封面：華強北夜市茶攤，年輕開發者與年長商人談論 Token 中轉站Cover: Huaqiangbei night market tea stall, young developer and older businessman discussing token relay Image AI-assisted (Z-Image-Turbo)
Hany Zheng Ge · September 25, 2026 · Futian
1. Prelude: An Old News Item That Sent a Chill Down My Spine
This morning, a friend tossed an old news item my way: "Monthly flow over ten million — what kind of business is a token relay station?"
Though called "old news," it was actually from a June investigation by 21st Economic Report and Jiemian News, resurfaced in September. Why the sudden buzz? Because last month Anthropic locked down regions — Fable 5 and Mythos 5 no longer available to overseas users. Overnight, countless token relay stations ran dry, and the whole circle realized: we've been living off other people's water pipes.
Reading it, my first reaction wasn't surprise — it was familiarity. This business? I've seen it before in thirty years of electronics trading — except back then it was called "grey market," now it's "token relay station."
2. First Act: Tea at Huaqiangbei, a Young Tells Me "Don't Buy Official"
Last week at Huaqiangbei, a young developer named A-Yuan invited me for tea. He said, "Uncle Zheng, you're still buying the official API? It's so expensive — let me introduce you to a token relay station. Claude Opus official is 170 bucks per million tokens; they sell it for 78, small stalls go as low as 20-30% of official price, the cheapest is just 2 bucks — almost free."
I asked, "If it's that cheap, how do they make a profit?" A-Yuan laughed, "Wholesale, of course." Token relay stations buy Coding Plans in bulk, amortizing the cost to 0.2 bucks per 10K tokens, while the official price is 1 buck; they sell at 0.5 bucks, a 60% gross margin. A $200/month Claude Max account gets split among a dozen or twenty users, pulling in two to three thousand bucks, peaking at four to five thousand — 25x leverage.
I said, "Isn't this just like the old days of tearing apart genuine devices to sell spare parts?" He replied, "Exactly, Uncle Zheng — where there's a wall, there's a smuggler. Overseas AI models aren't directly available in mainland China, and they're pricey, so smugglers naturally appear."
That night, lying in bed, I reflected: "Where there's a wall, there's a smuggler" — truly timeless.
3. Second Act: Three Tiers — the More You Earn, the Greyer It Gets
The report lays it out clearly: token relay stations come in three tiers.
The first tier is the legitimate players — think US OpenRouter — buying in bulk for official discounts, then reselling retail. Thin margins, but clean, sustainable long-term.
The second tier is the grey zone — and this is where most people operate today. Tactics are varied: registering shell companies to grab free credits, buying multiple Coding Plans to build pools, reverse-engineering Windsurf or Cursor's internal calls, tapping into public water pipes with private lines. As one station chief put it, "It's just like stealing water or stealing electricity." Selling at 20%-50% of official price, the profit hinges on the "grey" factor.
The third tier is the bottomless pit — selling fake goods. They claim to offer Claude, but the backend runs cheap open-source models. Tests show 45% of relay stations suffer "identity mismatch." There are also exit scams, and peers look down on them.
Having spent 26 years in foreign trade, I can instantly see: the three tiers correspond to three types of people. The first tier are in it for the long haul; the second are gambling on quick money; the third are in and out in one pot. Today's market? The second tier dominates — because Claude is simply too expensive. A single conversation burns five to six cents; ordinary users can't keep up, so cheaper alternatives naturally attract a crowd.
4. Third Act: The Greatest Harm Isn't the Price — It's "Killing Two Birds with One Stone"
A-Yuan told me about his friend A-K, one guy with two servers, charging over a million in monthly flow. How low is the technical bar? Open-source solutions like New API and sub2 API exist — buy a server, tell the AI to help deploy it, and even non-technical folks can get it running. Post an ad, and within 24 hours, eleven 200-person groups are flooded. They sell "empowerment courses" for 88 to 199 bucks, then sell deployment support to thousands of enterprises — selling the shovel is where the real profits lie.
