Anthropic's Prospectus Puts Account-Ban Revenue at 5-7% of Income—Pure Profit After Charging First
Paid Claude users face a business model where banning accounts is a revenue line, not safety: 5-7% of Anthropic's income is non-refunded payments after bans. Because a persistent device_id and backup traces survive a ban, any new account on the same machine dies instantly unless you wipe those local files first.
Anthropic's IPO prospectus discloses that 5-7% of its revenue comes from account bans, roughly $1.65–2.31 billion, and that money is nearly pure profit: users are charged first, then banned, with no refund. The number confirms what many users suspected about the economics of Claude's enforcement.
A reverse-engineering analysis of Claude's local code shows a persistent fingerprinting system. Every login writes accountUuid and a stable device_id to ~/.claude.json; that device_id survives account bans unless manually cleaned, so a banned machine instantly kills any new account. Claude also keeps a telemetry retry queue (~/.claude/telemetry/1p_failed_events.*.json) that buffers environment data during network fluctuations and sends it once stable, plus rolling backups (~/.claude/backups/.claude.json.backup.*) and OAuth tokens in the macOS keychain, all preserving traces of old accounts.
The server-side ban algorithm is a black box, and even a clean US-based physical machine gets banned when the system judges usage as non-compliant. The practical conclusion is that continuing to pay for Claude is risky; Codex, Kimi, and Grok now offer comparable capability.
A prospectus that lists ban revenue as 5-7% of income turns account enforcement into an audited business line; users' losses become shareholders' gains.
The local fingerprint stack (device_id, retry queue, backups, OAuth secrets) is designed so that a single ban can permanently taint a machine; wiping only one file is not enough.
Banning a clean US physical machine shows the algorithm judges behavior, not identity or location, so no amount of clean environment protects a user whose usage pattern looks non-compliant.
Shouldn't we reflect? Without the current environment, Huawei wouldn't be short of chips, and AI models wouldn't be short of cards.
I always thought banned accounts could get refunds; turns out they can't?