Coinbase now earns nearly half its net revenue from a line called "subscription and services," and the name suggests steady fees that keep coming when trading dries up. We took the line apart quarter by quarter. Most of it is two bets: one on the price of crypto, which it shares with trading, and one on short-term interest rates, which it does not. The arithmetic of each, and four dated tests that will tell us whether we are right.
Coinbase's "subscription and services" line is less diversified than its name, and less diversified than its six revenue lines suggest. When we count the common movements behind those lines over the past twelve quarters, instead of counting the lines, we find about two. The first tracks the price of crypto. It is the common movement of consumer trading, institutional trading, staking rewards and the line called "other subscription and services," and it accounts for 65% of the variation in the growth of all six revenue lines, against about 46% that chance alone would usually produce in a record this short. The second tracks the short-term interest rate, and it carries one line almost by itself: stablecoin revenue, the income Coinbase earns on the reserves behind the USDC dollar token.
So the line is not a buffer independent of trading, as the company's earnings deck calls it, and it is not merely trading in disguise either. It is a bet on crypto prices that it shares with trading, plus a bet on interest rates that trading does not carry. In the second quarter of 2026 the share rose to 48% mostly because trading fell faster: subscription and services fell 12% from a year earlier while transaction revenue fell 22%. Over the twelve quarters we measured, the two moved in the same direction nine times.
The rate bet has a simple formula. Stablecoin revenue is Coinbase's share of the reserve income, times the interest rate, times the amount of USDC in circulation. At today's figures each percentage point of the policy rate is worth about $320 million a year to Coinbase. If rates went back to the zero-to-a-quarter-point range of 2020 and 2021, where the effective rate averaged 0.08%, the stablecoin line would earn about $26 million a year instead of about $1.2 billion, and the whole subscription and services line would lose roughly half its size. Four tests on Coinbase and on Circle, the issuer of USDC, will be scored on August 31, 2027.
This inquiry began with a hypothesis I sent to Vista's Desk on October 7, 2026, drafted with the help of another AI assistant. It started from Coinbase's own quarterly report for the three months ended June 30, 2026, the Form 10-Q filed on July 30. My hypothesis was that a higher share of subscription and services revenue does not prove stable, recurring economics, or insulation from the crypto cycle. The filing showed the line at $555 million, 48% of net revenue, down from $632 million a year earlier, and it attributed most of the decline in stablecoin revenue to lower interest rates.
I asked the Desk to challenge it, and it did, in three useful ways. It brought in the company's own phrase from the earnings presentation released the same day, which describes subscription and services as "a durable buffer independent of trading volumes." It pointed out that "does not prove" is a statement no later date can falsify, and that the right instrument for the question was one investors use every day on portfolios and almost never on a single company's revenue: principal components, which count the hidden factors that move a set of series together. This inquiry supplies the last two: a test that can fail, and the factor count, run on every quarter of Coinbase's disaggregated revenue since stablecoin revenue was first reported on its own.
The argument itself has been made before, and the people who made it deserve their credit. Stratrix wrote on September 19, 2026, in "Coinbase Swapped Trading-Fee Dependence for USDC Yield, Rate-Sensitive and One Bill From Vanishing," that the largest new leg of the business is the yield on the cash behind a stablecoin, paid under a deal with one company. Two days later the Trefis Team asked "Is Coinbase's Steadier Side Really A Cushion?" and answered that "the steady line answers to crypto prices as well." Coinbase itself says that over the past year it has captured about half of all USDC economics, and analysts at Bernstein and the Bank for International Settlements have measured that share too. The method has older roots. Kevin Stiroh asked in 2004 whether the growth of fee income had diversified American banks away from interest income; his answer was that noninterest income "is quite volatile and increasingly correlated with net interest income," so that more revenue lines bought less safety than the count implied. Fred White applied portfolio theory to the mix of a state's taxes in 1983, and Bruce Kingma to the revenue sources of nonprofit organizations in 1993. We found no published application of principal components to one company's revenue lines to measure how diversified it is. That factor count, the record of signs, Coinbase's share of the reserve income measured in every quarter since 2023, and four dated tests are what is new here.
Walter Bagehot, the editor of The Economist, published Lombard Street in 1873, and in it he told the story of the first English bankers. They were goldsmiths who held other people's coin and, as Bagehot put it, "carried on upon a trifling scale what we should now call banking." The money in their vaults sat still, and money that sits still can be put to work. The goldsmiths, Bagehot wrote, "used to deposit their reserve of treasure in the 'Exchequer,' with the sanction and under the care of the Government." In plain terms they lent it to the Crown, the safest borrower in England, until Charles II "shut up the 'Exchequer,' would pay no one, and so the 'goldsmiths' were ruined." Bagehot called it a "monstrous robbery." Notice what failed. The goldsmiths' business was sound: hold other people's money, lend the idle part to someone safe, keep the difference. What failed was the borrower. The safest name in the kingdom stopped paying. The business model was sound. The borrower was not.
The model has outlived every borrower since. A customer leaves money somewhere for convenience, the holder invests it in something safe, and the holder keeps the difference between what the investment earns and what the customer is paid, which for most of history has been nothing. In the Gospel of Matthew the master in the parable of the talents scolds the servant who buried his coin: "Then you should have deposited my money with the bankers, and on my return I would have received it back with interest." The parable assumes what every banker since has known. Idle money earns something, and someone keeps it.
In the United States the law made the model explicit for nearly eight decades. Section 11 of the Banking Act of 1933 forbade member banks to "pay any interest on any deposit which is payable on demand," and the Federal Reserve wrote that ban into Regulation Q on August 29, 1933. It stood until July 21, 2011, when the Dodd-Frank Act repealed it. For seventy-eight years, banks were forbidden by statute to pay interest on demand deposits. After 1980 households could earn interest on checkable NOW accounts, but a business's checking balance stayed interest free by law until 2011.
The ban worked only while market rates were low enough that customers did not mind. When short-term rates rose well above the ceiling in the 1970s, the money market fund was born to pay depositors what banks could not. The first fund began selling shares to the public in 1972, and taxable fund assets were about $4 billion at the end of 1977. By November 1982 they were $235 billion, almost sixty times as much in under five years. Idle money, it turned out, was idle only as long as the gap between what it could earn and what it was paid stayed small. Bagehot had recorded the saying a century earlier: "John Bull can stand many things, but he cannot stand two per cent." It was already old when he wrote it down.
