Anthropic’s 5% to 7% revenue line is a detection problem, not a scandal

The sentence in the S-1 should not be read as “violating users contributed 5% to 7% of revenue.”

For the nine months ended September 30, 2026, about 5% to 7% of revenue came from customers later determined to have violated usage policies or commercial terms. Using $16.286 billion for the first half and the media-reported Q3 figure of $16.8 billion disclosed to investors, nine-month revenue is about $33.086 billion. The 5% to 7% range corresponds to roughly $1.65 billion to $2.32 billion. The S-1 states the revenue-recognition timing: the service had already been provided, and the restriction or termination came afterward. The money is consideration that accumulated during the lag between the violation, detection, investigation, and account closure.

For the same demand, if it is identified on day one, this revenue is close to zero. If it is identified only after three months, it is three months of billing. So 5% to 7% is first a financial projection of the detection and disposition cycle. “Compliance detection-lag revenue” is more accurate than “violating-user revenue.”

That projection is not clean. It needs to be broken into at least four ledgers.

First ledger: accounting and disclosure are separate

“The service was provided and revenue was recognized under the rules” is not the controversial part. The controversy is the phrase about recalculation after detection methods improve. The S-1 says this estimate may be adjusted because stronger monitoring will uncover previously undetected usage. That leaves legal room, but it does not equal a retrospective reduction of already-recognized revenue.

What is more likely to happen first is a reclassification of disclosure categories. Part of historical revenue gets relabeled as “later determined to be in violation,” which makes the share look higher in retrospect. What actually affects the income statement is refunds, credits or compensation, refunds of unused prepaid balances, and postpaid bad debt. The S-1 does not promise to restate revenue, nor does it quantify refunds. Pushing “the estimate may be adjusted” into “past revenue will be recalculated downward” overextends the prospectus language.

A change in disclosure categories and a change in accounting revenue are two different layers of risk. Mixing them lowers the legal rigor of the analysis.

Second ledger: the ratio mixes six variables

The 5% to 7% range mixes the paid scale of violating customers; violation type, including content violations, identity or regional violations, and distillation; revenue structure by usage, subscription, or prepayment; actual time from discovery to disposition; whether unused balances are refunded upon account closure; postpaid bad debt; and new revenue from re-registration.

The logic holds that earlier detection brings this revenue closer to zero. A high ratio cannot directly prove that the risk-control system is inefficient. A better reading is that the ratio reflects the product of high-risk customer structure, disposition cycle, and payment model. If API resellers, bulk accounts, and distillation-type demand make up a high share of the customer base, such demand is naturally short-lived, high-prepayment, and easy to re-register. Revenue accumulation during the detection lag will then be larger. This is not an excuse for risk control. It means the ratio has limited explanatory power as a single indicator.

One accounting detail matters. If unused prepaid balances are refunded when an account is closed, actual retained revenue will be lower than recognized revenue. If they are not refunded, the figure is closer to the full scale of lagged revenue. The S-1 does not disclose this split. The 5% to 7% range is more like an upper-bound estimate than a precise violation retention rate.

Third ledger: closure and re-registration are not net incremental revenue

The transparency-report data is striking. In the first half of 2026, 11.4 million accounts were banned, 398,000 appeals were received, and 42,000 were overturned, an appeal success rate of about 10.5%. A scale of 11.4 million cannot correspond to 11.4 million different individuals or enterprises. Resellers hold bulk accounts and switch to a new one when one is banned. That is an openly known way of operating in the industry.

The judgment that the same demand is repeatedly charged needs three qualifications.

First, account count does not equal paying customer count. How many of the 11.4 million are free accounts, trial accounts, or shell accounts registered in bulk is not broken out by the S-1 or the transparency report. A large number of bans may generate no revenue at all and are only risk-control actions.

Second, revenue from re-registration is not pure increment. The platform repeatedly receives subscription fees or prepayments, and it also bears risk-control costs, payment chargebacks and refund costs, payment-channel costs, and customer-service appeal costs. If the same demand is split into ten short-lived accounts, recognized revenue may rise, but net retention per unit of demand may not. Writing repeated charging directly as a profit contribution overstates the economic significance of re-registration.

Third, the 10.5% appeal success rate shows that wrongful bans exist. It cannot be reversed into “a large number of banned accounts are normal paying users.” People willing to write appeals and attach work proof are likely high-frequency, heavy users. The 398,000 appeals are only a fraction of 11.4 million bans. Among the silent majority, how many are genuine violations, how many are identity non-compliance, and how many are wrongful hits where users simply gave up cannot be distinguished from the available data.

Re-registration splits the same demand into multiple short-lived accounts and creates repeated recognized revenue during the detection-lag period. Net benefit must deduct refunds, chargebacks, and risk-control costs. Closure and re-registration cannot be treated as a clean revenue amplifier.

