Someone Else Pays for Your AI Access

you probably didn't think about this when you signed up.

you entered your card details, verified your phone number, maybe uploaded a government ID and took a selfie. friction. annoying. you moved on.

somewhere in cambodia or kenya, someone did the same thing. except they weren't signing up for claude. they were being paid — under $30 — to complete a verification step on behalf of someone they'll never meet, for a service they'll never use, in a supply chain they don't fully understand.

their face is now in a database they didn't choose. it will be used again. not for claude.

every time anthropic tightens access to protect its models, the evasion doesn't stop. it migrates.

geoblocking produced vpn services. phone verification produced sms farms. credit card requirements produced stolen card networks. biometric kyc — live selfies, government id matching — produced agents traveling to lower-income countries to recruit real people willing to complete in-person verification for cash.

the controls and the evasions are a paired system. you can't have one without the other. and the cost of the evasion doesn't stay where the models are. it moves to wherever people are poor enough to trade their biometric data for $30.

the fable shutdown made this visible in a new way.

on june 12, 2026, anthropic disabled fable 5 and mythos 5 for every customer worldwide — not because of an outage, not because of a flaw they found, but because the us government issued an export control directive at 5:21pm Anthropic's official statement and there was no way to segment foreign nationals from us persons in real time. so they turned it off for everyone.

gabriel attal compared it to iran blockading the strait of hormuz AI Frontiers Media. brussels talked. developers in san francisco talked about reliability. nobody talked about cambodia.

the transfer station economy — documented in may 2026 by oxford researcher zilan qian ChinaTalk, May 5 2026 — has been running this supply chain in public for years. github. taobao. telegram. chinese developers accessing claude at 10% of official price through api proxies that sit between them and anthropic's infrastructure.

the three ways the price gets that low:

first, account arbitrage — bulk-registered free credits, unused quotas, carved-up max plans.

second, model swapping — you pay for opus, you get haiku, sometimes you get glm. you can't verify which model answered you.

third, the logs. every prompt, every response, every tool call, every reasoning trace sitting on a proxy operator's server. for a developer using claude code, that's your repository context, your engineering decisions, your verified correct outputs. the markup business is customer acquisition. the logs are the margin.

but the third meal isn't just data extraction. the upstream supply chain that keeps the proxy pool running needs verified accounts. verified accounts need identities. identities increasingly need biometrics. and biometrics, when ai deepfakes get good enough to detect, need real humans.

so agents go to cambodia. agents go to kenya. they find people willing to complete verification for under $30. those faces enter a database. that database doesn't stay in the claude access supply chain.

the chinese developer paying 10% for tokens didn't order this. they're trying to build something with the same tools everyone else has, priced out by geography the same way a developer in lagos is priced out by latency and infrastructure. neither of them sees the person whose face just got harvested in cambodia. neither of them chose the system that makes that harvesting profitable. they're both downstream of a fight they didn't start, between parties who will never absorb the cost themselves.

the worldcoin black market documented this pattern before anyone was paying attention. iris scans harvested in cambodia and kenya, sold for under $30. the same infrastructure. the same geography. the same people absorbing costs they didn't choose.

this isn't new. content moderators in kenya process trauma for platforms they'll never use. data labelers in colombia annotate images for models trained in san francisco. the biometric harvesting is the same supply chain, one layer deeper.

a face verified to bypass anthropic's kyc today can be resold to open a fraudulent bank account tomorrow. it can generate a deepfake. it can be used for blackmail. the original subject in the global south bears the legal and reputational consequences of a transaction that had nothing to do with them.

i build in port harcourt. every api call i make crosses an ocean and costs latency i can't engineer away. i wrote about that recently — the physics problem nobody warned you about.

this is the other side of that piece.

the infrastructure gap isn't just latency. it's who absorbs the externalities of the access war. when two parties fight over who gets to use a model, a third party — somewhere with weaker institutions, fewer legal protections, and more financial pressure to say yes to $30 — pays the cost neither of the original parties wanted to carry.

that's not a side effect.

the controls will keep tightening. fable has been offline for seventeen days. mythos was partially restored on june 27 — only for critical infrastructure organizations the us government specifically approved. general users, developers, international subscribers are still waiting. gpt-5.6 is next in line for the same review process. each new restriction produces a new evasion layer, and each evasion layer reaches further down the economic ladder to find humans willing to be part of the supply chain for cash.

the people performing outrage about ai access — in brussels, in san francisco, in policy papers — are arguing about the front of the supply chain. nobody is arguing about the back.

someone else is paying for your claude access. you won't read about them in the policy papers.

AI helped me research, structure, and edit this piece. The arguments, the examples, and the opinions are mine. So is whatever's wrong with them.