§0 · What's already been said, and what we're adding

Buy-now-pay-later research runs across four layers that never quite meet. The consumer-behavior layer is the oldest and most developed: Marco Di Maggio, Justin Katz, and Emily Williams (NBER Working Paper 30508, 2022) built a transaction-level panel of BNPL users and found the product boosts total spending and, for liquidity- constrained households, facilitates a "liquidity flypaper effect" — the extra spending power sticks where it lands rather than smoothing across time the way a standard lifecycle model would predict. Bian et al. (2023) and deHaan et al. (2024) extend this into overborrowing specifically. All of it treats the BNPL user as an individual, in isolation from any other loan they might be carrying.

The merchant-economics layer is newer and thinner: a single paper, Tobias Berg, Valentin Burg, Jan Keil, and Manju Puri's "The Economics of Buy Now, Pay Later: A Merchant's Perspective" (NBER Working Paper 33152, November 2024), is the first to model BNPL from the seller's side rather than the borrower's. Using randomized and observational data from a German e-commerce furniture retailer, they show formally that BNPL functions as a bundling device — a product plus a merchant-subsidized zero-interest loan — that lets a merchant price-discriminate between customers by creditworthiness. It is rigorous and it is the strongest peer-reviewed grounding available for anything BNPL-related. It is also written for merchants and finance academics, and it says nothing about what happens when a household holds five of these loans at once.

The regulatory layer knows about that part, even if it doesn't use the same words for it. The CFPB's own market report (December 2025) and a stream of Federal Reserve consumer surveys document "loan stacking" — their own, more clinical term — for exactly the condition this investigation is about: obligations that exist, in aggregate, nowhere a lender, bureau, or the borrower themselves can see them. "Phantom debt" is not regulatory language; it was coined by Wells Fargo senior economist Tim Quinlan in a December 2023 report ("BNPL: The Phantom Debt"), and financial press and some Fed research since have adopted it because it is, simply, the more vivid and accurate term. FICO built an entire new scoring model — Score 10 BNPL, developed with Affirm — specifically because the aggregate had become invisible enough to require new infrastructure. But regulatory reports describe a condition; they don't connect it to why the condition exists at the scale it does, or put a number on what it costs the households carrying it.

The fourth layer isn't academic at all. It's a live, unresolved public argument. The advocacy group Protect Borrowers has publicly claimed that BNPL's merchant costs get passed onto all consumers through higher prices, "regardless of whether they use BNPL." The American Fintech Council's response, on the record: "There is nothing to substantiate the claim." The Financial Technology Association's, flatter still: "This claim is an opinion, not a fact." Neither side has published a number. Nobody has actually run it.

None of this is news, exactly — the pieces are all sitting in public.

What this investigation adds: the historical frame connecting BNPL's sudden ubiquity directly to the specific, dated gap between flat real wages and a 29–36 percent rise in essential costs since 2020, rather than treating the product's growth as a self-contained fintech story; the structural comparison to layaway, BNPL's own hundred-year-old predecessor, and the single variable that flipped between them; a first-attempt, honestly-bounded dollar estimate for the secondary merchant-extraction mechanism, built from actual 10-K and IPO-prospectus filings rather than industry press releases; and the synthesis none of the four existing literatures were built to supply on their own.

One scoping choice is worth stating directly rather than leaving for a reader to wonder about. This investigation focuses on fintech BNPL — Klarna, Affirm, Afterpay, Sezzle — not on card-issuer installment products like Chase Pay Over Time, American Express Plan It, or Citi Flex Pay. That is a deliberate scope, not an oversight, and the reason cuts against naive intuition: card-issuer installment plans are actually the larger market by usage, not the smaller one. Credit-card installment plans beat fintech BNPL roughly 3-to-1 — 31 percent of consumers used one in the past three months, versus 12 percent for fintech BNPL. What makes fintech BNPL the sharper case for this investigation isn't its size. It's that a Pay Over Time or Plan It balance is just a repayment structure bolted onto a card the household already has, and it reports to credit bureaus the same way any card balance does. Fintech BNPL mostly doesn't report to bureaus at all — which is the specific gap FICO's Score 10 BNPL model exists to close, and the specific gap this investigation is about. §3 returns to card-issuer plans directly, because the picture there is less reassuring than it first looks.

