consumers split a purchase into installments. The typical structure is four equal payments over six weeks, sometimes longer. The platforms charge merchants a fee rather than charging consumers interest on short-term plans. The products exploded in popularity during the pandemic, reached roughly 180 million loans in 2021, and have kept growing.
The official data looks fine. Charge-off rates are low and falling. The platforms tightened their underwriting after the frothy 2021 period. Affirm began reporting all of its loans to Experian in April 2025, with TransUnion reporting starting May 2025 — a meaningful transparency step. The Federal Reserve Bank of Richmond published a research note in February 2026 describing BNPL as an increasingly important but still relatively small component of the consumer credit landscape, and calling for continued monitoring.
That’s a reasonable, well-sourced view. Serious analysts hold it.
The problem isn’t that the consensus is wrong. The problem is that it’s reading a partial ledger.
The instruments that do report are already telling a different story
Start with auto loans.
In January 2025, 6.56% of subprime auto loan borrowers were at least 60 days past due on their payments. Fitch Ratings defines subprime as borrowers with credit scores of 640 or below, and has tracked this figure since the early 1990s. That 6.56% reading was a series record. For context, the same rate had been running below 3 percent at the pandemic-era low in 2021. The share of all auto loan balances transitioning into serious delinquency, meaning 90 days or more past due, hit 3% in the fourth quarter of 2024, the highest level since 2010. Cox Automotive estimates 1.73 million vehicles were repossessed in 2024, the highest total since 2009.
The economy, by every conventional measure, was not in recession during this period. Unemployment stayed historically low. That combination, recession-era repossession numbers alongside near-full employment, doesn’t have an obvious innocent explanation. Something is wrong with the consumer balance sheet that the headline economic data is not capturing.
Then look at mortgages.
The government-backed mortgage programs, specifically Federal Housing Administration loans and Department of Veterans Affairs loans, showed serious delinquency rates rising 70 and 57 basis points respectively year-over-year as of the fourth quarter of 2024, while conventional mortgages moved only 2 basis points. This data comes from the Mortgage Bankers Association’s National Delinquency Survey. By the end of 2025, FHA serious delinquencies were up 104 basis points year-over-year. This matters because FHA and VA loans have the most lenient qualifying standards. They serve borrowers with lower down payments, thinner credit histories, and less financial cushion. They are the products most likely held by borrowers who also carry buy now, pay later debt that no mortgage underwriter has ever seen. The fracture is not showing up uniformly across the mortgage market. It is showing up precisely where you would expect it to if a hidden layer of consumer debt were degrading borrower quality at the margins.
Credit cards tell a similar story. Charge-off rates hit a 13-year high in the third quarter of 2024 before stabilizing. The stabilization sounds like good news. The Federal Reserve Bank of Kansas City noted that subprime borrowers pulled back from credit card usage entirely during this period, not because their finances improved, but because they stopped taking on new card debt. That is a different kind of outcome than recovery.
Three instruments. Three different stress patterns. All concentrated in the same borrower cohort.
Why the BNPL number looks fine
There are three interlocking reasons the buy now, pay later platforms report clean books while the borrowers who use them are showing strain everywhere else.
The architecture. BNPL platforms have no obligation to report loans to the major credit bureaus. Most don’t. The Consumer Financial Protection Bureau had to use a special supervisory data order just to study the market, because ordinary reporting mechanisms didn’t produce enough information to analyze. Affirm, to its credit, now reports to two of three major bureaus. But Affirm acknowledged in its investor communications that those loans go into specialty files separate from the core credit profile a mortgage lender or auto lender would pull — meaning other lenders cannot see those loans in standard underwriting. Klarna, Afterpay, Sezzle, and Zip remain largely dark. A borrower can carry four simultaneous BNPL balances across four different platforms and no external lender sees any of it.
