AI Token Relays Resell Model Access at 87% Discount, Vectoral Finds
A multi-layer grey market based in China uses pooled accounts and open-source gateways to bypass geo-restrictions and billing controls on frontier AI models.
A sophisticated reseller network is diverting high-value API access from Anthropic, OpenAI, and other frontier AI labs to customers at steep discounts, according to a new report from Vectoral. The relay market, operating primarily out of China, has matured into a four-tier ecosystem that collectively draws 3.6 million monthly visits across its top ten platforms.
How the Relay Chain Works
Vectoral researcher Matt Lenhard identified four distinct layers in the token relay supply chain. Upstream merchants sell credit cards and accounts. Midstream operators pool these accounts. Downstream relays, also called transfer stations, proxy traffic to end users including developers, SaaS companies, and buyers engaged in model distillation.
The relays rely heavily on open-source gateway software, particularly the projects 'one-api' and 'new-api', which provide OpenAI-compatible interfaces for managing multiple API keys and billing. These tools were designed for legitimate multi-tenant deployments but have been repurposed to obscure the origin of relayed requests.
Pricing Gaps and Incentives
The economics are stark. Vectoral documented relays offering the equivalent of $3,333 in official Anthropic credit for 425 RMB, approximately $59. That represents roughly $0.13 of official usage per dollar spent, an 87 percent discount that undercuts AI labs' pricing by an order of magnitude.
Some operators have gamified customer acquisition. The site hvoy.ai runs a daily lottery awarding fifty $100 API keys. According to Vectoral, the draw uses a Partial Fisher-Yates shuffle seeded by Bitcoin block hashes for provable fairness. Users earn entry tickets through daily check-ins, with up to three tickets per round at 20 points each.
Why Labs Care
The relay market directly undermines AI providers' business models by leaking compute capacity at near-zero marginal cost. Lenhard wrote in the report: "A relay crosses the line when its channels are stocked with stolen, leaked, or pooled keys instead of the operator's own, and when it resells that access against the providers' terms."
The Vectoral report notes that some commercial buyers use relayed tokens for model distillation, training domestic models using outputs from frontier systems. This claim appears in the report with screenshots from a V2EX forum thread.
For AI labs, the relay ecosystem creates a continuous fraud detection challenge. Providers must deploy behavioral analysis and strict know-your-customer controls to prevent what the industry calls "denial of wallet" attacks, where fraudsters exhaust credit limits through bulk account creation using stolen or virtual cards.
The pattern mirrors earlier grey markets in digital advertising and software licensing, where impressions and seats were resold through proxy networks. As frontier models become more valuable, the incentive to exploit billing loopholes and geo-restrictions grows accordingly.