The Zero Percent Illusion

It is one tap in the same payment sheet you already use to buy a coffee. Apple shut down its own in-house Apple Pay Later product in 2024 and replaced it by surfacing a third-party installment loan directly inside the Apple Pay button at eligible checkouts — no app to download, no separate signup, no moment where taking on debt feels different from paying. That is the actual level of friction a BNPL loan requires today, and it is worth sitting with before asking why "0% APR, four easy payments" now shows up at checkout for a couch, a plane ticket, a grocery order, and an emergency-room copay.

The honest answer isn't really about payment technology. It's that a lot of households can no longer complete an ordinary purchase in a single motion — not because they manage money badly, but because rent is up 36 percent and groceries are up 29 percent since 2020, while the same paycheck buys roughly 2 to 3 percent more than it did then, after inflation. Buy-now-pay-later did not create that gap. It is the current, best-marketed, lowest-friction way of financing across it.

Klarna's own founder makes the industry's real case honestly, and it deserves to be heard on its own terms: the average Klarna balance is $87, against a $6,730 average U.S. credit-card balance. A fixed, four-payment installment plan genuinely does not compound the way revolving card debt does. For a household using one BNPL loan at a time, on schedule, this is a fair alternative to a worse product — not a trick. §8 gives this argument its full due.

The common financial-advice version of the same argument is blunter still: take the free money. If a household has the cash to pay outright but finances at 0% instead, while that cash earns any return at all elsewhere, the arithmetic genuinely favors financing — real, uncomplicated arbitrage, not a trick played on the household. §8 develops why that arbitrage is sound for some households and a trap for others, and the difference turns out to hinge on exactly one thing: whether the household can actually follow through on the terms.

What this investigation argues is not that BNPL is predatory by design. It argues that the entire "one loan, fixed terms, no compounding" case was built for a world where a household takes out one BNPL loan. That is no longer the typical case. 63 percent of BNPL users are already carrying more than one loan at the same time, and roughly a third are borrowing from more than one provider — a fact no single lender, no credit bureau until this year, and often not the household itself, has been positioned to see. The mechanism this investigation names is not the loan. It is what happens to the loan's own logic once it stops being singular.

§2 · The mechanism and the historical arc

Financing routine consumption during a wage-price squeeze is not a fintech invention. It is close to a hundred years old. Installment buying in America traces to the years after the Civil War, when manufacturers realized more households could afford a sewing machine or a horse-and-buggy if payment could be spread across time. Sears opened its entire mail-order catalog to installment payment in 1930; by 1940, more than 30 percent of Sears catalog sales were transacted on installment. Layaway — pay a deposit, the store holds the goods, complete the balance over time — reached its peak popularity specifically during the Great Depression, when it served, in the plain language of the historical record, as "a practical solution for cash-strapped consumers unable to purchase essentials outright." The pattern holds across a century: when wages fall behind the price of ordinary living, American retail responds by financing the difference. It does not respond by lowering the price.

What changed is one variable, and it is the variable this whole investigation turns on. Layaway required the store to hold the merchandise until the customer had paid in full — the buyer never took possession before the debt was retired. That single design choice self-limited exposure by construction. A household could not quietly run five layaway plans across five different stores at once, because it would have no goods to show for any of them until each was paid off, and each plan was capped by what the household could pre-pay before receiving anything at all. Buy-now-pay-later inverts this exact mechanic. The household receives the product immediately and pays afterward, across as many providers as will approve the purchase, with no shared visibility connecting any of them. That inversion — not a lower interest rate, not a better checkout experience — is the actual structural innovation, and it is what makes the stacking mechanism in §3 possible at all.