Coinbase's stablecoin revenue is this trade in a new wrapper. USDC is a token that its issuer, Circle, promises to redeem for one dollar. Behind every token Circle holds a dollar of reserves, mostly short Treasury bills and repurchase agreements, and the reserves earn interest. The issuer pays its holder nothing. The interest is divided between Circle and Coinbase under an agreement the two signed on August 18, 2023, and in 2025 a federal law, the GENIUS Act, made the arrangement's foundation explicit: no permitted issuer "shall pay the holder of any payment stablecoin any form of interest or yield (whether in cash, tokens, or other consideration) solely in connection with the holding, use, or retention of such payment stablecoin." The law names issuers, not distributors, which is why Coinbase still pays rewards to holders on its own platform, and how far that door stays open is a live question, which Congress has already tried and so far failed to settle (more on that below). Regulation Q kept the float's income with the bank. The GENIUS Act keeps it with the issuer and the people the issuer shares it with. History, as it does, rhymes.
On October 31, 2008, a message reached a mailing list of cryptographers under the name Satoshi Nakamoto. "I've been working on a new electronic cash system that's fully peer-to-peer, with no trusted third party," it began, and it pointed to a nine-page paper, Bitcoin: A Peer-to-Peer Electronic Cash System. Nine weeks later, on January 3, 2009, the first block of the new ledger was written. Inside it, its maker left a line from that day's edition of The Times: "The Times 03/Jan/2009 Chancellor on brink of second bailout for banks." A new money began its life with a note about the old money's rescues.
Three years later, in May 2012, Brian Armstrong, then a software engineer at Airbnb, started Coinbase on what its listing prospectus would call "the radical idea that anyone, anywhere, should be able to easily and securely send and receive Bitcoin." With Fred Ehrsam as co-founder, it went through Y Combinator, the startup accelerator, that summer. Nine years later, on April 14, 2021, its shares began trading on Nasdaq under the ticker COIN. It was a direct listing. The company sold no shares; its owners simply became free to sell theirs.
Trading coins made a plain problem. A trader who sold bitcoin wanted dollars, and dollars lived in banks, which keep business hours that a blockchain does not. The market's answer was a token that stood for a dollar and moved like a coin. Tether came first, in 2014. The Treasury's 2021 report on stablecoins said they were "primarily used to facilitate trading of other digital assets." USDC arrived in September 2018, launched through Centre, a consortium Circle and Coinbase founded together. Each token was to be backed by reserves held by the issuer and redeemable for one dollar. The reserves would earn interest, and the holder would not.
Not every digital dollar had dollars behind it. TerraUSD was what its makers called an algorithmic stablecoin. It held its peg, they said, because it could always be exchanged for a sister token, LUNA, and they advertised as much as 20 percent interest on it. In May 2022 it came loose from the dollar, and in the words of the Securities and Exchange Commission its price and its sister tokens' "plummeted to close to zero." A reserve can be counted. An algorithm has to be believed.
USDC's own test came from the other side of the ledger, on a weekend. A reserve kept as a bank deposit is a loan to the bank. On Friday, March 10, 2023, California's regulators closed Silicon Valley Bank. That night Circle said that $3.3 billion of the roughly $40 billion behind USDC was still inside; wires sent the day before to move it out had not been processed. On Saturday the banks were shut. The blockchain, as always, was open. Redemptions had to wait for the banks, so holders who wanted out sold on the open market, and USDC traded as low as about 88 cents. Coinbase's filings later spoke of "a temporary disruption to USDC services for several days." On Sunday the Treasury, the Federal Reserve and the Federal Deposit Insurance Corporation announced that depositors would have "access to all of their money starting Monday, March 13." The token climbed back toward its dollar, and all of Circle's money at the bank was later moved to other banks.
Bagehot would have known that weekend at once. His goldsmiths deposited their reserve in the Exchequer, with the safest borrower in England, until in 1672 the King stopped paying. Circle deposited part of its reserve with a bank, and the bank failed. The reserve was real. Its address was the trouble. The business model was sound. The borrower was not. Three and a half centuries changed only the ending: the goldsmiths were ruined, and the depositors of 2023 were promised every dollar within two days.
Five months later the partners rearranged the house. On August 18, 2023, Coinbase handed its half of Centre to Circle in exchange for 3.5% of Circle's fully diluted equity, Centre was wound up, and Circle took sole governance of USDC. The same day the two signed the collaboration agreement whose arithmetic this inquiry takes apart.
Circle's shares began trading on the New York Stock Exchange on June 5, 2025, under the ticker CRCL, after an offering at $31 a share. Six weeks later, on July 18, the GENIUS Act became law, with its rule that an issuer may not pay holders interest. On August 5, 2026, Circle's chief executive, Jeremy Allaire, told investors that "our agreement with Coinbase has renewed on its existing terms," which under the contract's three-year renewals carries it to August 2029.
Congress has not finished. On May 14, 2026, the Senate Banking Committee voted 15 to 9 to advance a version of the CLARITY Act whose Section 404 would forbid any digital asset service provider, an exchange like Coinbase among them, to pay interest or yield "solely in connection with the holding" of a payment stablecoin. Rewards for bona fide activity would stay legal if they did not work like a bank deposit's interest, even when figured on a balance and how long it was held. On September 15 the bill failed a procedural vote in the full Senate, 49 to 50. The rewards Coinbase pays to keep USDC on its platform live, for now, in the room that vote left open.
Coinbase reports its revenue in two families. Transaction revenue is what it earns when customers trade: fees from consumers, fees from institutions and a smaller line of other transaction revenue. Subscription and services is everything else that is not interest on its own corporate cash, and in the second quarter of 2026 it held four tenants.
Stablecoin revenue, $292 million, Coinbase's share of the income on USDC's reserves. Blockchain rewards, $83 million, the rewards Coinbase earns for staking customers' tokens on networks such as Ethereum and Solana, which it shares with those customers. Interest and finance fee income, $66 million, interest on customers' cash and on the loans Coinbase makes. And other subscription and services, $114 million, which includes Coinbase One, the paid membership that is the only part of the line a reader would recognize as a subscription.
