
Why would anyone run a business that sells AI access below cost and then publicly post their own losses? This week, on China's largest developer forum, one operator did exactly that: a frank, spreadsheet-driven postmortem of a "relay station" (中转站) — a third-party service reselling LLM API access at a discount to official prices — that lost roughly ¥30,000 (about $4,200) in its first month. The thread pulled in nearly 1,300 replies and became one of the week's most-read discussions on LINUX DO.
The short answer to the mystery is that the discount-API business has almost no moat and almost all of the risk: abuse, fraud, provider crackdowns, and a customer base that only came for the lowest price. It's a cautionary tale worth understanding even if you never touch one of these services — because the same economics explain why cheap API access keeps appearing, and keeps disappearing.
In Chinese dev slang, a relay station is a middleman service that sells you access to frontier models — GPT, Claude, Gemini, DeepSeek — through its own API endpoint, priced as a fraction of the official per-token rate. Listings routinely advertise 0.1x–0.3x of official pricing, sometimes with generous bonus credits on top. The pitch is aimed at students, indie hackers, and developers in regions where paying the official price in dollars is painful: same models, one-third the cost, no foreign credit card needed.
The operator who posted this week's confession started from that exact position. As discussed on LINUX DO, he described going from an ordinary forum member to running a commercial relay, motivated by his own memory of being a broke student and worker with nobody reimbursing token costs. The business model, in his telling, was to buy up consumer subscriptions in lower-priced regions, slice their quotas into reseller "packages" — standard, professional, flagship tiers — and sell them at 0.25x of the official rate with bonus credits stacked on the bigger plans. On paper, the math worked: buy quota cheap, sell it slightly less cheap, pocket the spread.
On paper, it always works. In practice, the thread is 1,300 replies long because everyone in the ecosystem has a story about why it doesn't.
The operator broke his losses into the classic gray-market triad, and commenters added a fourth.
1. The provider changed the deal mid-game. The single biggest blow, per the operator, was a quota cut from the upstream platform: weekly allowances on his top-tier accounts were roughly halved, from about 2,200 units to about 1,100. His cost base per account barely moved, while sellable quota dropped by half — flipping several account tiers from profitable to loss-making overnight. This is the structural trap of the whole business: relay operators build on subscription products they don't control, and the providers of those subscriptions are not in the business of subsidizing resellers. One quiet policy change upstream and your margin spreadsheet is fiction.
2. Accounts get banned, throttled, or "degraded." Forum users use a slang term — 降智, roughly "getting dumbed down" — for accounts that stop behaving normally: quality drops, limits tighten, access dies. The operator budgeted for some of this and still got it wrong. Every banned account is not just a lost subscription fee; it's the customer support load of everyone who bought quota on it, the refunds, and the reputation hit.
3. Fraud and abuse eat the margin. A discount API endpoint is a magnet for the exact users a legitimate business would least want: people reselling the reseller, automated scraping, users who charge back after draining their credits. Several commenters described the same pattern — you price aggressively to win volume, and the volume you win is the volume that loses you money. The honest low-use customers the operator was trying to serve were, by his own account, a minority of traffic.
4. The price war has no floor. With dozens of relays competing for the same price-sensitive buyers, the going rate keeps ratcheting down. The operator described effective cost ratios pushed down toward 0.07x–0.14x of official pricing just to stay listed — at which point a single bad month doesn't just hurt, it wipes out the quarter. One commenter summarized the consensus: the only relays that survive are the ones big enough to negotiate real upstream deals, and those aren't really relays anymore.
None of this is in the fine print of a relay's landing page, but it sits in the providers' own legal text. Anthropic's consumer terms, for example, are explicit: "You may not share your Account login information, Anthropic API key, or Account credentials with anyone else," nor "make your Account available to anyone else." (Anthropic Consumer Terms)
That means every relay built on consumer subscriptions lives one detection wave away from total shutdown — and every customer of such a relay is renting access that can vanish with no refund path. The operator's thread is unusually honest about this asymmetry: his customers' risk was his cost, until it became his loss.
The irony the thread kept circling back to is that the price gap relays exploit has been shrinking for two years. If what you want is legitimately cheap API access, the official rate cards are worth a look before you touch anything gray:
| Route | What you get | Verified pricing |
|---|---|---|
| DeepSeek official API | V4.1-Flash / V4-Pro, OpenAI-compatible endpoint | Flash $0.15 in / $0.60 out per 1M tokens off-peak; Pro $0.66 / $1.98 (deepseek.ai/pricing) |
| OpenRouter | One key, many models, transparent per-token pricing | Market rates with clear model routing |
| Claude Pro | Official subscription with high usage limits | $20/month (Anthropic's pricing pages, checked Sep 2026) |
DeepSeek's official API, for instance, prices its flagship Pro model at $0.66 per million input tokens off-peak — a rate that undercuts what relays were charging for the same model a year ago. The arbitrage window that made relays irresistible in 2024 has narrowed to the point where the risk premium barely pays for itself. That, more than any single fraud incident, is why the operator's math broke: he was competing against providers who keep cutting their own prices.
There's also the self-hosting middle path this community loves: running your own gateway like GPT-Load to pool your own subscriptions and keys behind one endpoint. It's the legitimate version of the relay pattern — same credential-multiplexing convenience, none of the reselling.
Threads like this one are how the Chinese dev community does collective price discovery. The operator didn't post a polished marketing story; he posted a loss report, with numbers, in a forum where a thousand practitioners could audit his assumptions in real time. The discussion that followed — part sympathy, part "we told you so," part detailed post-mortem of where the fraud came from — is a better education in API economics than most pricing whitepapers.
The deeper point is about why the gray market exists at all. The buyers weren't criminals; they were students and salaried developers whose employers reimburse nothing for AI tools — a situation the operator specifically set out to fix. When official pricing is denominated in dollars and your salary is in yuan, a relay offering 0.25x pricing isn't a luxury, it's the only way the math works. The sustainable fix isn't moralizing at buyers; it's what providers have actually been doing — cutting official prices, expanding free tiers, and making the legitimate path cheap enough that the gray one starves.
The same price sensitivity is what drives the coding-subscription shuffle this community is famous for: developers stack, cancel, and re-route subscriptions constantly, hunting for the cheapest legitimate way to keep coding. Until the gap between salaries and dollar-denominated pricing closes completely, relay stations will keep opening, keep undercutting each other, and — as this week's most-read confession shows — keep discovering that selling something for less than it costs is a great way to learn arithmetic. The rest of us get the spreadsheet for free.
As discussed on LINUX DO — relay-operator loss figures and business mechanics are the operator's own reported numbers, as debated in the thread; provider pricing and terms cited above were verified against the providers' official pages.