Summary of Key Points
AI intermediaries serve as “bridge platforms” that connect ordinary users with overseas large models (such as Claude and GPT), with extremely low technical barriers (they can be set up in just one day). There are two types of intermediaries: legitimate and grey. Legitimate ones generate revenue through platform fees (e.g., OpenRouter), while grey ones exploit various tactics like subscription splitting and model substitution to amass huge profits. These intermediaries address the practical need of domestic users to access overseas models, but they also introduce risks such as compliance issues, data breaches, and the difficulty in distinguishing between genuine and counterfeit models, representing a “semitransparent infrastructure” in the AI era.
Detailed Analysis
1. What is an AI intermediary? – A “low-barrier AI delivery service”
Think of an intermediary as an “AI version of a courier service.” Normally, you would have to send requests directly to OpenAI/Claude to use their models, but domestic users cannot do so without an overseas phone number and credit card. The role of an intermediary is to receive your requests, verify the funds in your account, forward them to the respective models, and then return the results to you. Technically, setting up such a service is almost barrier-free: open-source tools like One API and New API can be used, and with some configuration adjustments, it can be up and running in just one day. As a result, intermediaries across the internet all have a similar format – unified interfaces, support for multiple models, payment in RMB, and prices starting from a certain discount. However, there are four main types of businesses operating within this space:
- Legitimate agents: Purchase models officially and resell them legally.
- Enterprise gateways: Enterprises buy models themselves, and intermediaries help allocate usage quotas internally.
- Regional price difference sellers: Buy models overseas and resell them to domestic users.
- Grey intermediaries: Use unauthorized accounts or substitute models to reduce costs (this is the most common practice).
2. How do intermediaries make money? – Profits come from “grey operations” and information asymmetry
Legitimate intermediaries earn relatively less (for example, OpenRouter generates over $50 million annually through a 5.5% platform fee), while grey ones are the main source of substantial profits:
- Subscription splitting: Buying large numbers of Claude/GPT subscription accounts (often obtained illegally, such as those sold on Taobao) and dividing the available usage into smaller “quota units” for sale. For instance, a Claude Max 5X subscription account might be sold at ten times the official API price, with the operator betting that the account won’t get banned; if it does, they simply purchase a new one.
- Model substitution: Selling cheaper domestic models (like GLM 5.1) as more expensive overseas models without users noticing the difference (given the significant price gap). This practice is less common now due to the lower resale value of Claude accounts.
- Distribution: Larger intermediaries sell models at lower prices to agents, who then resell them at a higher mark-up, turning from retailers into wholesalers.
- Information asymmetry: The currency used by intermediaries (e.g., points) is not real dollars; the exchange rates and fee structures are set by the operators. For example, charging 5 times more for a 1-yuan recharge could result in the same actual cost as charging only 0.2 times more, allowing operators to profit.
Here’s an example: A small operator might generate monthly revenues of 30,000–300,000 yuan, while larger ones can earn tens of millions with a team of less than 20 people, achieving gross margins of 50–60%. However, some intermediaries suffer losses; for instance, an operator in Shanghai was detained for illegally obtaining resources and still ended up in the red after making compensation.
3. What are the pitfalls of using AI intermediaries? – Three major risks
- Compliance risks: Using unauthorized accounts or illegal methods to access models can lead to legal consequences (as seen with the Shanghai operator).
- Data breach risks: All your requests, including code, keys, and prompts, pass through the intermediary’s servers, potentially allowing data to be retained and sold. Researchers have noted that while API fees are a source of revenue, logging and analyzing user activity generate true profits, as this data can be used for model training and further monetization.
- Difficulty in distinguishing between genuine and counterfeit models: Substituted models often perform significantly worse. For example, an interface labeled as Gemini 2.5 might only score 37 points in medical tests (compared to the official 83 points). Users are usually unaware of this, and even with plugins that allow model switching, they may not realize it has been done without their knowledge (for instance, the Cursor plugin automatically switches the model behind the scenes).
4. Why have AI intermediaries become so popular in China? – A need driven by practical constraints
- Inaccessibility of overseas models: OpenAI/Claude do not support domestic users, requiring an overseas phone number and credit card for registration, with increasingly strict KYC (know-your-customer) requirements (Anthropic even uses email to determine location). Intermediaries first solve the “usability” issue before focusing on offering lower prices.
- Surging demand for programming agents: Previously, models were used for occasional queries; now, programming agents require continuous consumption of tokens (e.g., for file reading and tool invocation), making subscription plans insufficient. Intermediaries provide a cheaper and more convenient solution.
- Modular operation: Account providers and technical solutions are readily available, and if an intermediary gets banned, they can simply switch to a new domain name with minimal effort.
5. The paradox of AI intermediaries: both infrastructure and a “black box”
AI intermediaries are used extensively (many developers rely on them), yet they lack the transparency of traditional infrastructure. You don’t know whether the models being used are genuine, whether data is being stolen, or whether the platform might suddenly shut down. This highlights an important issue: in an AI-driven world where models serve as a form of production capacity, those who can reliably distribute these services hold significant power. Currently, this role could be held by legitimate platforms like OpenRouter or small operators within WeChat groups. In the future, legitimate platforms may replace grey intermediaries, but for now, this industry is still in its wild growth phase.
Conclusion
AI intermediaries represent a “double-edged sword” of the AI era: they enable ordinary users to access advanced models, yet they also introduce risks related to compliance and security. As regulations and technology evolve, this semi-transparent market may gradually become more regulated. However, for the time being, they remain a necessary solution for many users.
(End of translation)