Two of the four tenants, stablecoin revenue and blockchain rewards, account for 68% of the line. Neither is a fee a customer agrees to pay on fixed terms. The 10-Q says plainly that stablecoin revenue, like Coinbase's interest income, is "not accounted for as revenue from contracts with customers" under the accounting rule for customer revenue; it is income under the agreement with Circle, which the 10-K says "is not a customer of the Company." Blockchain rewards rise and fall with the price of the tokens staked and with the reward rates of the networks themselves. The filing put the decline in blockchain rewards in the quarter at $56.5 million from lower token prices, "driven primarily by Solana," and $16.1 million from lower reward rates.
Each tenant also has a cost beside it that the revenue line does not show. Coinbase pays its customers "USDC rewards" to keep their tokens on its platform, recorded in sales and marketing: $119 million in the quarter, up from $103 million a year earlier, while stablecoin revenue fell. Stablecoin revenue less those rewards fell 16%, three times the 5% decline of the revenue line alone. Blockchain rewards less the fees Coinbase passes on to stakers fell 45%. These are proxies we built, revenue less one disclosed cost next to it, and they are not margins: the filings do not allocate staff, technology or overhead to individual lines. But they say which way the economics moved, and it was down more steeply than the revenue line shows.
The name, in short, describes how the revenue is billed less than what moves it. Plato let Socrates make the point in the Cratylus, after a long argument with a man who believed that whoever knows the names of things knows the things: "the knowledge of things is not to be derived from names. No; they must be studied and investigated in themselves." That is the method of what follows.
Begin with the largest tenant, because its machine can be written down exactly. The 10-K describes how Coinbase is paid. Each day the reserves behind USDC earn income; from it the managers' fees and certain expenses are taken, leaving what the agreement calls the Payment Base. Circle keeps a portion as issuer. Coinbase and Circle each earn the income on the USDC held on their own platforms. Other approved partners earn what they have agreed. And Coinbase receives "50% of the remaining Payment Base," the income on all the USDC held anywhere else. The fifty percent was redacted in the copy of the agreement filed in 2025; the copy filed with the 10-K for 2025 shows it.
All of that collapses into one line:
Stablecoin revenue = θ x r x S
S is the average amount of USDC in circulation in the quarter, the float. r is the short-term interest rate the reserves earn; we use the federal funds rate as the yardstick, because the reserves are invested in the shortest Treasury bills and repurchase agreements, whose yields track it closely. And θ, the Greek letter theta, is Coinbase's share of the income: the fraction of r x S that ends up in its revenue. The contract fixes the formula for theta, not its value: theta moves with where the USDC sits, because Coinbase earns the income on tokens held on its own platform and half of the residual on tokens held elsewhere, with the portion Circle retains as issuer, which the agreement sets on a schedule tied to circulation, and with what other partners take.
We can measure theta every quarter from the filings and the decks: annual stablecoin revenue, divided by the average federal funds rate times the average USDC in circulation. Here is the second quarter of 2026, walked through. Stablecoin revenue was $292.1 million in the quarter, or $1,168.6 million a year at that pace. The federal funds rate averaged 3.63%. USDC in circulation averaged $77 billion, according to the earnings deck. The reserve income on the whole float was therefore 3.63% x $77 billion = $2,795 million a year, and Coinbase's share was $1,168.6 million / $2,795 million = 41.8%.
| Quarter | Stablecoin revenue, $ millions | Fed funds, % | USDC in circulation, $ billions | Theta, % |
|---|---|---|---|---|
| Q1 2023 | $198.9 | 4.51 | 41.3 | 42.7 |
| Q3 2023 | $172.4 | 5.26 | 26.0 | 50.4 |
| Q2 2024 | $240.4 | 5.33 | 33.0 | 54.7 |
| Q2 2025 | $332.5 | 4.33 | 61.0 | 50.4 |
| Q4 2025 | $364.1 | 3.90 | 76.0 | 49.2 |
| Q2 2026 | $292.1 | 3.63 | 77.0 | 41.8 |
Over the fourteen quarters from early 2023, theta stayed between about 40% and 55%. Coinbase itself says it has captured about half of all USDC economics over the past year; the series measures that share in every quarter. It has stayed within a band, and what moved stablecoin revenue most was the rate and the float. The band has one clear step: theta rose from 40.5% to 50.4% in the third quarter of 2023, the quarter the new agreement with Circle was signed. When the float shrank in 2023, from $41 billion to $25 billion, revenue fell even as rates rose. When the float tripled between the end of 2023 and the end of 2025, revenue more than doubled even as rates fell. And in 2026, with the float flat at about $77 billion, revenue fell with the rate.
Two cautions about theta. Since the first quarter of 2026 Coinbase reports the income on its own corporate USDC in a different line, so the last two quarters understate theta on the old basis; adding back the $18 million and $28 million the deck reports for those quarters gives 47.4% and 45.8%. Even on the old basis theta drifted down from about 54% in 2024 to about 46% in mid-2026, and the deck shows where the drift sits. On the USDC held in its own products, Coinbase earned about 96% of what the federal funds rate would pay on those balances in the second quarter of 2026, against 95% a year earlier. On the USDC held everywhere else, its take fell from about 37% to about 28%. The documents do not say how much of that went to the partners admitted under the Stablecoin Ecosystem Agreement of November 2024, how much to Circle's retention schedule, and how much to the gap between the federal funds rate and what the reserves actually earned. It is one of the things a person who has negotiated such terms would know. The question is live: on September 17, 2026, Circle reported in an 8-K a five-year agreement under which it pays Binance a monthly incentive fee on USDC held through Circle's wallet service, together with a $100 million share purchase by Binance, replacing the agreements of November 2024 and August 2025. The filing does not say whether that fee comes out of the pool Coinbase shares in.
