Washington can revoke Anthropic’s models at the border. It can’t stop them from showing up in Shenzhen at a tenth of the price.
What Happened
In a detailed report for ChinaTalk, Oxford China Policy Lab researcher Zilan Qian documents a thriving Chinese market for Claude access — sold through “transfer stations” (中转站), API proxies that route requests through overseas nodes to defeat geoblocking and KYC. Tokens change hands at roughly 1 RMB per dollar of usage: a 70–90% discount on Anthropic’s official rates.
The users aren’t only frontier labs. They’re professors, students, startups, and hobbyists. Anthropic’s own February 2026 report, cited in the piece, flagged 20,000-plus fraudulent proxy accounts. And by April 2026, Singapore — a convenient routing node — quietly topped the world in per-capita Claude consumption.
The key insight: Every control Anthropic added — geoblocking, phone verification, credit-card gates, live biometric KYC — spawned a matching evasion industry: SMS farms, fake-ID shops, even a black market in iris scans. The barrier didn’t stop access. It set the price.
One Fish, Three Meals
Qian’s sharpest framing is how a single proxied account gets monetized three ways at once — the user is simultaneously a paying customer and an unpaid raw material.
1 · Markup on access
Bulk-registered free accounts, resold unused quota, corporate-discount arbitrage, split subscriptions, and — in some cases — stolen-card purchases.
2 · Silent model-swapping
Proxies quietly route to cheaper models — Sonnet for Opus, or a Chinese model entirely — and relabel the output. One study found a “premium” proxy scoring 37% on a benchmark the real model clears at ~84%.
3 · Harvesting the prompts
Every prompt, response, and reasoning chain becomes a commodity. Claude Code logs — real repos, real decisions, human-verified outputs — are gold for distilling rival models. Datasets of frontier reasoning traces already circulate with murky sourcing.
The Structural Read
This is the other face of the story we covered when Austria asked the EU to host Anthropic. There, the lesson was that permission is the scarcest asset in AI. Here, the lesson is that permission leaks.
The Permission Layer — the governance tier where states decide which capabilities cross which borders — looks airtight from Washington. From the demand side it’s a sieve. Access control changes the cost of getting a model; it does not change whether you can. And cost is something a 20,000-account proxy network is built to absorb.
Zilan Qian · ChinaTalk
“Access blockage rarely stops determined users. It raises the cost of access — which in turn creates profitable markets.”
And the harms don’t stay neatly inside the geopolitical box. The same plumbing that lets a blocked developer reach Claude harvests biometric data sold to fraudsters, leaks prompts used for targeted scams, and feeds criminal markets — with many victims in the Global South. Qian’s blunt conclusion: the harms are not separable.
Three Implications
1 · Export controls are a tax, not a wall
They raise costs and degrade quality for the blocked side — but the capability still flows. Policy designed as a wall behaves like a tariff.
2 · “Buying Claude” may not mean buying Claude
Silent model-swapping means enterprises routing through cheap proxies can’t trust what’s answering them. Provenance becomes a first-order procurement question.
3 · Your prompts are the product
Reasoning traces and code logs are the highest-value distillation fuel there is. Anyone using an unofficial gateway is exporting their own IP one query at a time.
The Bottom Line
Anthropic built one of the most aggressive access-control stacks in software, and a market formed on the other side of it within months — cheaper, sloppier, and quietly siphoning the exact data that makes the models valuable. Restricting a model doesn’t contain it. It just decides who profits from the workaround, and what gets harvested on the way through. Permission is scarce, as Vienna argued this week. It is also, as Shenzhen demonstrates, for sale.
Source: “How to Buy Cheap Claude Tokens in China,” Zilan Qian, ChinaTalk.









