Circle Internet Group (NYSE: CRCL): what the price requires, and the one observable that decides it

01The view

Underweight, as an unsized research view. Not on the distribution waterfall. Hold retention at the company's own core guidance of about 39% and start supply from the $73.6bn actually outstanding rather than an assumed $80bn, and at the growth USDC has delivered in 2026 today's enterprise value requires 50.1% annual cash-flow growth after 2030. At the 40% compound growth the company guides, it requires 14.0%.

Those are the only two states that matter, and one observable separates them. The valuation is a levered bet on USDC circulation. Retention, which the earlier version of this work treated as the central question, moves the answer by six points across the range the evidence supports; the supply path moves it by twenty-seven.

What we had wrong. The earlier version of this work rested on a 36.9% retention assumption. That sits below the company's core guidance of 38–40%, below the 38.5% the June quarter actually printed, and 5.1 points below the estimate used by the most bearish covering analyst. It was a bias rather than conservatism. Correcting it in the bulls' favour is the right thing to do, and it does not rescue the valuation, because the binding variable is elsewhere.

Why this is a variant view. No covering house frames it this way. The published Underweight models retention above company guidance and supply up 20%, so its $37 target is a multiple assertion rather than a model output. The published Outperform's $140 requires 37.2× enterprise value to EBITDA, which is a re-rating rather than an earnings outcome. The Neutral $71 was set when the stock was $63.26 and has not been revisited 45% higher. Consensus is 47% Overweight with a mean target of $103.69, above spot, while the one observable that decides the valuation has fallen 2.3% this year.

What this document still does not contain. A twelve-month price target, a probability-weighted value, or a position size. The reverse discounted cash flow in Section 13 is a present-value diagnostic and is never converted into a target. The value grid in Section 21 is an illustration and carries no probabilities.

02What the business earns

Circle issues USDC and earns interest on the reserve assets backing it. It then pays most of that interest away under a contractual distribution waterfall. The retained share, not supply growth, is what reaches a shareholder.

USDC in circulation averaged $76.5bn over the June quarter and ended it at $73.3bn. Those are two different measurements and are never mixed anywhere in this document; Section 13 sets out why that matters.

Source: Q2 2026 Form 10-Q, condensed consolidated statements of operations and key operating indicators, period ended 30 June 2026 [S2].

03The distribution waterfall: what is contractual and what is calibrated

Contractual and disclosed. A 10 basis point issuer fee accrues to Circle on USDC in circulation. Circle and the counterparty each earn on balances held on their own platform. The counterparty then receives half of the residual after approved ecosystem participants are paid.

Calibrated, and not disclosed. The share of the residual absorbed by the wider ecosystem before that split is a model fit, solved at 36.8% so that the model reproduces the reported quarter. It is one free parameter fitted to one observation, and it is identified as such wherever it is used. The filings do not state whether approved participants are paid proportionally to rates, in fixed dollars, or in tiers, and that single unknown drives the whole of Section 4.

S is average USDC supply, y the reserve yield, c Circle's own platform share, b the counterparty's platform share, and L the ecosystem leakage term fitted at 0.368. k is the retention applied to interest above the issuer fee.

Source: Q2 2026 Form 10-Q, Item 3 quantitative and qualitative disclosures about market risk, and distribution-cost detail [S2].

04The marginal retention response: the assumption that anchors the forecast

Our published assumption was the outlier, and it was wrong. Company guidance for the FY2026 net revenue margin is 41.7–43.7% headline; excluding the $180m token presale that implies a core margin of about 39%, the middle of the reiterated 38–40% range [S3]. The June quarter printed 38.5%. The most bearish covering analyst models 42.0% for 2027. A 36.9% model average sits below all three, and Section 1 restates the conclusion at the 39% core guide. The assumed response already drives the forecast, and treating it as something that never anchors anything would be inaccurate. Circle's disclosed +100bp sensitivity shows $737m of incremental reserve income against $360m of incremental distribution costs: net retention of 51.2% of the increment, under its specified June-end assumptions.

What the third row does to the rest of this document. Holding the model's net reserve income at a 3.49% yield and applying the disclosed marginal response only to the increase toward the 2027 curve yield gives EBITDA of about $926m instead of $816m and economic cash flow of about $405m instead of $320m. At that level the expectations comparison in Section 6 inverts: EBITDA would sit above accessible public estimates rather than in line with them.