Now the worked examples. At the second quarter's float and theta, each percentage point of the rate is worth:
41.8% x 1% x $77 billion = $322 million a year
The company's own figure points the same way, at a higher share. Its 10-K for 2025 says that a change of 150 basis points (1.5 percentage points) in the average rate on the reserves would have changed 2025 stablecoin revenue by $540.3 million, which is $360 million a point on a 2025 float that averaged about $65 billion. That implies a share of about 55%, at the top of the range we measure for 2024 (51% to 55%) and above 2025's (49% to 51%); the 10-K's sensitivity and our quarterly measure are not built the same way. At a rate of 2%, the line would earn $644 million a year. At 1%, $322 million. At a quarter of a point, the ceiling the Federal Reserve held from March 2020 to March 2022, $80 million; at the 0.08% the effective rate actually averaged over those two years, about $26 million. Since September 17, 2026, after the Federal Reserve raised its target by a quarter point, the effective rate has been about 3.88%, worth $1,249 million a year at the second quarter's float and share.
Interest rates go into that formula once, but they go into Coinbase's costs as well. The USDC rewards Coinbase pays customers reprice with rates, too: the 10-Q attributes a $45.4 million increase in rewards to larger balances, "offset in part by a reduction in the rewards rate." When the rate falls, Coinbase pays its customers a little less to keep their USDC on its platform. The lower cost cushions the profit. It does nothing for the revenue line.
The earnings deck makes much of the share: subscription and services "represented 48% of net revenue." A share is a ratio, and a ratio can rise because its numerator rose or because its denominator fell. In the second quarter of 2026 both parts fell, and the denominator fell faster. Subscription and services fell 12.2% from a year earlier; transaction revenue fell 21.6%; net revenue fell 17.3%. On the restated basis the share rose from 45.3% to 48.1%.
Here is the arithmetic that shows what the share measures. Had transaction revenue merely held at its year-earlier $764 million, with subscription and services where it actually landed, the share would have been 42.1%, lower than a year before. The share went up because trading went down. It measured relative resilience in a bad quarter for trading, which is real, and nothing about insulation.
The resilience was itself partly a one-off. The deck credits the other subscription and services line with the "achievement of performance earn outs," and its outlook for the third quarter lists the "Roll-off of Q2 performance earn outs" as a negative. Interest and finance fee income plus other subscription and services rose 0.5% from a year earlier, the steadiest part of the line, and some of that was the earn-outs. One quarter of stability, partly one-off, is one observation; it is not a pattern.
The deck also gives a third-quarter outlook for the line of $500 million to $580 million, below the second quarter's $555 million at the midpoint and 19% to 30% below the third quarter of 2025, restated. In May the company had guided to $565 million to $645 million for the second quarter, and the line came in at $555 million, below the range. A buffer that misses its own forecast in a falling market is behaving like the market.
The share is a static picture. The question I asked is dynamic: when crypto turns, does this line turn with it? To answer it we need the lines' history and a way to count the forces that move them together.
Principal components is that way. It was described by Karl Pearson in 1901 as the line of closest fit through a cloud of points, named and formalized by Harold Hotelling in 1933, and it has been a staple of portfolio work since Harry Markowitz showed in 1952 that diversification depends on how holdings move together, not on how many there are. The idea is simple to state. Take several series. Find the single direction, a weighted mix of all of them, along which they vary the most together. That is the first component. Then the next direction, unrelated to the first, that captures the most of what is left. And so on. If one component explains most of the variation, the series are mostly one force wearing several names.
In symbols, write the correlations among the lines' growth rates as a matrix, Σ, and solve:
Σ v = λ v, and the first component's share = λ1 / (λ1 + λ2 + ... + λ6)
Each λ, lambda, is the variance one component explains; each v holds the weights, called loadings, that say how much each line takes part in it. With six lines measured on a correlation scale, the lambdas sum to six. If the lines were independent, each lambda would be close to one and the first component's share close to a sixth.
We built the series from every 10-Q and 10-K since 2022: six lines, consumer transactions, institutional transactions, stablecoin revenue, blockchain rewards, interest and finance fee income, and other subscription and services, each as its growth over the same quarter a year before. Year-over-year growth removes seasonality, and comparing each quarter with its own year-earlier figure on the same basis keeps the reclassifications out of the arithmetic: the 2026 quarters are compared with 2025 restated, and every earlier pair with the presentation of its own filing. That gives twelve observations, from the third quarter of 2023 to the second quarter of 2026. We used the correlation matrix, not the covariance matrix, so that the biggest line cannot dominate merely by being biggest.
The result: the first component explains 65.3% of the variation in the six lines' growth. The second explains 20.9%. Together they explain 86%; the remaining four explain the rest between them.
Twelve quarters is a short history, and growth rates measured over overlapping years are strongly linked from one quarter to the next, which makes spurious correlation common, so we asked how often chance alone would give a first component this large. We built six independent random series two thousand times, each as year-over-year changes in overlapping windows exactly as ours are, and computed the first component's share each time. The median was 46.4%, and the 95th percentile 60.0%. Coinbase's 65.3% is above that: chance gave a share this large in about one draw in seventy. It is a clear result, not an overwhelming one, because twelve overlapping quarters are a short record. Leaving out each quarter in turn moves it only between 64% and 73%.
There is one more way to say it. A standard measure of how many independent series a set behaves like is the square of the sum of the lambdas divided by the sum of their squares; it is six for six independent lines and one for six copies of the same line. For Coinbase's six revenue lines it is 2.1. Attilio Meucci made the same idea the "effective number of bets" of a portfolio. Six names, about two forces.
Which forces? Here the loadings speak, with the caution the method demands. The first component loads almost equally on consumer transactions, institutional transactions, blockchain rewards and other subscription and services, at weights near one half each, with a smaller weight on interest and finance fee income and nearly none on stablecoin revenue. Its quarterly scores move with the growth of the Bitcoin price at a correlation of 0.88, and with Ether's at 0.88. The second component is stablecoin revenue almost alone, with a loading of 0.86, and its scores move with the change in the federal funds rate at a correlation of 0.58.
A loading describes co-movement; it does not name a cause. The words "crypto factor" and "rate factor" are labels we attach after comparing the components with outside series, and with twelve overlapping observations the comparison is a description, not an estimate. But the labels fit what the filings themselves say moves each line. Consumer trading moves with prices, which bring customers in. Staking rewards are paid in tokens, so their dollar value is the token price. The "other" line holds Coinbase One, the earn-outs and the services the deck credits, and its growth has tracked consumer trading at a correlation of 0.96. And stablecoin revenue, as the formula shows, is the rate times a float.
