Every digital payment rail charges something fixed for each transaction. A fixed fee doesn't care how small the payment is, so as the payment shrinks the fee becomes the whole thing. Somewhere near a dollar the arithmetic stops working and the transaction simply never happens.
You can watch it happen. The chart below sweeps from the largest payments in the economy down to the smallest, one rail at a time. Volume climbs the whole way as payments get smaller — right up until it doesn't.
Every U.S. payment rail, drawn from its published annual transaction count and average payment size: 253.8 billion payments a year across wires, ACH, checks, RTP, FedNow, cards and cash. Volume peaks near $25 and then falls off a cliff at $1.75, where a card fee passes a fifth of the payment. The gold area is what the same trend implies below that line.
Nobody wires $40 and nobody buys a house with a debit card. Each rail occupies a band of transaction sizes, and the bands are set almost entirely by how the fee is structured: a flat charge pushes a rail upmarket, a percentage charge pushes it down. Here is the actual U.S. picture.
| Rail | Payments / yr | Total value / yr | Average payment | What it costs to send | Year |
|---|---|---|---|---|---|
| Fedwire (wire transfer) | 217.3 M | $1,148 T | $5,283,000 | $25–35 retail | 2025 |
| FedNow | 8.4 M | $0.85 T | $101,435 | $0.043 | 2025 |
| Checks | 9.2 B | $24.5 T | $2,653 | ~$1–2 all-in | 2024 |
| ACH | 35.2 B | $93 T | $2,642 | fractions of a cent at the operator | 2025 |
| RTP (The Clearing House) | ~0.50 B | $1.3 T | $719 (2024) | $0.045 | 2025 |
| ATM cash withdrawal | 3.4 B | $0.72 T | $210 | $0–4 | 2024 |
| Credit cards | 67.1 B | $6.51 T | $97 | 2.4–3.5% + $0.10–0.30 | 2024 |
| Debit cards (non-prepaid) | 99.3 B | $4.34 T | $44 | $0.21 + 0.05% capped | 2024 |
| Prepaid cards | 21.3 B | $0.65 T | $31 | 2%+ + fixed | 2024 |
| Cash | ~21 B est. | — | ~$22 est. | no marginal fee | 2024 |
| x402 (agent payments) | 165 M cumulative | $0.05 T | $0.30 | no protocol fee | 2026 |
Each rail plotted at its average payment size and its annual transaction count. The relationship is close to a straight line on log-log axes over eight orders of magnitude — and it stops dead at the bottom.
A fee of 2.9% plus 30¢ sounds like a percentage with a rounding error attached. It isn't. Written as a share of the payment it becomes 2.9% + $0.30 ⁄ v, and that second term goes to infinity as the payment gets small. On a $100 order the fixed part costs you 0.3 points. On a $1 order it costs you 30 points.
Effective fee rate by transaction size. The horizontal line is a viability threshold — the point where the fee is eating so much of the payment that the payment stops being worth making. Where each curve crosses it is that rail's floor.
Until 2010 the card networks forbade merchants from setting a minimum purchase amount, because a minimum makes a card less cash-like. Retailers went to Congress and asked for relief, and the Durbin Amendment gave it to them: a merchant may now refuse a credit card on any purchase below ten dollars.
Visa, Mastercard and Discover all rewrote their operating rules to match. That handwritten "$10 minimum" sign taped to the register at the coffee shop is the fee floor made law — a formal admission by the payments industry that below a certain ticket size, the rail cannot pay for itself.
The other admission is quieter. The Federal Reserve notes that one of the reasons cash survives is the absence of marginal transaction costs, and seven cash payments per consumer per month have refused to disappear for four straight years. Seventy percent of them are under $25. Cash is what the sub-dollar economy uses because it is the only rail with a fixed fee of zero — and it doesn't work over the internet.
Between $10 and $500 the payment size distribution is well measured and behaves like a power law: each halving of transaction size brings roughly a fixed multiple more transactions. Fit that curve where the data is good, extend it into the region the card rails can't serve, and the gap between the extension and reality is an estimate of the missing economy. Then move the fee and watch the floor move.
Solid line is measured. Dashed line is the fitted power law extended below the floor. The shaded band is the region your fee opens up that card rails cannot reach.
Value per decade rather than transaction count, because the count is the eye-catching number and the value is the one that matters. Both curves climb with transaction size — there is more money in big payments, obviously. What matters is where they separate.
Not as far as you might hope, if dollars are what you care about — and much further than you might hope, if transactions are. The two answers come apart, and the gap between them is the most useful thing in this model.
Share of the total opportunity captured, as the fixed fee falls from fifty cents to a hundredth of a cent. Two curves, because value and volume are unlocked at completely different points.
A power law extrapolated three orders of magnitude past its data deserves suspicion. So here is an entirely independent way to size the same thing: look at what we built instead of sub-dollar payments, and add it up.
The sub-dollar economy is not missing. It is being monetised by proxy, through three workarounds that all exist because you cannot charge someone eight cents.
Advertising is the big one. An ad impression is worth a fraction of a cent, and the only way to collect a fraction of a cent from a reader is to sell their attention to a third party instead. U.S. digital advertising took in $294.6 billion in 2025.