Borrower prioritization. When a household budget gets tight, debts don’t get paid equally. Borrowers make choices. For BNPL, the consequence of falling behind is immediate and concrete: access to the service disappears. No more checkout installment plans. For a borrower living close to the margin, that access has real value. It is how purchases get made. So BNPL gets paid first. Credit cards absorb the delinquency instead. The Richmond Fed identified this mechanism explicitly in its 2026 research: BNPL loans require autopay, the payment amounts are small relative to other credit products, and preserving access creates a strong behavioral incentive to stay current. The BNPL balance sheet looks healthy not because borrowers are healthy, but because distressed borrowers are protecting their BNPL access at the expense of instruments that actually report.
Extend and pretend. When a BNPL borrower does fall behind, platforms have a strong incentive to restructure the payment rather than recognize the loss. An extended plan keeps the debt classified as performing. A charge-off forces a loss onto the books. So platforms extend, the debt stays technically current, and the official charge-off rate stays low. Klarna’s own history offers a preview of what happens when this dynamic encounters public market disclosure requirements for the first time. When Klarna completed its initial public offering on the New York Stock Exchange in September 2025, its prospectus emphasized strong underwriting and declining credit losses. Two months later, in its first earnings report as a public company, Klarna disclosed a net loss of $95 million as credit loss provisions jumped from 0.44% to 0.72% of gross merchandise volume, a 64% increase in a single quarter. A securities class action followed, alleging the offering documents had understated credit risk. The provisions that appeared suddenly in November had not accumulated suddenly. They had been there.
We’ve seen this before
During the pandemic, federal student loan payments were paused for 43 months. A subsequent 12-month on-ramp period then kept missed payments from being reported to credit bureaus even after payments resumed. The result was that the serious delinquency rate on student loans sat below 1% for years.
The on-ramp expired in October 2024.
In the first quarter of 2025, according to the Federal Reserve Bank of New York, 13.7% of student loan borrowers, nearly six million people, had at least one loan 90 days past due or in default. That matched pre-pandemic levels exactly. The stress had not gone anywhere. The reporting architecture had simply made it invisible, and when the architecture changed, the number reappeared.
Buy now, pay later is sitting in the equivalent moment right now. One major provider is reporting to the bureaus. The data goes into specialty files that underwriters cannot access. The other major platforms remain dark. The adjacent instruments, auto loans, FHA mortgages, credit cards, are already showing the strain.
The question isn’t whether this resolves. It’s what the number looks like when the measurement catches up.
What would change my mind
This argument has specific, testable implications. Three things would meaningfully undercut it.
First: Fannie Mae or Freddie Mac incorporates BNPL debt into the debt-to-income calculations used in conventional mortgage underwriting, applies that methodology retroactively to loans originated between 2022 and 2024, and finds that fewer than 3% of those loans would have been declined. If the invisible debt load isn’t distorting mortgage qualification at scale, the displacement argument loses its most consequential leg.
Second: By the fourth quarter of 2026, Affirm’s post-April 2025 loan vintage, now visible in bureau specialty files, shows a serious delinquency rate below 2%. That would suggest either that extend-and-pretend isn’t operating at the scale the adjacent data implies, or that Affirm’s underwriting genuinely separates it from the rest of the market. Either outcome matters.
Third: Klarna’s credit loss provisions, which jumped 64% in the quarter after its initial public offering, stabilize below 0.60% of gross merchandise volume for two consecutive reporting periods in 2026, without a meaningful contraction in its U.S. loan book. If provisions normalize without portfolio shrinkage, the post-IPO deterioration was noise, not signal.
Until one of those conditions is met, the gap between what BNPL reports and what borrowers experience remains the most interesting unresolved question in consumer credit.
Related: What Zillow Can’t Show You — the same reporting-architecture gap appears in housing: prime-cohort mortgage stress shows up in non-agency RMBS data that the headline MBA survey doesn’t capture, concentrated in the same Sunbelt borrower cohort carrying BNPL balances.
If you found this useful, the best thing you can do is forward it to one person who would push back on it. I’d rather be wrong in public than right in private.