So the line called subscription and services is two bets in one name. Blockchain rewards and the other line, a little over a third of its revenue, sit on the same crypto factor as trading. Stablecoin revenue, a little over half, is a bet on r, the short rate, which trading does not carry at all, written on a float that grows when crypto activity grows. Interest and finance fees, the remaining eighth, take part in both. That second bet is real diversification, and it is the part of the hypothesis the data refine: the line is not insulated from crypto, but it does carry one force that trading does not.
The record of signs makes the same point in plainer terms. In nine of the twelve quarters, subscription and services and transaction revenue moved in the same direction from a year earlier. The three exceptions are instructive. In the third quarter of 2023 trading fell 21% while subscription and services rose 59%, because the federal funds rate had risen above 5% and a new agreement with Circle, signed that August, raised Coinbase's share of the reserve income. In the second quarter of 2025, a near miss, trading slipped 2% while the line grew 10%. In the fourth quarter of 2025 trading fell 37% and the line still grew 14%, on a float that had doubled in a year. In all three, what held the line up was the rate, the float or the terms of the Circle agreement, never a customer's subscription.
The correlation of the two growth rates over the twelve quarters is 0.74.
A reader who values subscription and services as recurring revenue, at the multiple the market gives to subscription businesses, is assuming two things the label does not say: that the rate stays where it is, and that crypto activity keeps the float and the staking balances growing. We can price the first assumption with the formula.
At the second quarter's run rate, subscription and services was $2,221 million a year and stablecoin revenue $1,169 million of it. Hold the float at $77 billion and theta at 41.8%, and move the rate:
| Federal funds rate | Stablecoin revenue a year | Change from the second quarter's pace |
|---|---|---|
| 3.63% (Q2 2026) | $1,169 million | none |
| 2.00% | $644 million | $525 million less |
| 1.00% | $322 million | $847 million less |
| 0.25% (the 2020 to 2022 ceiling) | $80 million | $1,088 million less |
| 0.08% (the 2020 to 2022 average) | $26 million | $1,143 million less |
At the effective rate of 2020 and 2021, the line would lose about $1.14 billion a year, 51% of its current annual size, without a single customer leaving. Lower USDC rewards would cushion the profit, but not the revenue line.
This is not a remote case. It is where the line came from. In 2021, with rates near zero, subscription and services was 7% of Coinbase's net revenue and stablecoin revenue was less than $10 million for the year. In 2023, when the rate passed 5%, subscription and services was 48%, and stablecoin revenue was about seventy times its 2021 level. The 10-K for 2023 said the increase was "attributable to higher average earned interest rates on USDC reserves, which rose 325 basis points." The 48% of 2023 came from a rising rate; the 48% of 2026 came from falling trading. The same share, for different reasons, which is why a share alone tells so little.
Rates are not the only hinge. Circle's own 10-K adds a sentence that every holder of either company's shares should read twice: "Any relationship between interest rates and USDC in circulation is complex, highly uncertain, and unproven." Lower rates could, in principle, raise demand for a token that pays nothing, because the cost of holding it falls; they could also lower it, if crypto activity, which drives much of the float, slows. The float S is the term the formula cannot forecast, and it is the term most exposed to the crypto factor.
And theta, the term whose rules the contract sets, rests on an agreement whose initial three-year term ran to August 18, 2026. The deck says the renewal conditions were "already met and the partnership will renew on the same terms." That is the company's statement, and the agreement filed with the 10-K describes automatic three-year renewals when its thresholds are met. On August 5, 2026, Circle's chief executive, Jeremy Allaire, told investors that "our agreement with Coinbase has renewed on its existing terms," which carries it to August 2029. The contract protects the share. It does not protect the rate or the float.
Every business that has earned income on other people's idle balances has met the same two dangers, one at each end of the rate cycle. At zero, the income disappears. At a high rate, the balances leave.
The first danger has a precise history. When short rates sat near zero after 2008, money market funds could not cover their own fees without paying their investors a negative yield, so their managers waived the fees. The Investment Company Institute estimates that "from 2009 to 2015, advisers waived an estimated $36 billion in money market fund expenses." In 2021, when rates were again near zero, an average of 97% of money fund share classes waived expenses, and waivers came to an estimated $8.4 billion in that year alone. Charles Schwab's 10-K for 2021 shows what that did to one large manager: its money market funds earned $457 million in fees before waivers and waived $326 million of them, 71%. Income on idle money is a business only when money earns something.
The second danger arrived on schedule when rates rose. Schwab's 10-K for 2023 described its clients "allocating cash out of sweep products into higher-yielding investment solutions." Its bank deposits fell from $444 billion at the end of 2021 to $290 billion at the end of 2023, a decline of 35% even after it added brokered deposits to fund itself, and its net interest revenue fell 12% in 2023, a year of the highest short rates in a generation.
Coinbase meets both dangers in a form of its own. At zero rates, the formula shows how little the stablecoin line would earn. At high rates the pressure to share appears as USDC rewards: the more the float earns, the more customers must be paid to keep their tokens on the platform, where Coinbase earns the income on them directly instead of half of a residual. The rewards Coinbase paid rose from $35 million in 2023 to $441 million in 2025. It is Schwab's cash-sorting problem, solved in advance by paying customers to stay.
This reminds me of the oldest lesson in that history, the one Bagehot drew from the goldsmiths and that the money market funds relearned in 1982 and again in 2021. The income on idle money is real, and it can be very large, but it belongs to whoever controls three things at once: the money, the rate, and the customer's willingness to leave it where it is. Coinbase controls the third in part, through its rewards. It controls none of the first two. The Federal Reserve sets r, and the crypto market sets S.
The economics of deposits is one of the better-studied subjects in finance, and it says the same thing from the other side. Itamar Drechsler, Alexi Savov and Philipp Schnabl showed in 2017 that when the Federal Reserve raises rates, banks widen the spread between what they earn and what they pay on deposits, and deposits flow out; in 2021 they showed that a bank's deposit franchise lets it pay rates that are low and insensitive to the market, which offsets the interest rate risk of its long-term lending. The contrast with a stablecoin is the instructive part. A bank holds long, fixed-rate loans that keep earning when rates fall. USDC's reserves are invested in the shortest bills and repurchase agreements, so their income reprices within weeks, and nothing on the asset side locks it in. A bank's franchise is hedged. A stablecoin's is not.