A second, narrower mechanism rides alongside the first and deserves its own precise treatment, because it is the one with a formal economic model behind it. BNPL bundles a product sale with a loan the merchant, not the borrower, subsidizes. This is the load-bearing difference from a credit card, where the cardholder's own creditworthiness sets their own interest rate. Berg, Burg, Keil, and Puri show — with a worked model and real transaction data — that this is a textbook bundling-based price-discrimination device, in the analytical tradition running from George Stigler (1963) through Adams and Yellen (1976) to McAfee, McMillan, and Whinston (1989): because creditworthiness correlates negatively with willingness to pay, subsidizing the loan lets a merchant capture a low-willingness-to-pay customer at an effective discount while a high-willingness-to-pay customer keeps paying full list price. In their data, offering BNPL increased sales 20 percent, concentrated among the lowest-creditworthiness customers. BNPL providers additionally write "no-surcharge" clauses into merchant contracts — the identical legal structure behind decades of credit-card interchange economics — which means a merchant cannot recover its 4-to-7-percent BNPL fee by billing it only to the customers who triggered it. Whatever cost gets recovered has to move through the single price posted for everyone. The closest studied analog for how much of a fee like that actually reaches the shelf price is the Durbin Amendment debit-interchange literature, and its finding cuts against intuition: merchants passed through at most 28 percent of their interchange savings to consumers when the fee went down, and a comparable Spanish study found only about 17 percent pass-through. Applied honestly to BNPL, that puts the realistic range at 15 to 28 percent — real, but partial, and far short of the "your price doubled" story either side of the Protect Borrowers/fintech-industry fight has been implicitly gesturing toward without ever naming a number.

§3 · The instrument

The household's actual financial position, once BNPL usage moves past a single loan, is not "I have a buy-now-pay-later balance." It is a set of simultaneous, cross-provider obligations that exists nowhere as one number — not on a credit report until FICO's aggregation model finishes rolling out at bureau scale, not on a bank statement, not in the household's own running sense of what it owes, because each purchase felt like a self-contained decision at the moment of checkout. This is structurally distinct from every instrument this publication has priced before:

Here, no single party holds the total — not the household, not any one lender, and, until FICO's new model reaches full scale, not even the industry meant to police it. Sixty-three percent of BNPL users are already carrying more than one loan at once; a third are borrowing across more than one provider. "Phantom debt" — coined by a Wells Fargo economist, not a regulator, but adopted widely because it is exactly the right term — describes debt that is entirely real to the household paying it and entirely invisible to everyone positioned to see it coming.

Card-issuer installment plans — Chase Pay Over Time, Amex Plan It, Citi Flex Pay — complicate this picture rather than resolving it, and the complication is precise enough to state exactly. A Pay Over Time or Plan It balance does show up on a credit report, but only folded into the card's single combined balance. The bureau sees the total dollar amount owed on the card. It does not see what fraction of that total is locked into a fixed installment plan — a committed monthly payment, often with its own fee, that behaves nothing like ordinary revolving debt — versus genuinely flexible balance the household could pay down to the minimum in a tight month. Two households can report an identical balance and an identical utilization ratio to a lender while holding completely different actual flexibility, if one of them has committed half of that balance to a fixed schedule and the other hasn't. The invisibility this investigation is built around doesn't disappear once a BNPL-shaped product moves onto a card the household already has. It moves one layer deeper — from "how much do you owe," which is visible, to "how much of what you owe is actually flexible," which isn't visible to anyone reading the report.

§4 · Where you sit in the distribution

Fifteen percent of U.S. adults used BNPL in 2024, up from 12 percent in 2022; roughly 96.3 million Americans are projected to use it in 2026. Usage is not evenly spread. It concentrates in households with income between $20,001 and $50,000, among Black and Hispanic consumers, among women, and among adults under 45 — nearly one in five Americans under 45 has used BNPL, against 8 percent of those 60 and older. Renters use it at a rate 52 percent higher than homeowners. The credit-score skew is sharp: usage among people with scores below 620 runs at roughly three times the rate of people above 720.

The clearest single number in this section is also the most direct rebuttal to the idea that BNPL is mainly a convenience choice: 29 percent of BNPL users, per the Federal Reserve's Report on the Economic Well-Being of U.S. Households in 2025 (May 2026), cited "only way I could afford it" as their primary reason for using the service — not the cheaper option, not the more convenient one, the only route to completing the transaction at all. That figure is not evenly spread either — it tracks income almost linearly:

Separately, and by coincidence at the same headline rate: 29 percent of BNPL users have used the loans specifically to buy groceries — roughly double the share from two years earlier, and 38 percent among Gen Z users specifically (LendingTree/QuestionPro survey, March 2026). Among that grocery-and-food-delivery subgroup specifically, the Fed's own necessity figure climbs to 45 percent — households financing food are the most likely of all to say they had no other way to pay for it.

