TMTPost: The End of Huge Profits from Reselling Tokens—How Many Ways Can AI Relay Stations Survive?

This year's "Double 11" shopping carts may need to add one more item: Tokens.

Recently, major model companies such as Alibaba Cloud, DeepSeek, Zhipu, MiniMax, and Kimi have successively plunged into e-commerce platforms, and AI subscriptions and Token top-ups have become new categories for "Double 11." Tokens pushed onto e-commerce shelves are beginning to be packaged and promoted like mobile phone data plans. It is reported that Tmall is still communicating with vendors and plans to design bundled packages around needs such as learning, family, and one-person studios.

Today, large models have become common productivity tools, and there is no shortage of paying users among both enterprises and individuals. Behind this consumption boom, the business around it is reaching a fork in the road: consumer demand is growing, but not all businesses selling Tokens can continue to make money.

Discounts and windfall profits, happening on the same screen

Over the past few years, from aggregation platforms connecting hundreds of models to independent vendors providing inference services, and to small API relay stations that attract users with low prices, all have stood between model supply and user demand and taken different paths to making money.

Overseas, OpenRouter is a representative among model aggregation platforms, with a gross margin of 71%. In August this year, payments giant Stripe announced that it had agreed to acquire OpenRouter. The two sides did not disclose the transaction price, but media reports said the deal was worth more than $8 billion. In its acquisition announcement, Stripe said OpenRouter connects more than 400 models and more than 80 suppliers.

And the value it was recognized for is not just helping users buy Tokens, but helping developers choose among model capabilities, price, speed, and reliability according to task requirements. For enterprises, there is no need to build a separate access system for each model; instead, they can manage calls through a unified platform and dynamically select models according to specific tasks.

Domestic player "SiliconFlow" is another model worth observing. According to its public business introduction, SiliconFlow provides products such as serverless Token services, dedicated instances, inference acceleration, and private deployment, connecting different models, computing resources, and system software so that models can run on GPUs and AI chips from different vendors, and then providing inference capabilities to developers and enterprises in the form of Token services or other technical solutions.

As of June this year, SiliconFlow's platform had more than 10 million registered users and more than 13,000 enterprise customers, with average daily Token throughput of about 578.5 billion times and a single-day peak exceeding 1.07 trillion times. SiliconFlow has now completed seven rounds of financing, is valued at 7.74 billion yuan, and has submitted a listing application to the Hong Kong Stock Exchange.

It is worth noting that in the large-model industry chain, "relay stations" are not a business category with clear boundaries. There is also a group of small API relay stations that, relying on model interfaces, payment convenience, and price differences, have quickly gathered a group of users. Such relay stations provide prepaid top-ups, unified interfaces, or packaged quotas, making it more convenient for users to call overseas models while offering highly attractive prices.

Tech Planet found that the API quotes given by some relay stations on the market are almost directly "slashed to the bone" compared with official standard prices. Taking Claude Fable 5.1 as an example, the official standard API charges $10 per million input Tokens and $50 per million output Tokens. On the quote page of a certain relay station, the model of the same name costs only $1.5 per million input Tokens and only $7.5 for output, roughly 15% of the official price.

Image source: quote page of a certain API relay station.

Even dedicated price-comparison websites have emerged around relay stations, helping users compare quotes, actual exchange rates, and top-up costs across different platforms. On the consumer side, scrolling through screens full of "factory discounts," it is hard not to be tempted.

At the same time, Tech Planet learned that in addition to directly selling API quotas, some relay stations have launched another model closer to account trading: selling finished accounts bundled with subscription packages. For example, a certain platform sells old Claude accounts with Pro subscription packages for 150 yuan per month and promises guarantees for the account's first login and subscription service, but cannot guarantee subsequent account-ban risks.

"There are too many Claude account bans in China," said the operator of that platform. This risk does not come out of thin air. Transparency data disclosed by Anthropic shows that from January to June 2026, the company banned a total of 11.4 million accounts; during the same period, it received about 398,000 appeals, of which about 42,000 appeal outcomes were overturned.