On the stablecoin side, Gary Gorton and Jeffery Zhang argued in their 2023 paper "Taming Wildcat Stablecoins" that stablecoins recreate the private banknotes of the nineteenth century; Gorton and four colleagues found, in a paper first circulated in 2022 and published in 2026, that stablecoin owners are compensated for run risk by lending their coins to crypto speculators, one mechanism by which the float and crypto activity may move together, and a reason to expect the two factors we found to be less separate in a crash than in ordinary times.
Yes, and it began with the chips. CoreWeave was formed in September 2017 as The Atlantic Crypto Corporation and mined cryptocurrencies, chiefly Ether. Ethereum stopped paying miners when it moved to proof of stake on September 15, 2022; CoreWeave closed its mining business on September 30 and turned to renting AI computing.
Bitcoin miners followed. On November 3, 2025, IREN, which then earned 97% of its revenue from bitcoin, signed a $9.7 billion, five-year contract to rent Nvidia chips to Microsoft, and Cipher Mining a $5.5 billion, 15-year lease of space and power to Amazon Web Services.
In May 2025 Coinbase launched x402, a protocol that lets software, AI agents included, pay for web services in stablecoins. An agent paying in USDC must hold some first, adding to S; I have found no figure for how much.
The last link, the rate, is conditional. Inquiry No. 7 used Ramsey's rule, r = ρ + θ x g (another theta), to argue that faster AI-driven growth would raise long real rates, a link it called disputed. Coinbase earns on the short rate. The bridge is the neutral rate, the policy rate for an economy in balance. Governor Michael Barr said in February he had raised his estimate of it "because of higher productivity," and that the AI boom is "unlikely to be a reason for lowering policy rates." Kevin Warsh, the Fed's chair since May, has written that AI "will be a significant disinflationary force." If Barr is right, r stays higher, worth about $320 million a year to Coinbase per point. If Warsh is right, the formula runs the other way.
I also measured whether the two parties move together, on weekly prices from 2016 to October 2, 2026. Before 2020 they did not: Bitcoin's weekly correlation with the Nasdaq-100 was 0.04. Since then it has been 0.26 to 0.40, and highest in the last twelve months. But take the Nasdaq-100's moves out of both, and Bitcoin's correlation with Nvidia falls to 0.01 since ChatGPT's release and 0.08 in the last year. The two parties share a mood, the market's appetite for risk, and nothing more. In the 19 weeks since ChatGPT's release when the Nasdaq-100 fell 3% or more, Bitcoin fell in 14, by a median 5.5%; in the 33 weeks it rose 3% or more, Bitcoin rose a median 2.3%, against 9.1% for Nvidia. Crypto shares the AI party's hangovers more than its highs. The miners are the exception, because they changed parties. Before 2024 IREN's weekly return moved one for one with Bitcoin and hardly at all with chip stocks; since 2024 it moves 1.30 times with the chip stocks and half as much with Bitcoin. CoreWeave, the former Ether miner, now moves with the chips and hardly with Bitcoin. Coinbase went the other way, toward Bitcoin.
Four hypotheses follow from the machine and the factor count. None is a recommendation to buy, sell or hold anything; each says what a later filing will show if the reading here is right, and what would prove it wrong.
The line moves with trading. Coinbase's subscription and services revenue will move in the same direction as its transaction revenue, year over year, in at least three of the four quarters from the third quarter of 2026 to the second quarter of 2027, on the restated basis. The evidence: nine of twelve quarters so far and a correlation of 0.74. The company's own third-quarter outlook of $500 million to $580 million for the line, 19% to 30% below a year earlier, already leans one way for the first of the four quarters, so the later three carry the test. Falsified if, in two or more of the four quarters, the line moves against transaction revenue, or holds within 5% of its year-earlier level while transaction revenue moves by more than 10%.
The factors stay few. Adding those four quarters to the twelve, the first principal component of the six lines' growth will still explain more than 54.7% of the variance, the level that six independent series, built as overlapping year-over-year changes over sixteen quarters, exceed only one time in twenty. The evidence: 65.3% on twelve quarters, between 64% and 73% when any one quarter is dropped. Falsified if the first component's share on the sixteen quarters is 54.7% or less.
The share of the reserve income holds, and the rate moves the revenue. In each of the four quarters, Coinbase's stablecoin revenue, annualized and divided by the average federal funds rate times the average USDC in circulation, will give a theta between 38% and 50% on the reported basis. The evidence: 40% to 55% in each of fourteen quarters on the old basis, 44.8% and 41.8% in the two quarters since the reclassification. Falsified if any quarter falls outside the band; a theta outside it would mean the contract, the mix of where USDC sits, or the partners' shares had moved enough to change the machine. Circle's new five-year agreement with Binance, signed in September, is the first thing to watch: if its fee is paid from the shared pool, the lower edge of the band is where it would show.
The issuer is a rate bet, too. Circle Internet Group's reserve income will be at least 90% of its total revenue and reserve income, the top line of its statement of operations, in its 10-K for 2026. It was 96.0% in 2025, 99.1% in 2024 and 98.6% in 2023, and Circle's own filing says "a decrease in interest rates reduces reserve return rates." The question is whether Circle builds enough fee revenue to dilute the rate bet at the issuer. Falsified if reserve income is below 90%.
What follows for each kind of holder is plain enough. Anyone who holds Coinbase for the subscription line is holding a rate position they may not know they have, and should know the size: about $320 million of yearly revenue per point. Anyone who holds it for crypto exposure has more of it than the trading line suggests, because a third of the subscription line rides the same factor directly and the float behind another half grows with it. And anyone who holds Circle holds the same rate position at greater concentration. The pattern travels: any company that reports income on customers' idle balances under a label that sounds like fees, at a broker, an exchange or a payments company, can be taken apart the same way, with the same formula and the same factor count.