But I told A-Yuan, "Young man, you see them making money, but do you see how they make it?" The report has a telling line: "For token relay stations, the baseline is no adulteration, no selling of data — but both are hard to uphold. If you don't count costs, someone will send you money — but few can stomach it."
What is "killing two birds with one stone"? One face uses a fake model to suck your token money; the other face packages your chats, code, and business data, selling it for 0.1 bucks per pop to data companies. A station chief openly admitted: people daily ask to buy user logs, specifically wanting Claude 4.6+ with 30+ rounds of conversation. You save a few dozen bucks, but the loss could be your entire business. Codebases get planted with trojans to steal your cloud credentials and encrypted assets — these aren't hypotheticals; they're happening now.
I told A-Yuan, "When you write code, guard those core passwords with your life — never let them pass through a token relay station." He replied, "I know — important stuff always runs on self-built domestic models." I said, "That's old-gangster wisdom."
5. Fourth Act: The Wind Has Already Shifted
Why is this business doomed? Three signals, each stronger than the last.
First, the authorities clamp down. Anthropic cut third-party harness in April, then on June 12th locked down regions entirely — the two strongest models no longer serve overseas users. Since token relay stations rely on Claude for about 90% of demand, losing it cuts the business in half.
Second, the heavies step in. In May 2026, Shanghai detained someone for three-odd weeks for illegally scraping and reselling low-price interfaces. On May 1st, Sun Yuchen launched B.AI, touting blockchain login and anonymous payments as "the flattest net," inviting questions about sidestepping identity chains; the same day, Fu Sheng rolled out EasyRouter, quickly renamed to "enterprise-grade model gateway," declaring no mainland services. Both big players rushed to white-wash — you tell me how deep this water really is.
Third, the domestic counter-offensive. In Q1's OpenRouter ranking, DeepSeek, Moon's Dark Side, MiniMax, Zhipu, and Qianwen sit at the top. Domestic models are both cheap and capable, and they're already exporting overseas. The old-station veterans led by Xia Da have already shifted tone: "The profiteering window is over; from now on, it's domestic token exports and slow, steady flow."
Zheng Ge's Three Points
First: Smuggler businesses will always exist, but lasting wealth must be clean money.
Back then, shanzhai phones, smuggled goods, and water-grey wares were all born from explosive demand and lagging regulation. The gaps will always be there, but getting rich off them? Sooner or later, you'll have to pay it back.
Second: Information gaps expire.
Today you profit from others' ignorance or inability to reach the official channel; tomorrow, the official locks the door or drops the price — and you're left with nothing. The true moat isn't whether your supplies are grey or clean; it's whether your service is steady and your data is guarded.
Third: As API sellers, we must walk the straight path.
I myself sell compute and sell APIs — this piece is, in part, written for myself. Legitimate bulk procurement, private enterprise deployment, and domestic-model overseas export — three roads. Margins are thin, but you can sleep soundly. Relying on reverse engineering and black cards to hold the line? When the hype passes, the first to die is whoever leaned on those tricks.
One-Line Summary
Token relay stations are the grey-market arbitrage spawned by the explosion of AI compute power and lagging regulation — profit rests on information asymmetry and capital sedimentation, but the cost is exposed user data and legal risk for operators. Smugglers will always be around, but if you want to last, you must run clean.
📞 Get in touch
WhatsApp (HK): +852 6641 7912
WhatsApp (US): +1 (562) 550-9968
Email: [email protected]
HK: 183 Flower Market St, Mong Kok
US: ALEXBI INC, Long Beach, CA
Qishanmen Tech · Star Compute
鄭哥，90年代大學畢業，26年橫跨電子、家電、服裝、珠寶、教育、AI算力多個行業。做過中學教師、翻譯、外貿，現在搞七善門科技和星辰算力。口頭禪：「點解唔得？」Hany, university grad from the 90s, 26 years across electronics, appliances, fashion, jewelry, education, and AI compute. Former teacher, translator, trader. Now running Qishanmen Tech and Star Compute. Catchphrase: "Why not?"