§5 · Who actually pays

The primary population is BNPL users themselves — disproportionately the lower-income, lower-credit-score, renting, under-45 households §4 describes. The industry's steelman says this population wins outright: an effectively subsidized price, and a fixed installment structure that does not compound the way credit-card debt does. The data complicates that claim without discrediting it entirely. Self-reported late payment has climbed three years running — 34 percent of BNPL users in 2024, 41 percent in 2025, 47 percent in 2026 (LendingTree's BNPL tracker) — a trajectory, not a one-time blip. Sixty-three percent are carrying multiple simultaneous loans.

Charge-off (actual default) rate, for comparison: flat at roughly 1.8–2 percent across the same period — CFPB's dollar-weighted 2023 rate was 1.83 percent. Late payment is climbing; loans failing outright are not.

Worth stating precisely, because it cuts against an easy overstatement: late payment is not the same as default. The more precise and more honest claim is narrower: a fast-growing share of BNPL users are struggling to pay on time, even while relatively few are losing the loan outright. That gap between "paying late" and "defaulting" is exactly what stacked, cross-provider exposure would produce — a household juggling several obligations misses individual due dates without any single loan actually failing, because it's making do across all of them rather than failing any one of them.

A sharper version of the same point shows up in the two largest providers' own numbers. Klarna's US financing balances 30-plus days past due improved 36 basis points off their Q2 2025 peak; Affirm's 30-plus-day delinquency on monthly installment loans sits flat at 2.8 percent (both per Q1/Q2 2026 earnings disclosures). Each individual lender's own book looks stable, even improving. That is not a contradiction of the phantom-debt thesis — it is exactly what the thesis predicts. No single lender's delinquency metric was ever going to capture a household's total exposure across five different apps; that is precisely the blind spot this investigation is about, and Klarna's and Affirm's own healthy-looking numbers are the clearest evidence that the aggregate risk is invisible from inside any single provider's ledger, not proof that the risk doesn't exist. And the product mix itself is drifting toward something the founders' framing does not describe: of the Federal Reserve's own $156.7 billion 2025 US BNPL issuance total, $47.1 billion — about 30 percent — was in the "longer-term loans" category rather than the fixed, short Pay-in-4 structure Klarna's founder describes, a segment more likely to carry disclosed interest rather than the fixed 0% terms the industry's own steelman rests on. A population already carrying more financial precarity than average (higher unpaid card balances, more past-due bills, heavier use of payday loans and pawn shops, per the CFPB's own data) is the same population accumulating stacked exposure across a product mix that is quietly drifting away from the fixed, non-compounding structure the "healthier than a credit card" case depends on.

A secondary population — non-BNPL payers, who skew higher-income and higher-credit-score by the same data — may absorb a modest, partial share of the aggregate merchant-fee cost through the single posted price, per the no-surcharge mechanism in §2. If the 15-to-28-percent pass-through estimate holds, this transfer runs in the opposite distributional direction from every other investigation in this publication: it would mean higher-income households mildly subsidizing price access for lower-income ones, which is not obviously an extraction in the usual sense — it may be closer to an accidental, uncounted, unvoted-on redistribution running through contract law rather than policy. That is worth stating plainly rather than smoothing into the shape of a familiar villain. The central harm this investigation names is not that transfer. It is the stacking risk borne by the BNPL-using population itself, which the industry's own single-loan convenience narrative has no account of.

§6 · The scale

The honest scale metric for this investigation is prevalence, not a single dollar figure, and that is a genuine departure from how this publication usually closes its numbers section — worth saying to the reader directly rather than papering over. Phantom debt, by definition, is not counted anywhere; inventing a precise dollar total for it would manufacture a false confidence this essay is specifically arguing against. What is measurable: roughly 96.3 million projected U.S. BNPL users in 2026; 63 percent of them carrying multiple simultaneous loans; about a third borrowing across more than one provider; and a self-reported late-payment rate that has climbed three years running — 34 percent (2024), 41 percent (2025), 47 percent (2026) — even as actual charge-off rates stay low, around 1.8 to 2 percent. These numbers describe the actual shape and trajectory of the exposure in a way a single manufactured dollar figure could not.