Diversification is the oldest advice in investing and the easiest to fake. Six revenue lines look like six streams. Ask how many forces move them, and Coinbase has about two, both outside its control: the price of crypto and the short-term interest rate. The second is genuinely different from the first, which is good news for anyone worried that subscription and services is only trading with a different name. The bad news is that the second is a bet on a number the Federal Reserve sets, and that the label, the deck's language and the 48% share all hide it.
The asymmetry is plain in the formula. If rates stay near their present 3.9% and the float grows, the line grows; the deck's outlook names the float as its first positive driver. If rates fall toward where they were in 2021, the line loses about half its size whatever the float does. And if crypto falls, the float and the staking balances fall with it, along with trading. The scenario in which subscription and services truly buffers trading is the one in which crypto falls while rates rise, as in 2023. It happened once. It may happen again, but it is a macroeconomic coincidence, not a contract.
The signposts are dated. The third-quarter report, probably in late October or early November, will show whether the line moves with trading again. The Federal Reserve's decisions over the next year set r. Circle's 10-K, in early 2027, will show the issuer's dependence on reserve income. And the quarterly decks will keep reporting USDC in circulation and on the platform, which is S, and the corporate balances that let theta be measured on both bases.
Likely, we can get some of it from primary research, which is what Vista is all about. The people who negotiated stablecoin distribution terms, priced staking programs and ran cash sweeps already know how these lines behave at the turns that twelve quarters have not yet shown.
As a matter of general past practice, how were reserve-income shares and distribution payments set, and what moved a distributor's effective share of reserve income from one year to the next?
The answer that would change the view"The distributor's share moved mostly with negotiated terms, not with rates or balances, and the terms were renegotiated often."
In general, how much of staking revenue's movement came from token prices, how much from network reward rates and how much from the number of tokens staked?
The answer that would change the view"Staked units, not prices, drove most of the revenue, and they kept growing through price declines."
As a matter of general past practice, at what gap between market rates and the rate paid did client balances start to leave, and how fast?
The answer that would change the view"Balances that earn nothing stayed put even when the gap passed four points, because convenience mattered more than yield."
In general, how did investors value fee-like revenue earned on customers' balances, and did the market distinguish it from contracted subscription revenue?
The answer that would change the view"Investors already priced balance income at a rate-sensitive multiple and never treated it as subscription revenue."
In general, which revenue lines were forecast from crypto prices, which from rates, and which from contracts, and how did the forecasts fare at the turns?
The answer that would change the view"The subscription lines were forecast from contracts and held at the turns, independent of prices and rates."
For a client engagement, Vista combines this report with that primary research.
We will score this inquiry on August 31, 2027, after Coinbase reports the second quarter of 2027 and Circle files its 10-K for 2026, and publish the result on the scorecard whether it flatters us or not. Each test sides with the reading here, that subscription and services is two bets, one on crypto and one on rates, and not a buffer independent of trading:
All four pass and the reading here gains ground; three is too early to say; two or fewer, and the line is more of a buffer than its history shows.
Bagehot's goldsmiths lent the money that sat in their vaults to the safest borrower in England and were ruined when he stopped paying. The money funds of 1982 drew $235 billion out of the banks when the rate rose, and gave billions back in waivers when it fell. Coinbase earns on the money that sits in USDC, and the borrower this time is the United States Treasury, as safe as borrowers come. The danger is not default. It is the rate, which the Treasury does not set either. A rose by any other name, as Juliet is usually quoted, would smell as sweet. A rate bet by the name of subscription is still a rate bet.
The bet. Anyone who values Coinbase's subscription and services as recurring revenue is betting that short-term rates stay near their present level and that crypto activity keeps the USDC float and the staking balances growing. The line's name does not carry that bet. Its formula does: theta times r times S, where the contract sets only the rules for theta.
The payoff. If rates hold and the float grows, the line grows with the float, and the 48% share is earned again on a larger base. If rates fall toward 2021's level, the stablecoin line alone loses about $1.1 billion a year, roughly half of the whole line's present size; if crypto falls, the staking and other lines fall with trading.
Our read. The evidence sides with the hypothesis, refined. In nine of twelve quarters the line moved with trading; one common factor, which tracks the price of crypto, accounts for 65% of the variation in six revenue lines; a second, which tracks the rate, carries the stablecoin line almost alone. The line diversifies Coinbase across two macroeconomic bets. It does not diversify it away from them. If AI keeps the neutral rate high, as Inquiry No. 7 argues it may, the rate bet pays; if crypto falls, the name will not help.
What settles it, and when. Coinbase's four quarterly reports from late 2026 to mid-2027 and Circle's 10-K for 2026; we score the four tests on August 31, 2027. The first conversations we would have are with the people who negotiated stablecoin distribution terms and ran cash sweeps, because between them they know how these balances behave at the turns that the last twelve quarters have not shown.
Four lines. In the second quarter of 2026: stablecoin revenue, $292 million, Coinbase's share of the income on the reserves behind USDC; blockchain rewards, $83 million, from staking customers' tokens; interest and finance fee income, $66 million; and other subscription and services, $114 million, which includes Coinbase One. Stablecoin revenue and blockchain rewards are 68% of the line, and neither is a fee a customer pays on fixed terms.
Stablecoin revenue is Coinbase's share of the reserve income, times the rate, times the USDC in circulation. At the second quarter of 2026's share of 41.8% and a $77 billion float, each percentage point of the federal funds rate is worth about $322 million a year. Coinbase's own 10-K for 2025 put a 150 basis point change at $540.3 million for that year, on a smaller float and a higher share. Since September 17, 2026 the effective rate has been about 3.88%.
Only in part. In nine of twelve quarters from 2023 to mid-2026 the line moved in the same direction as transaction revenue. One common factor, which tracks the price of crypto, accounts for 65% of the variation in the growth of Coinbase's six revenue lines; a second, which tracks the short-term interest rate, carries stablecoin revenue almost alone. The line diversifies Coinbase across two bets without diversifying it away from them.
Mostly because trading fell faster. In the second quarter of 2026 subscription and services fell 12% from a year earlier and transaction revenue fell 22%. Had transaction revenue held at its year-earlier level, the share would have been 42.1%, lower than a year before.
