China's gray market undercuts Claude at 90% discount via proxy networks

Anthropic's access controls rank among the strictest of any major AI provider for China—phone checks, foreign credit card verification, billing address validation, and even ID verification with live selfies for select users. Yet none of these barriers have stopped a thriving underground market. The transfer station infrastructure succeeds precisely because it is modular: account brokers, SMS verification platforms, proxy operators, and downstream resellers each handle a single link in the supply chain. When Anthropic bans one provider, the others remain intact and can reconnect within hours. Newer identity checks, which the company introduced to tighten control, already have documented workarounds—AI services generate realistic fake IDs, and deepfake technology defeats biometric checks, with real people in low-income countries sometimes recruited to complete KYC verifications for under $30.

The pricing advantage flows from multiple sources: free credits, discounted enterprise and education plans, token-quota sharing, and most significantly, the ability to monetize the data streams flowing through proxies. Every request—prompt, response, tool call, and iteration—is visible to the proxy operator and potentially valuable for training or distillation. Datasets with Claude Opus 4.6 outputs already circulate on HuggingFace with unclear provenance, and Chinese developers indicate that token sales are merely customer acquisition; the real margins lie in selling logs. While Qian's analysis cannot confirm systematic log monetization or identify buyers, she argues that rock-bottom pricing becomes viable only with this additional revenue stream, making users both paying customers and unpaid data producers.

The broader implication cuts across policy and security. Access restrictions intended to prevent geopolitical technology transfer have created exactly the circumvention markets they sought to prevent. The same infrastructure that allows a developer in Beijing to buy cheap Claude tokens also enables a bad actor to reach frontier models without being traced—since Anthropic initially sees only the proxy's account and IP, not the end user. This weakens monitoring systems designed to detect coordinated abuse across accounts. Beyond AI, the circumvention ecosystem feeds criminal markets: biometric data collected for KYC workarounds is resold for financial fraud and deepfakes, and account farming operations support spam, phishing, and credit card fraud. Anthropic, OpenAI, and Google have begun coordinating against distillation, but industry opinion is divided; in late July 2026, 25 companies including Nvidia, Microsoft, and Meta warned against premature restrictions on distillation, arguing it is a normal business practice. With the US government unlikely to regulate distillation alone, AI labs are left to enforce their own terms against methods that are admittedly illegal yet still deliver the distillation data they fear most.