Behind The Scenes: How we mapped the AI relay market

After my first encounter with token fraud at my previous employer, where I worked as a soft engineer, I realized this market needs to be mapped in a way that is reproducible and verifiable. So we decided to write down our methodology separately, walking you through how you can retrace our steps and discover similar results.

Our investigation of the AI relay market brought together firsthand experience of inference abuse, discussions on a Chinese developer forum (V2EX), publicly advertised relay pricing, and third-party website traffic estimates.

Those sources answer different questions. A price listing shows what a seller advertises. A forum comment records what an operator claims. A response header can help identify gateway software.

In April 2026, I witnessed approximately $40,000 in usage associated with bulk account creation and free-credit abuse, followed a few weeks later by approximately $60,000 routed through a support chatbot’s endpoint.

Conversations with engineers at other companies helped motivate the investigation, but they do not constitute a representative survey. We do not use informal accounts of companies losing millions of dollars as a measured estimate of industry losses.

Over the course of my research, I found the V2EX discussion through a Reddit thread. That was the discovery path into the relay ecosystem:

https://www.reddit.com/r/ClaudeAI/comments/1rm0xt7/how_chinas_ai_token_reseller_ecosystem_works/

At the time, Fortune had already covered Stripe’s concerns about inference-related token fraud in May 2026, which provided background for our relay measurements.

The main qualitative source is the V2EX thread “A comprehensive guide to AI transfer station jargon”, posted by v2exgo, who identifies terminal.pub as his own relay.

The thread offers participant accounts of the market, but its author also has a commercial interest in it. We attribute statements about sourcing, refunds, customers, and earnings to their usernames participating in the discussion in this thread rather than treating them as independently verified facts. All the comments mentioned in our post can be found in the archived version of the thread.

For translation from simplified Mandarin to English, we used a Google translator extension to translate the entire page and the screenshots included in the blogpost are all taken from the translated page.

Machine translation can obscure technical slang, currency conventions, and distinctions between account pools, resellers, and upstream providers, but we have preserved the meaning as conveyed in the translation within the larger context of this underground market.

Identifying the 49 relays

To identify the relays mentioned in this post, we used a crawler to discover relay websites and scraped the price-comparison site linked in a reply to the main post on V2EX, getcheapai.com.

This is a discovery-based sample. It is not a random sample of the market, and 49 is not an estimate of the total number of relays. This is our first attempt at maintaining a directory of active relays and in follow-up investigations we aim to expand to a more comprehensive list.

Here is a list of URLs for the top 3 providers, which we have not embedded in the main post to avoid advertising them:

https://nowcoding.ai/

https://icodeeasy.cc/en/

https://claude-zhongzhuan.cloud/

How to interpret the pricing comparison

Relay balances can use dollar symbols for internal credit units. Those units are not necessarily USD. Comparing a relay with official API pricing therefore requires accounting for both the cost of purchasing credit and the multiplier applied when that credit is consumed.

For a matched amount of usage, the comparison is:

Discount = 100 × (1 − relay cost in USD ÷ official cost in USD).

If the relay price is in RMB, it must first be converted using a specified exchange rate and date. Renminbi or RMB is the official currency of the People’s Republic of China.

For example, a package advertised at 425 RMB for the equivalent of US$3,333 in official usage implies:

425 ÷ 3,333 ≈ 0.1275 RMB per US$1 of official usage.

That is approximately 0.13 RMB, not US$0.13. A USD discount requires the additional currency conversion.

Package comparisons also depend on whether the buyer can consume the advertised allowance. Expiration, daily limits, model restrictions, unavailable capacity, and unused balances can make the realized price higher.

Identifying the gateway software and market categories

The research used a fingerprint in HTTP response headers to identify one-api or new-api deployments.

Headers can be removed, copied, or changed by forks and proxies. The results mentioned in the original post are fingerprint-based identifications.

The supply-chain description follows the forum’s account of merchants, account pools, relays, and customers. We separate pools from customer-facing relays to explain their different functions, which does not mean that one business occupies exactly one layer.

The following categories will evolve as more research is published on this blog:

• Free-credit abuse.
• Open inference endpoints.
• Chargebacks and stolen cards.
• Prepaid and virtual cards.
• Denial-of-wallet.

The combined figure of 3.6 million monthly visits was obtained from Bishopi, an all-in-one AI-powered platform designed for domain investors, startups, and anyone interested in researching and monitoring domains for SEO or competitor intelligence data.