Three qualifications, stated rather than buried. The third row is a sensitivity, not a replacement forecast: balances, platform mix, cost categories and contractual behaviour all need reconciliation before it could become one. The cause of the discrepancy is not uniquely identified as fixed versus variable partner payments; those are two candidate explanations of a gap that may have others. And a higher marginal retention raises the level of earnings while increasing exposure to falling rates, so it is not simply a more favourable assumption.

The rate curve is sourced. The conversion of that curve into Circle earnings remains conditional on where in this bracket the true response sits.

Source: Circle Q2 2026 Form 10-Q, Item 3 [S2], filed for the period ended 30 June 2026.

05The rate input

Both operating paths run on one curve, taken from the CME final settlement bulletin #179 of 17 September 2026, page 10 side 01. Twelve 2027 monthly fed-funds contracts, calendar-day weighted, imply an average effective federal funds rate of 4.5094%. Later years use the same bulletin's quarterly SOFR contracts, converted to a flat daily equivalent and weighted by calendar days.

Stated assumptions. The 14bp and 17bp gaps are calibrated to the June quarter and assumed to persist; neither is a market-quoted spread. Futures convexity, risk premium, business-day compounding and intra-quarter rate paths are not modelled. These are proxies, not forecasts.

Because the curve is identical across both paths, it sets the level and not the ordering. At a curve one hundred basis points lower the required post-2030 growth rates become 60.3% and 28.6%. The ordering survives; the level does not.

06Against expectations

Aggregate agreement is not agreement on drivers. The same $805m reproduces at 35.9% retention with $640m of expenses, at 36.9% with $676m, and at the June-quarter 38.5% with $731m. No inference about how the market models retention is available from published aggregates, and none is drawn here. Being within roughly one percent on revenue and one on EBITDA establishes that there is no headline earnings variant; it does not establish that the drivers agree, and a phrase such as “the earnings are not in dispute” would overstate what the aggregates prove.

The estimate snapshots used are not a certified closing consensus and individual analyst adjustments are not disclosed. They are read as a positioning reference, never as a valuation anchor.

07Against the published houses

Three notes are available: Goldman Sachs (Neutral, $71, 5 August, written when the stock was $63.26), Bernstein (Outperform, $140, 24 August) and Morgan Stanley (Underweight, $37, 15 September). Each reports profitability on a different definition, so each has been restated onto the definition used here: adjusted EBITDA excluding stock compensation, with net interest income removed.

What this strengthens. Goldman decomposes the token distortion exactly as Section 16 predicted: headline other-revenue guidance was raised to $310–330m by the $180m Arc presale while the core guide was cut from $150–180m to $130–150m, so the raise was a one-off masking a reduction. Goldman also quantifies the migration at about $68.5m of annual net revenue, roughly 9% of 2027 adjusted pre-tax profit, with the full mechanism: about $3.7bn on the venue, roughly 90% moved onto the counterparty's platform, balances previously unincentivised and therefore shared evenly. And the measurement discipline of Section 15 is vindicated in the wild: Goldman cites USDC market capitalisation up 12% year on year while Bernstein's own series shows circulation down 2.3% year to date. Both are true on their own basis, and they point opposite ways.

What this weakens. First, the retention assumption, treated in Section 4. Second, the claim of being within one to two percent of consensus: across four models 2027 adjusted EBITDA spans $776m to $909m, a 17% range, and the $805m vendor proxy used as the reference sits below every named house except Goldman. Agreement with a vendor mean is not agreement with the market. Third, the share count: implied diluted shares run 253.9m, 276.8m and 290.8m across the three houses, a 14.5% spread on a denominator this work treated as verified, worth about four points of required growth.

What it demonstrates rather than changes. Morgan Stanley reaches a higher EBITDA than this model with five points more retention, twelve percent more supply and twenty-seven basis points less yield. Four models cluster near one mean out of incompatible inputs, which is the clearest possible illustration of the point Section 6 makes in the abstract: an aggregate cannot be read as agreement on drivers.

08Required growth by retention and by supply

The grid below re-solves the identical specification across the two variables that matter, starting from the $73.6bn of USDC actually outstanding on 21 August rather than from an assumed base. Enterprise value is held at $22,788m, discounted at 12% with 3% terminal growth.