In Token windfall profits, there is a knife hidden

"Pie doesn't fall from the sky" is a consensus many API relay station users understand without saying.

As for low-price resources, most relay stations attribute their sources to a wide variety of upstream channels: some rely on procurement scale to obtain discounts, some use preferential quotas issued by model vendors, and some buy subscription packages and then split the call quotas within them for sale, thereby earning the price difference. These publicly disclosed upstream channels are already considered "legitimate" routes; what truly deserves vigilance may be the model behind the low price itself.

At API relay stations, the model names displayed by the platform are filled in by the platform itself. In other words, the model actually called in the backend may not be the real "full-blooded" model. Many users have posted that when actually using certain relay stations' models for complex tasks, performance often shows an obvious gap.

This doubt is not entirely a subjective guess on the part of users. In March this year, the research team at CISPA Helmholtz Center for Information Security in Germany published the paper "Real Money, Fake Models: Deceptive Model Claims in Shadow APIs," conducting a systematic audit of third-party "shadow APIs." The results showed that the outputs of some services differed significantly from the official APIs, with a maximum performance gap of 47.21%; in model fingerprint tests, 45.83% of the tests failed identity verification.

And within the user circles of API relay stations, the situation of "swapping" models has long been an open secret. As a result, model detection tools have emerged on the market, along with various specialized prompts for identifying watered-down models at relay stations. Before topping up a relay station's service, many users choose to first test the interface using relay station review websites.

Image source: a relay station review website.

"The waters at relay stations run deep; you have to find one you trust," one user shared from their own experience: after discovering obvious problems with the performance of the model provided by a certain relay station, they did not receive a proper resolution. The platform operator had initially said a refund was possible, but after the user provided their email, their account was simply deleted.

In addition, behind the low prices there are also compliance and data security risks. Since April this year, regulatory authorities have begun frequently issuing related risk alerts. On September 15, in the typical enforcement cases announced by the Cyberspace Administration of China, API relay stations were publicly named for the first time. Two websites operated by a Jiangsu technology company used the "API relay station" approach to call the interfaces of multiple large model products and provide chat and Q&A services externally. An investigation by the cyberspace authorities found that the company had not conducted security assessments as required and that content security risks existed; it subsequently ordered the company to rectify in accordance with the law and dealt strictly with the responsible persons, issuing them warnings.

Even domestic models have been drawn into this controversy. Stars 404, a partner at hvoy.ai, recently shared publicly that among the relay services they had observed, the "watering down" problem was more prominent with domestic models, and specifically named Kimi and GLM as "both heavily affected areas."

Today, domestic models are also becoming an important choice for developers worldwide. OpenRouter's weekly data shows the weight of domestic models in the global call market. In the week from October 5 to 11, 2026, the world consumed 172 trillion tokens, of which Chinese models accounted for 86.74 trillion, more than 50%.

The greater the demand, the more the intermediary services around model calls should have a market. Yet even with compliant business operations and greater technical investment, making money is not guaranteed. For example, SiliconFlow, a domestic "Token factory" that not only distributes but also adds inference optimization, achieved revenue of 55.33 million yuan in 2025, with cost of sales reaching 68.632 million yuan, an overall gross margin of -24%, and a full-year net loss of approximately 345 million yuan.

Large models are hot, but Token is not a business that makes money as long as there is traffic.

The business of the AI middle layer is getting harder to do, and the Token market has also ushered in new paths and opportunities.

For example, Nexus, launched by the AI inference platform Fireworks, focuses on helping customers choose suitable models according to task requirements through methods such as automatic model routing, in order to reduce AI programming costs. Related promotional materials claim that costs can be reduced by 50% to 75% at most, with the specific effect depending on task type, model selection, and evaluation criteria. What it targets is the efficiency space in the model invocation process.

Demand for Tokens is still growing, but as more and more players begin selling Tokens, what will truly determine who can stay in the game is who can, beyond price, continuously provide value that users are willing to pay for.