The relay market powering token resellers and fraud

Original Article ↗ Hacker News Discussion ↗

Grey-market “token relay” services are emerging that resell access to frontier AI models like Anthropic and OpenAI at a fraction of official prices, often by exploiting flat-rate subscriptions, free trials, and in some cases stolen or refunded credit cards. Commenters debate whether this is primarily theft, contract breach, or arbitrage, and draw analogies to ticket scalping, all-you-can-eat buffets, and historical phone-phreaking and ad-fraud schemes. The thread also explores why subscription economics make such abuse attractive, how providers might detect or limit it (e.g. canaries, device fingerprinting, KYC), and the broader ethical and economic implications for both AI companies and legitimate users.

Pricing confusion and token economics

  • Several commenters question the math in the article’s example: 425 RMB (~$59) for $3,333 in Anthropic credit.
  • Multiple people argue the phrasing “$0.13 of usage per $1 spent” is backward; the intended meaning seems to be ~$1 of upstream usage for ~$0.13 paid to the reseller.
  • Others recalculate and claim the discount is even steeper (~$1 for ~$0.017), suggesting a missing zero.
  • Some note that if credit cards are stolen or chargebacks used, the effective “cost per token” is the fraud cost, not the face value.

What “tokens” are and aren’t

  • A user’s attempt to relate tokens to GPU hardware is rejected; others explain tokens as units of text processed by models, not hardware-generated currency.
  • Token counts roughly measure computation per model, but aren’t directly comparable across providers.

Fraud, theft, and ethics debate

  • Strong disagreement over whether this is:
    • straight credit card fraud/theft of services,
    • “mere” breach of contract/ToS,
    • or morally neutral arbitrage of mispriced subscriptions.
  • Some stress stolen cards, chargebacks, and “theft of services”; others argue the core issue is circumventing usage controls, not always stolen cards.
  • There’s a parallel to ticket scalping, all-you-can-eat buffets, car or apartment subletting, and gym-pass sharing.

Subscription models and arbitrage

  • Many see flat-rate, usage-capped subscriptions as the root cause: selling something below its market clearing price invites arbitrage.
  • Arguments:
    • Subscriptions as loss-leaders are inherently fragile to abuse.
    • Fixed per-token pricing is more robust but less attractive to heavy users.
    • Suggestions include tiered discounts at higher usage, prepaid transactions, and reconsidering subsidized plans.

Detection and mitigation of token fraud

  • Described tactics include:
    • Device fingerprinting (criticized as weak).
    • “Canary” values: buying from resellers, embedding unique markers in prompts, then tracing them back to accounts, banning, and training ML models on the patterns.
  • Commenters note this is a classic cat-and-mouse game similar to ad fraud and subsidized electricity abuse by miners.

Grey markets for broader cloud/AI credits

  • Discussion extends to resale of unused or abused AWS/Azure startup credits, with reports of inference sold at ~4% of nominal price.
  • Creating new entities to repeatedly access free credit programs is highlighted as a parallel abuse channel.

User experience and quality concerns

  • Buyers of resold access often can’t verify which model they’re actually getting; they rely on reputation.
  • Some suggest silent model downgrades (e.g., cheaper models sold as premium ones) likely occur but are hard to detect.