Supply dominates retention by a factor of about three. Moving retention across its entire remaining bracket, from the company's own guide all the way to the disclosed marginal response, is worth nineteen points. Moving supply growth from twelve percent to thirty is worth twenty-seven. That reverses the emphasis the earlier version of this work carried, in which the distribution waterfall was the central question. It is second order.

The price needs roughly the growth the company guides. At 40% compound supply growth and the company's own retention guidance the requirement is 14.0%, which is an ordinary business outcome. At 12% it is 50.1%, which is not. There is no comfortable middle: supply would have to compound at about 28% to bring the requirement to 25%, and at about 33% to bring it to 20%. USDC circulation is down 2.3% this year.

09Economic cash conversion, and what it does not establish

Expensed and capitalised compensation are distinct charges, and the related payroll tax is cash rather than noncash. Nothing is charged twice, and the $245m line is not described as wholly noncash. Cash capital expenditure and capitalised development are already deducted, so this is a figure before acquisitions and incremental required capital, not before all growth capital.

This is not normalised consensus. It imports this model's compensation assumptions. Tax, working capital, token receipts, corporate interest, leverage and other conventions remain unresolved, and the definition of the published forecast has not been reconciled: it is stated with $97m of capital expenditure, which does not match the $92m used here. The full $245m compensation line is deliberately not subtracted, because cash payroll taxes may already be deducted inside a conventional free-cash-flow measure. The residual could reflect an actual forecast difference, a definitional difference, or both. The cash-conversion disagreement is unestablished pending normalisation.

One claim that is not made anywhere in this document. That a headline EBITDA multiple implies any particular cash conversion. It does not: the same multiple can reflect different conversion, growth, duration and required-return assumptions, so no conversion rate is identified from it.

Source: public cash-flow forecast snapshot [S4], retrieved 18 September 2026; definition not reconciled. The accessible snapshot does not certify a closing consensus or disclose individual analyst adjustments.

10The compensation-normalised multiple

ICE and Nasdaq keep ordinary stock compensation inside non-GAAP operating profit. Circle's adjusted EBITDA excludes it, together with the related payroll tax: $245m, or 30% of 2027 modelled EBITDA. Charging it puts the denominators on one basis.

What this establishes, and what it does not. It corrects a major peer-denominator difference, and it is the cleanest claim in this work because it depends on no forecast. It does not put every vendor estimate, investment holding, growth profile or capital structure on a fully identical basis. It corrects one difference, not all of them, and the residual differences are not quantified here. It also cuts against the bearish reading: a 14× headline reference is 20.0× once normalised, which is a premium to two of the three peers. The implied enterprise value at that reference is $11,418m.

11Two operating paths to 2030

These are two internally consistent continuations, presented as bundles. They change six assumptions together, so the difference between them cannot be attributed to any one of them, and Section 12 shows how much of it one driver accounts for.

12One-driver sensitivity

Changing only Path A's supply growth from 20% to 30%, and holding every other Path A assumption unchanged:

That single change explains about 51% of the A/B cash-flow gap in this ordered sensitivity. It is not a unique attribution, since the drivers interact, so the figure is the share attributable in this ordering rather than a decomposition. It shows that supply growth remains economically important and that any framing attributing the gap to platform mix alone is wrong on its own model. Quoting 33.3% against 14.1% without 21.6% overstates how much of the distance is about retention.

13What the price requires: a present-value diagnostic

Ten explicit flows from 2027 to 2036, the first discounted one period from today, r = 12%, g = 3%. Enterprise value is held at $22,788m and the only variable solved is the 2031–36 growth rate.

What a solved rate is, and is not. Each figure is whatever growth makes that path match today's enterprise value at a 12% required return. Both paths match current enterprise value by construction, so neither validates any expected return and neither can be used to support one. Saying that today's price comfortably supports Path B reverses the logic: 14.1% was solved to match today's enterprise value, so it establishes no independent valuation cushion.

