Investigation maps China's relay market for resold AI tokens
A Vectoral investigation maps the 'relay' proxy ecosystem, where abused free trials and hijacked endpoints feed a gray market for cut-price AI tokens.
An investigation by Matt Lenhard of Vectoral, highlighted on July 26 by researcher Simon Willison, maps the "relay market" — a gray-market ecosystem, active primarily in China, in which proxy operators pool illicitly obtained API credentials and resell large-model access at steep discounts.
Where the tokens come from
The supply side runs on several channels: systematically abusing free-trial credits; hijacking unprotected customer-support chatbots, which can be proxied into the LLM APIs behind them; and, in some cases, stolen credit cards and chargeback fraud. The trade is built on open-source gateway software such as one-api and new-api, which makes standing up a reselling proxy nearly turnkey. Demand is equally structural: buyers want cheap tokens, a way around geographic restrictions — US frontier models are not officially available in China — and bulk model outputs for distillation.
A laundering layer for API traffic
The relay layer does more than move discounted tokens; it obscures who is actually consuming a model and from where. That has direct policy relevance: the US Treasury threatened sanctions this week over alleged distillation of Anthropic's Fable models, but enforcement presumes that API traffic can be attributed. A mature relay ecosystem is precisely the infrastructure that breaks that assumption. For developers, the finding is more immediate: any LLM endpoint exposed without authentication is not a minor leak but a monetizable asset that reselling operations actively scan for.
The investigation matters because a great deal of AI policy and commercial practice — geo-restrictions, usage policies, per-customer rate limits, anti-distillation clauses — quietly assumes that access controls at the API layer work. This is documented evidence that a scaled, professionalized market already defeats those controls, and that the gap between where frontier models are officially available and where they are actually used is far narrower than the market maps suggest.