Prepaid balances and app-store wallets are the second. You cannot buy a 40¢ item, so the platform sells you $4.99 of gems and meters it out, keeping 15–30% for running the float. Global in-app purchase revenue was $167 billion in 2025.
Subscriptions and bundles are the third: charge $11.99 monthly because you can't charge 30¢ an article, and accept that most subscribers are paying for things they never read.
The model's figure for the one-cent-to-one-dollar band, against the observed size of the markets that exist to work around it.
None of this is a new idea, and the last several attempts failed. Any honest version of this argument has to say why.
In 1999 Nick Szabo made the argument that has held up best. The binding cost of a small payment, he said, is not the fee — it is the thought. Deciding whether a thing is worth eight cents costs more than eight cents in attention, and no amount of protocol engineering reduces it. Andrew Odlyzko's The Case Against Micropayments made a parallel case: the obstacles were never technical, they were economic, sociological and psychological. Users prefer flat fees because flat fees are restful. This is why bundles won.
The objection is correct and it is also narrower than it looks. Mental transaction cost is a cost borne by a human deciding, one decision at a time. It scales with the number of decisions, not the number of payments. Two things break that link.
Budgets. One decision — "spend up to $5 a month on articles" — authorises thousands of payments. The deciding happens once.
Agents. Software has no mental transaction cost. When a program pays another program for an API call, there is no attention to conserve, and the only remaining cost is the one Szabo said would stop mattering: the fee.
This is not speculative any more. The x402 protocol has settled roughly 165 million transactions across about 69,000 active agents, at an average size near $0.30. Independent analysis of the Base deployment found that payments between 10¢ and $1 made up nearly half of all volume in early 2025. The volumes are small and a meaningful share is still testing rather than commerce. The shape of the distribution is the point: the moment the fixed fee disappears, payments appear below the floor.
Which suggests the real sequence isn't that cheap payments create a micropayment economy for people. It's that cheap payments create one for machines, and people arrive later, through agents that already have a budget.
The model is four lines of arithmetic. All of it is here so you can disagree with it precisely.
Let S(v) be the share of consumer payments larger than v. The Federal Reserve's Diary of Consumer Payment Choice reports that about 25% of payments are under $10, about 50% under $25, and 13 of 43 monthly payments exceeded $50 — so S(10)=0.75, S(25)=0.50, S(50)=0.30. Fitting a Pareto form S(v)=Cv−α to the outer two anchors gives α = 0.569, which then predicts S(25) = 0.445 against an observed 0.50. Close enough to use, not close enough to be smug about.
Because α is below 1, the total value under the curve converges as transaction size approaches zero even though the transaction count diverges. The dollar figures are therefore finite and well-behaved; the count figures are the ones to treat carefully.
The Diary reports 48 payments per consumer per month at an average of $141. Across roughly 262 million U.S. adults that is 151 billion consumer payments a year worth $21.3 trillion — which lands within a few percent of U.S. personal consumption expenditure, so the base is sane. A is then set so the model reproduces the observed payment count between $10 and $100, the best-measured stretch of the curve.
The fit is not tuned below $10, so its behaviour there is a result rather than an input. In the $25–$500 range the fitted curve lands within 25% of the observed counts. At $5–$10 it predicts 3× the observed volume, at $2–$5 about 10×, and below $1 several hundred times. That widening gap, appearing exactly where the card rails stop working, is the finding.
A payment of size v is treated as viable when (F + p·v)/v ≤ θ, so the floor sits at v* = F/(θ − p). At the default θ = 20%, a 2.9% + $0.30 card has a floor of $1.75 and a $0.0001 fixed fee has a floor of five hundredths of a cent.
The sweeping chart in the opening is built differently from everything below it. It works bottom-up from each rail's own published annual count and average payment size — 253.8 billion payments a year across all nine rails, including business ACH, business cards and checks. The model in this section works from consumer payments only, 151 billion a year. Same method, wider base, so the opening chart's figure for the sub-$1.75 band comes to about $1.0 trillion against the $725 billion here. Read the pair as a bracket rather than a single answer: roughly $0.7–1.0 trillion depending on whether you count business and government payments alongside consumer ones.
The exponent is doing the work. Between α = 0.44 and α = 0.74 — both defensible from the Fed's own anchors — the 1¢–$1 estimate moves from $280 billion to $1.2 trillion. Use the slider; treat the headline as an order of magnitude.
Fees are not the only floor. Attention, trust, latency, chargeback risk, tax and accounting overhead, and the sheer bother of a checkout all bind too. Removing the fee is necessary, not sufficient.
Bundling is often genuinely better. A power law extended downward assumes latent demand for unbundled things. Some of it isn't there: people really do prefer one subscription to four thousand decisions.
Much of the value is already collected. As the cross-check section argues, advertising and app stores are capturing a large share of this band today. A cheaper rail redistributes and improves that capture more than it conjures new value.
The consumer base excludes business and machine payments. This makes the estimate conservative in one direction — agent-to-agent volume is not in the calibration at all — and it means the transaction counts should not be read as human behaviour.
Sub-$1 data is thin because sub-$1 payments barely happen. The share of consumer payments below $1 is the one input taken from the shape of the Fed's histogram rather than a published figure. It sets the depth of the observed cliff, not the size of the counterfactual, but it is the softest number on the page.