Path A is the more demanding continuation, not an impossible one. Path A itself compounds economic cash flow at 33.4% through 2030; what is demanding is holding that pace for another six years, and the early period does not validate the extension. A categorical impossibility claim would need an operating argument rather than a percentage, and none is offered. Neither post-2030 path has been substantiated. Terminal value is 69% of the Path A present value and 59% of Path B. Quarterly results can update confidence in which continuation is running; they cannot settle a decade of growth duration. This diagnostic is not converted into a twelve-month price target anywhere in this document, and no target is offered.

14The 2036 operating endpoints

A required growth rate is an abstraction. The endpoint it implies is something that can be underwritten, against balances, retention, cost structure and the investment needed to reach it.

The endpoints are close because the enterprise value each must justify is the same. Path A needs a far higher growth rate from a far lower 2030 base; Path B needs a modest rate from a high base. The paths therefore differ far more in the shape of the required continuation than in its destination, which makes this an argument about duration and investment rather than about whether a number is achievable. What has not been done: the investment required to reach either endpoint has not been underwritten, and neither has been tested against a capacity or market-size constraint. Both are listed as open items in Section 22.

15Platform share: two series, two measurements

The improvement from 7.4% to 19.5% is a daily weighted-average series. The quarter-end holding of $12.442bn is about 17.0% of circulation, below that quarter's average. Annual improvement on an average series and the latest endpoint convey different information, and neither observation alone establishes the next trajectory.The momentum in the average series favours the stronger path. No commercial evidence has been identified for why penetration should subsequently slow enough to favour the weaker one, and none is asserted. The strongest conclusion the data supports is that the trajectory is unresolved. Every platform threshold elsewhere in this document is stated on a named basis, quarterly average or endpoint, and the two are never compared to each other.

Source: Q2 2026 Form 10-Q, key operating indicators and their stated definitions [S2].

16Q3 2026: a monitoring event, not a quantified catalyst

The migration is now quantified, and it is a permanent step-down rather than an event. About $3.7bn of USDC sat on the venue when the agreement was signed and roughly 90% has since moved onto the counterparty's platform. Those balances were previously unincentivised, so their economics were shared evenly between the two parties; once they sit on the counterparty's platform that share is lost. At a 4.1% reserve yield the arithmetic is $3.7bn × 90% × 4.1% × 50% = about $68.3m, against the $68.5m a third party publishes, described there as roughly 9% of 2027 adjusted pre-tax profit [S4]. On this model that is 8.4% of 2027 EBITDA and 21% of economic cash flow.

It settles the size and the mechanism, which the earlier version of this work could not. It does not make the quarter a catalyst: the effect is a lower recurring base rather than a surprise, its existence was disclosed in August, and it does not establish a full quarter at completed migration levels. The label stays a monitoring event.

The token recognition bridge. About $180m of token revenue is guided for the second half, with no quarter split disclosed. September's launch supports a Q3 possibility but does not establish the amount. Roughly $62m of deferred consideration would remain after $180m is recognised, and that bridge has to be carried explicitly into any 2027 comparison. Cash received, consideration deferred and revenue recognised are three distinct things and are kept distinct here; $242m of presale sits in deferred revenue.

The surprise that would force estimates lower: recurring other revenue below $90m for the quarter, with token recognition disclosed separately, combined with distribution cost at or above 62% of reserve income. That combination cannot be explained by the one-off and would move 2027 estimates.

Recent industry events, sized honestly. Arc launched on 16 September with its institutional validator cohort rolling out in phases, and the SEC issued conditional relief for tokenised NMS-stock secondary trading on 17 September. Neither establishes incremental fee profit for Circle, and neither requires the use of USDC or Arc. They are relevant commercial developments, not quantified earnings catalysts.

Sources: Q2 2026 earnings call transcript and Q2 2026 analyst call transcript, pp. 6–7 [S3], 4 August 2026; company release and SEC release [S5]. No Q3 2026 earnings date has been announced.

17The evidence that would settle it

Platform share appears twice because the average and the endpoint told different stories in the June quarter. A threshold on one basis is not a threshold on the other, and mixing them is how a mixed signal gets read as a clean one. Every threshold is stated on the basis Circle uses in its own key operating indicators, so each is readable directly from the quarterly disclosure. These observables update confidence in which continuation is running. They cannot settle a decade of growth duration, and no number of quarterly prints will. What is still missing is an evidence-based probability for either path; none is offered, because the trajectory is unresolved and inventing one would give the conclusion authority the work does not support.