The secondary, dollar-denominated mechanism is measurable, and here is the calculation in full. U.S. gross merchandise volume in the specific 0-percent-APR and Pay-in-4 short-installment segment — not Affirm's broader interest-bearing loan book, which is a different, transparently-priced product accounting for 72 percent of that company's volume — ran approximately $110 billion in 2025, per the Federal Reserve's own June 2026 breakdown of "Pay in 4" ($78.3 billion) and comparable short-term loans ($31.2 billion). At the cross-validated merchant-fee rate of 4 to 7 percent (Klarna's base rate is 3.29 percent plus $0.30; Affirm and Sezzle run closer to 6 percent plus $0.30; Afterpay charges 4.99 percent plus $0.30), total U.S. merchant fees on this segment come to $4.4 to $7.7 billion a year. Applying the empirically grounded — not assumed — interchange pass-through rate of 15 to 28 percent yields an estimated transfer to non-BNPL-benefiting consumers of roughly $700 million to $2.2 billion a year. For scale, separately: BNPL late-fee revenue collected nationally was $80.3 million in 2023, down from $93.8 million in 2021 — small, and falling, and not the number this essay is built around.

No existing study — academic, regulatory, or advocacy — has published either figure. Both are first attempts, built from primary filings rather than industry estimates, and both are presented here with their derivation and their limits stated in the open, not buried in a footnote.

§7 · The asymmetry

This publication grounds every investigation in George Akerlof's 1970 analysis of information asymmetry and Xavier Gabaix and David Laibson's 2006 work on shrouded attributes, and this one is no exception — but the shape of the asymmetry here is unusual, and worth naming precisely. It is not simply that the counterparty knows more than the household, the way an employer knows the exact float value of a delayed paycheck. It is that, until very recently, no party could see any individual household's aggregate exposure — not the household, not any single lender, not even a regulator working from survey-level statistics rather than a real-time ledger — because the system was built without any of them needing to. Independent underwriting was not an oversight. It was, and remains, the default architecture of the entire product category. FICO's Score 10 BNPL model — built in partnership with Affirm specifically to correct this — is the clearest evidence available that the asymmetry is structural rather than rhetorical: an entirely new piece of credit-scoring infrastructure had to be invented because the aggregate genuinely could not be seen by anyone before it. It is still, as of this writing, mid-rollout rather than fully live: the scores exist and are offered to lenders at no additional cost, but full effect depends on BNPL data reaching the credit bureaus at scale, which has not yet happened. The gap between "the fix has been built" and "the fix is deployed" is itself worth naming — solving the technical problem and closing the actual blind spot are not the same milestone.

None of this is evidence that markets fail. It is closer to the opposite, and it would be dishonest to treat it otherwise. BNPL is a private-sector innovation that emerged with no mandate behind it, competing on price and terms in a market consumers were entirely free to ignore, and by the ordinary test — do people keep choosing it — it is working: usage has grown by every measure cited in §4. The invisible hand is visibly, demonstrably operating, and it is disciplining exactly what is visible to it. Late-fee revenue fell from $93.8 million in 2021 to $80.3 million in 2023 even as usage climbed (§6) — providers competed that friction away because it was a number consumers could see and compare across apps. That is the market doing precisely what price competition predicts when information is available to the parties transacting. What Akerlof's 1970 extension of that same framework shows — and what this investigation is really about — is that the invisible hand can only discipline what its participants can see. A consumer comparing BNPL providers can see the fee. No consumer, no individual lender, and until FICO's model finishes rolling out, no regulator either, could see a household's total exposure across providers, because the product was built so that none of them would ever need to. The market is not failing to correct this. It has no mechanism by which it could notice.

§8 · The case for change

The steelman, in full: buy-now-pay-later is a real improvement over revolving credit-card debt in the single-loan case. Fixed terms. No compounding. And, per Berg, Burg, Keil, and Puri's own empirical finding, it expands market access to households a uniform-price merchant would otherwise turn away entirely — not a marginal benefit, but the difference between a sale and no sale. Klarna's $87-average-balance comparison against a $6,730 average card balance is honest evidence for the population using the product as its own founders describe it, and this investigation does not dispute it.

The steelman has a sharper form than the industry usually states it, and it deserves to be made explicitly rather than left as an aside. If a household holds the cash to pay for a purchase outright but finances it at 0% instead while that cash earns any return at all elsewhere — even the yield on an ordinary savings account — the arithmetic is not close. Financing at 0% while keeping the cash working is straightforwardly better than paying cash, and better still than paying early: it is the same treasury logic a business applies to supplier payment terms, capturing the spread between the cost of credit (zero) and the return on cash (whatever it is) as pure, riskless arbitrage. "Take the free money" is not naive financial advice. It is correct, for a household with the capacity to execute it.