At the 0.08% the effective federal funds rate averaged in 2020 and 2021, with the float and Coinbase's share held where they were in the second quarter of 2026, the stablecoin line would earn about $26 million a year instead of about $1,169 million, a loss of about $1.14 billion, roughly half of the whole subscription and services line's present annual size. At a quarter of a point, the top of that period's range, it would earn about $80 million.
Every figure was computed in code (model.py, standard library only) from two data files built from Coinbase's filings at sec.gov and its investor relations decks: data-quarterly.csv, every revenue line from the first quarter of 2022 to the second quarter of 2026 in thousands of dollars, each row naming the filing whose presentation it uses and checked so that the lines sum to their totals; and data-drivers.csv, the quarterly average of the daily federal funds rate (FRED series DFF) and of Coinbase's daily Bitcoin and Ether prices as published by FRED (CBBTCUSD, CBETHUSD), with the average USDC in circulation from the Q2 2026 earnings deck (pages 13 and 40) and earlier letters. Fourth quarters are the fiscal year less nine months, cross-checked against the shareholder letters. Theta is annualized stablecoin revenue (four times the quarter) divided by the average federal funds rate times average USDC in circulation; the 2026 quarters are also shown with the deck's corporate-balance revenue added back. The principal components are the eigenvalues and eigenvectors of the correlation matrix of the six lines' year-over-year log growth, twelve observations from Q3 2023 to Q2 2026, each quarter compared with its year-earlier quarter on the same basis (2026 against 2025 restated; Q1 and Q2 2025 against 2024 on their original basis; custodial fees folded into other subscription and services throughout; for Q3 and Q4 2023, whose 2022 comparatives keep other transaction revenue inside consumer, consumer is compared as consumer plus other transaction revenue on both sides); the chance benchmark builds six independent random walks 2,000 times with fixed seeds and takes their overlapping four-quarter changes, twelve (and sixteen) observations, exactly as the data are built. The effective rate since September 17, 2026 (3.88%) and its 2020 to 2022 average (0.08%) are from the same FRED series; the 2025 average float of about $65 billion is from Circle's 10-K for 2025, and the on-platform and off-platform takes divide the deck's page 40 revenue by the federal funds rate times the page's balances. The effective number of lines is the square of the sum of the eigenvalues over the sum of their squares (after Meucci's effective number of bets). The drivers' correlations use the year-over-year change in the average federal funds rate and the log change in the average Bitcoin and Ether prices. The proxies for economics subtract USDC rewards (Note 17 of the 10-Q) from stablecoin revenue and blockchain rewards fees (the transaction expense table) from blockchain rewards; they are not margins. Every historical figure and quotation was checked against the primary document named in the sources.
| Input | Value | Source |
|---|---|---|
| Coinbase revenue lines, Q2 2026 and Q2 2025 restated (thousands) | Subscription and services $555,145 and $632,243; stablecoin $292,147 and $308,914; transaction $599,156 and $764,270; net revenue $1,154,301 and $1,396,513 | Coinbase Form 10-Q for the quarter ended June 30, 2026, Note 5 |
| Q2 2026 sub-lines (thousands) | Blockchain rewards $83,342; interest and finance fee income $66,128; other subscription and services $113,528 | Same |
| USDC rewards and blockchain rewards fees, Q2 2026 and Q2 2025 (thousands) | $119,108 and $102,521; $53,102 and $89,157 | Same, Note 17 and the transaction expense table |
| Revenue lines, Q1 2022 to Q2 2026 | Six lines and totals by quarter, each row on its filing's basis | Coinbase 10-Qs and 10-Ks, 2022 to 2026 (data-quarterly.csv) |
| Average USDC in circulation, Q2 2023 to Q2 2026 | $25 billion (Q4 2023) to $77 billion (Q2 2026) | Coinbase Q2 2026 earnings deck, pp. 13 and 40; earlier letters |
| Corporate-balance stablecoin revenue, Q1 and Q2 2026 | $18 million and $28 million | Same, p. 40 |
| Federal funds rate, quarterly average | 3.63% in Q2 2026; 5.33% at the 2023 to 2024 peak | FRED series DFF |
| Bitcoin and Ether prices, quarterly average | Daily Coinbase prices | FRED series CBBTCUSD and CBETHUSD |
| Coinbase's rate sensitivity, 2025 | 150 basis points = $540.3 million of stablecoin revenue | Coinbase Form 10-K for 2025, Item 7A |
| Q3 2026 outlook and Q2 2026 outlook | Subscription and services $500 to $580 million; Q2 outlook $565 to $645 million | Coinbase Q2 2026 earnings deck, pp. 31 and 23 |
| Chance benchmark | 2,000 draws of six independent random walks, taken as overlapping four-quarter changes over 12 and 16 observations, fixed seeds | Computed in the model |
The model (model.py), its output and the facts files behind each input are kept with the inquiry's working files; the arithmetic in the text was checked against them.
This inquiry is research, not investment advice. It states what the filings and models show under labeled assumptions; nothing here is a recommendation to buy, sell or hold any security, and the decision belongs to the reader. As of October 4, 2026, Russ W. Rosenzweig, Vista's founder, owns shares of and holds options on Coinbase, CoreWeave, Nvidia, Microsoft, Amazon and Airbnb, all named in this inquiry, and owns shares of Charles Schwab; he owns no shares of Circle Internet Group, IREN or Cipher Mining. Holdings through mutual funds and exchange-traded funds are not counted. One passage and two inputs were corrected on October 7, 2026, the day of publication: the first version said the 10-K's implied share of about 55% was "the 2024 and 2025 level," when we measure 51% to 55% for 2024 and 49% to 51% for 2025; the table of inputs gave the average USDC float as $30 billion to $77 billion, when its low was $25 billion in the fourth quarter of 2023; and it described the chance benchmark as independent normal series, when it uses random walks in overlapping windows, as the text says. How this inquiry was made: written by Russ W. Rosenzweig with Vista's AI research desk. It began with Russ W. Rosenzweig's hypothesis about Coinbase's second quarter of 2026, drafted with the help of another AI assistant, and the desk challenged it before this inquiry was written; the models were built and every figure computed in code, each public source was checked against its primary document, and the draft was fact-checked against those documents again before he read it. I read it and listened to it before it was published. Why I write with AI.