18What would change the view

There is no position to cover, reduce or exit. These triggers change the view and the model, which is the only thing an unsized research call can do. The primary one is observable continuously and independently of company disclosure, which is why it is not a quarterly line.

Why the earlier circulation trigger was not a falsifier. Four weekly observations at $90bn do not invalidate an $80bn annual average: four weeks at $90bn and forty-eight averaging $79.17bn produce exactly $80bn. The trigger above is stated as a sustained thirty-day average for that reason, and its consequence is stated in required growth rather than as an instruction.

The asymmetry of the test. The bull case requires a sustained reflation to be confirmed. The bear case requires only the absence of one. That asymmetry is the reason this is now stated as a view rather than as a watchlist entry, and it is also the reason the view is unsized: the same asymmetry would make a position uncomfortable to hold through a crypto rally that lifted circulation for reasons unrelated to Circle.

19Three reasons this view is wrong

Written against the underweight.

1. Twelve percent supply growth is our assumption, not an observation. The 2026 tape is roughly flat, but a flat year inside a multi-year buildout is not a trend. Circulation troughed at $71.9bn in early August and added supply into late August. If 2026 turns out to be a digestion year after a doubling, the 25–30% column is the right one and the requirement is 23–28% rather than 50%. Test: the thirty-day average through the December quarter.

2. We were already wrong once on this name, on the parameter we were most confident about. The published version of this work rested on a 36.9% retention assumption that sat below the company's guidance, below the printed quarter and below the most bearish covering analyst. Being wrong in the same direction twice on the same name is a base rate a reader should price, and it is the honest reason this view is unsized. Test: whether the supply assumption, like the retention assumption, turns out to sit outside every external reference.

3. Circulation may be the wrong denominator entirely. USDC reached a record share of stablecoin transaction volume in the June quarter even as balances fell, and the payments network is scaling toward $23bn of annualised volume with monetisation only beginning in the second half. If the business migrates from balance economics to transaction economics, a model keyed to circulation measures the wrong thing. Test: other revenue excluding token recognition, against the $130–150m core guide.

The honest weight on these: the first is live and unresolved, the second is a statement about the analyst rather than the company, and the third would take years to appear in earnings. None of them changes what the December circulation print means.

20Executability

The view is unsized and no position policy is applied, so an executability gate has nothing to gate. It is run here because the methodology requires it before valuation, and because the liquidity claim needs narrowing.

The historical volume record supports the conclusion that a position of ordinary size can be traded in ordinary market conditions. It does not support an unconditional conclusion that institutional exit liquidity is unconstrained: volume in a stressed tape is not volume in a normal one, and no stressed-tape evidence has been gathered. Derivative expressions are out of scope: no option-implied move, implied volatility or spread statistic is computed or quoted anywhere in this document.

21Appendix: an illustrative value grid, with no probabilities

Net liquid assets $2,619m; fully diluted shares 276.8m. The current price is $91.74.

Why there are no probabilities here. A probability-weighted value would require evidence-based probabilities over the combined supply, retention and cost outcomes, and no such evidence base exists. The reverse discounted cash flow supplies neither the probabilities nor the exit multiples, and both solved paths match today's enterprise value by construction, so they cannot validate any expected return. Assigning a weighted value would give a headline more authority than the work supports. Every EBITDA case above is produced by the same operating model used throughout this document, including the retention bracket of Section 4. This grid shows the arithmetic sensitivity of value to two inputs and nothing more. It is not a valuation conclusion and no weighted value is derived from it.

22Sources, stated assumptions and open items

Stated assumptions. Reserve-to-benchmark spread 14bp for 2027 and 17bp thereafter, assumed to persist. Ecosystem leakage 36.8%, fitted. Normalised tax 23%. Incremental working capital at 10% of incremental fee revenue. Discount rate 12%, terminal growth 3%, first flow one period from today. Direct fee costs at 6% of other revenue. USDC reserve balances are not treated as corporate working capital. None of these is company guidance.

What would change the conclusion. A like-for-like cash-flow expectations bridge; a reconciled or explicitly bracketed rate-response model; and evidence linking the observed platform and supply trends to the probability and timing of a revaluation. Those are the decision-critical tasks.