Fourth ledger: the regional-restriction portion is identity non-compliance

In September 2025, Anthropic updated its terms. Entities directly or indirectly owned more than 50% by Chinese capital, regardless of whether they are registered in mainland China, Singapore, or the United States, fall within the restriction. The S-1 does not break down the 5% to 7% by violation type. A reasonable inference is that a substantial part is not users generating illegal or abusive content. It is identity or equity structure being judged as an unsupported region.

This demand has different characteristics. Writing code, editing documents, and doing research are ordinary productivity tasks with strong willingness to pay and limited alternatives. This is the root reason the reseller business has remained viable. For these users, a ban is not punishment for a violation. It is being cut off for identity non-compliance. The incentive to switch to another shell and keep paying is very strong. This raises the re-registration component of detection-lag revenue and weakens the narrative that bans compress revenue in the face of demand.

This must be clearly labeled as a reasonable inference, not a fact disclosed in the S-1. Writing “much of the 5% to 7% is regional-restriction revenue” goes beyond the documentary basis. Writing “it cannot be ruled out that a considerable share is identity non-compliance revenue, because the reality of demand and willingness to pay do not support the assumption that a blanket ban will work” is defensible.

Disclosure is a legal act; materiality is not a hard 5% cutoff

A revenue source that accounts for 5% to 7% of revenue, that the company itself is actively compressing, and that involves account bans and regional restrictions would, by common materiality standards, be the kind of risk that must be disclosed. Under U.S. securities law, materiality is not a simple 5% threshold. It is a judgment combining likelihood and magnitude. This ratio, together with the forward-looking statement that detection and disposition were strengthened in August and September, that the share is expected to decline, and that this may affect overall revenue and growth, forms a fairly complete advance risk communication.

Its first function is to provide a defense. If Q4 or later growth looks ugly because violating revenue is being compressed, the prospectus already says it may reduce future overall revenue and growth. When plaintiffs’ lawyers go through it line by line in a class action, this is a ready-made basis for defense. The second function is the PR and safety narrative: “we banned accounts to the point that it hurt revenue, and we still banned them.” For regulators and enterprise customers, this can be read as a statement. For investors, it is first a growth headwind already priced into the risk factors.

Disclosure is first a legal act, and only second a PR act. Reading it as a declaration of conscience misses what the document is doing.

The October 12 announcement: misjudgment costs begin to shift to users

The direction of the October 12 announcement is consistent with that sentence in the S-1. In specific areas such as model distillation, starting September 24, Anthropic resumed charging for requests blocked in advance by the safety system. The stated reason is that the false-positive rate is below 0.1%. The company also acknowledged that normal requests that are wrongly blocked will be charged.

The logical extension of this rule is that the cost of violations is borne by violators, and part of the cost of misjudgment is borne by users. For the platform, this internalizes the externality of risk control. If you send a request that may trigger a block, even if it is ultimately judged benign, the cost of the blocking action itself is no longer absorbed unilaterally by the platform.

Three qualifications matter. First, 0.1% is a request-level false-positive rate, not an account-level one. On a base of billions of requests, the absolute number is not small. Second, the charging scope is limited to specific areas such as model distillation, not a global resumption of charging for blocked requests. Third, whether the appeal and refund mechanisms after wrongful blocking are symmetric is not elaborated in the announcement. This rule may contribute revenue in the short term. The backlash in fairness of terms and customer relations cannot be judged only from the revenue side.

Sensitivity: fill in the base, then read Q4 growth

The core judgment holds. The 5% to 7% range is detection-lag revenue, not fines on a violating group. Stronger detection will compress future growth. Disclosure is first a legal act. Three points need restraint.

Do not treat 5% to 7% as a clean indicator of risk-control efficiency. It mixes customer structure, payment model, disposition cycle, and refunds or bad debt.

Do not treat repeated charging from re-registration as net benefit. Account count does not equal paying customer count, and repeated recognized revenue must deduct refunds, chargebacks, and risk-control costs.

Do not treat the regional-restriction revenue share as a disclosed fact. It is reasonable to infer that identity non-compliance contributes a considerable share, but it cannot be quantified, and it cannot be written as something the S-1 says but does not say.

Here is the sensitivity in prose. If the detection cycle compresses from three months to one month, the lagged revenue retained from the same demand would shrink to roughly one-third, and the compressed amount would be about two-thirds of the original lagged revenue. If it compresses from three months to close to zero, retained revenue approaches zero and the compressed amount approaches the full original lagged revenue. If detection is strengthened only in August and September, the nine-month figure would contract only partially, depending on Q4 recognition timing and refund rates.

On the $33.086 billion base and a 5% to 7% range, lagged revenue is about $1.65 billion to $2.32 billion. If Q4 compression is $0.5 billion to $0.8 billion, the drag on Q4 year-over-year growth is approximately the compressed amount divided by prior-year Q4 revenue. If the prior-year Q4 base was in the $15 billion to $17 billion range, the drag would be about 3 to 5 percentage points. With a $20 billion base, it would be about 2.5 to 4 percentage points. The base is not disclosed, so that input is left open. Fill in the prior-year Q4 base and the refund rate before deciding how to read Q4 growth.