Anthropic's S-1 Reveals 7% Revenue from Banned Users: Inside the Business Logic

On October 26, 2026, Anthropic filed its S-1 registration statement with the SEC, targeting a $1.5–2 trillion valuation. The filing disclosed that an estimated 5% to 7% of revenue in the first nine months of 2026 came from customers later determined to have violated usage policies or commercial terms.

在 2026 年前 9 个月里，Anthropic 估计约有 5% 到 7% 的收入，来自后来被认定违反其使用政策或商业条款的客户

Categories of violating users generating substantial bills before bans

Gray-market automation scripts: Bulk-registered accounts bypassing rate limits for high-frequency automated content generation (SEO spam, bulk marketing emails), consuming massive token volumes before suspension.

Security researchers and developers conducting "jailbreak" tests: Adversarial prompt attacks involving deep, multi-turn conversations that are heavy token consumers.

Enterprise customers with compliance gaps: Companies integrating Claude into workflows with data compliance risks or output policy violations, incurring high enterprise subscription fees before being audited and removed.

The author notes that many legitimate users also get caught in the ban net, including those accused of model distillation.

"高强度滥用"在财务数据上的滞后体现

Why Anthropic voluntarily exposed this "toxic" revenue stream

The author identifies three strategic reasons:

Regulatory and compliance signaling: The S-1 serves regulators as much as investors. Anthropic's brand moat is "AI safety and alignment"; tolerating abuse revenue would undermine that narrative. The filing states enforcement was tightened in August and September, and the violating revenue share is expected to decline.

Financial "bad debt" risk: Banned users generate chargebacks (via Apple/Google subscriptions), refund disputes, and high support costs. Net margin on this revenue is near zero or negative after hidden costs.

Growth engine pivot: Anthropic no longer depends on marginal users. H1 2026 revenue reached $16.3 billion, up over 1300% year-over-year. The valuation thesis rests on high-value enterprise adoption — specifically Claude Code users who drive net dollar retention above 900%. Sacrificing 5% low-quality revenue for 900%+ NDR is a calculated trade-off.

The author observes that refunds after bans are rare, which explains why the revenue share remains high.