That capacity is the entire question, and it is where the steelman runs out. A business capturing the identical arbitrage on supplier terms does it through a treasury or accounts-payable function built for exactly this purpose: a consolidated ledger of every obligation, its due date, and the cash available to meet it, forecast forward in one place. A household attempting the same move typically has no equivalent — no consolidated view across four or five different BNPL apps, no forecasting, nothing but memory and however many push notifications each provider happens to send. The arbitrage is real. The infrastructure to execute it safely does not exist for most households, which is the same gap §3 names as phantom debt, arrived at from the opposite direction — not "why is this dangerous" but "why can't more households do this rationally, the way a business would." Roughly 96 million people are effectively being asked to run an ad hoc treasury operation with none of a treasury desk's tools, and the climbing late-payment rate in §5 is what that gap looks like once it goes wrong.

What the steelman does not account for is what happens once usage moves from "one loan, paid on schedule" to "the default way to pay for nearly everything" — which is precisely how the product is now marketed and deployed, at checkout, for groceries and rent and healthcare as readily as for furniture. At that scale, the fixed-terms, no-compounding case stops being the operative fact. The operative fact is that no one — not the household, not any single lender — can see the total across providers, and 63 percent of users are already living the aggregated version of the product that the industry's own marketing has no language for.

Why hasn't the market corrected this on its own?

On the dimensions consumers can actually see — price, fees, terms — it has: providers compete on rate, and have driven late-fee revenue down even as usage climbed (§6), which is the ordinary discipline of an open market working exactly as intended. Fragmented underwriting sits entirely outside that competitive discipline: no individual BNPL provider has a commercial incentive to make a customer's exposure to its competitors visible, since doing so could only ever lower its own approval rate, and no consumer can shop for a lender's transparency about a number the lender itself does not have. It is also worth being precise about what "no government intervention" means in this specific case, since the mechanism does not sit in a policy vacuum: the no-surcharge clauses in §2 are enforceable because ordinary contract law enforces them, and the CFPB's withdrawal of the 2024 BNPL Interpretive Rule on May 12, 2025 — followed by its statement that it will not prioritize Truth-in-Lending enforcement against BNPL providers — was an affirmative policy decision, not government's mere absence. For the stacking mechanism specifically, what's missing is not a fix for a market failure in the usual sense — it's a disclosure requirement on the one dimension the market has no mechanism to discipline itself on, layered onto commercial law that already fully exists. (The state-level levers below, aimed at the regulatory gap more broadly, are a different and more direct kind of intervention; they are not standing in for a disclosure fix, they are addressing a separate hole federal withdrawal left open.) And the two populations that would need to complain for ordinary market pressure to build have no reason to: the approved, successfully-repaying user has no complaint to file, and the user accumulating stacked exposure across providers mostly cannot see it happening until a payment is missed.

With the federal government pulling back, states have moved into the gap, and this is the strongest precedent available — not a foreign regulator's proposal, but domestic law already on the books. New York and Illinois have both passed BNPL-specific consumer-protection statutes. New York's is the sharper of the two: it requires BNPL providers to be licensed and subjects BNPL loans to the state's 16 percent interest-rate cap — the first time a major mechanism in this investigation has been directly addressed by enacted, not proposed, law. A U.S. House Financial Services Committee discussion draft on CFPB reform, released July 24, 2026, is open for public comment through August 21, 2026 — a live, current opportunity, not a hypothetical one.

There is real precedent that the underlying invisibility problem is solvable, not merely diagnosable. FICO's Score 10 BNPL model, built with Affirm, is the first working piece of cross-provider aggregation infrastructure — proof of concept that the technical barrier is not the obstacle, even though (per §7) it has not yet reached full deployment. Three targeted levers, each aimed at a specific mechanism rather than offered as an undifferentiated menu:

§9 · Methodology + sources

Sources are tiered by verifiability. Tier 1 is regulatory filings and government data. Tier 2 is academic research with attributed methodology. Tier 3 is commentary and live policy debate. All editorial constants used in the essay are documented on the methodology page.