(This article is authored by Tech Planet and published with authorization from Titanium Media.)
Article | Tech Planet, Author | Hua Wei
This year’s “Double 11” shopping cart may include one more item: Token.
Recently, major large-model companies such as Alibaba Cloud, DeepSeek, Zhipu, MiniMax, and Kimi have successively entered the e-commerce platform arena, with AI subscriptions and token top-ups emerging as new product categories for “Double 11.” Tokens, now featured on e-commerce shelves, are beginning to be bundled and promoted much like mobile data plans. Reportedly, Tmall is also in discussions with vendors, planning to design bundled packages tailored to needs such as learning, home use, and solo studio setups.
Today, large models have become commonplace productivity tools, and both enterprises and individuals alike form a growing paying user base. Behind this consumption boom, the business ecosystem surrounding them is reaching a crossroads: while consumer demand continues to grow, not every venture selling tokens can sustain profitability.
In addition to the consumer-facing interfaces actively cultivated by large-model vendors, the AI middleware layer that has emerged around these models is currently undergoing a major transformation.
Discounts and exorbitant profits occur on the same screen.
Over the past few years, entities ranging from aggregation platforms that connect hundreds of models to independent vendors offering inference services, and even small API gateways that attract users with low prices, have all positioned themselves between model supply and user demand, each forging its own distinct monetization strategy.
Overseas, OpenRouter is a leading model aggregation platform, with a gross margin of 71%. In August this year, payment giant Stripe announced that it had agreed to acquire OpenRouter. The two parties did not disclose the transaction price, but media reports indicate that the deal exceeded $8 billion. In its acquisition announcement, Stripe stated that OpenRouter connects more than 400 models and over 80 vendors.
And the value it’s recognized for goes beyond simply helping users acquire tokens—it enables developers to strike a balance among model capabilities, cost, speed, and reliability based on their specific task requirements. For enterprises, this means they no longer need to build separate integration systems for each model; instead, they can manage API calls through a unified platform and dynamically select models tailored to particular tasks.
Domestic player “Silicon Flow” represents another pattern worth watching. According to its public business overview, Silicon Flow offers serverless token services, dedicated instances, inference acceleration, and private‑deployment solutions, connecting diverse models, computing resources, and system software. This enables models to run on GPUs and AI chips from different vendors, with inference capabilities then delivered to developers and enterprises via token services or other technical approaches.
As of June this year, Silicon Flow Platform has over 10 million registered users and more than 13,000 enterprise customers, with an average daily token throughput of approximately 578.5 billion and a single-day peak exceeding 1.07 trillion. Today, Silicon Flow has completed seven rounds of financing, with a valuation of RMB 7.74 billion, and has submitted a listing application to the Hong Kong Stock Exchange.
It’s worth noting that, within the large‑model industry chain, “intermediary platforms” do not constitute a clearly defined business category. A number of smaller API intermediaries have rapidly attracted a user base by leveraging model APIs, convenient payment options, and price differentials. These platforms offer prepaid accounts, unified interfaces, or bundled quotas, making it easier for users to access overseas models while offering highly attractive pricing.
Tech Planet found that some API pricing offered by third-party hubs is almost a steep discount compared to the official standard rates. Take Claude Fable 5.1 as an example: the official standard API charges $10 per million input tokens and $50 per million output tokens. Meanwhile, on one particular hub’s pricing page, the same model is priced at just $1.5 per million input tokens and $7.5 per million output tokens—roughly 15% of the official rate.
Image source: A pricing page on an API gateway.
Even around transit hubs, specialized price-comparison websites have emerged, helping users compare quotes, actual exchange rates, and top-up costs across different platforms. From the consumer’s perspective, scrolling through screens filled with “official‑brand discounts,” it’s hard not to be tempted.
Meanwhile, Tech Planet has learned that, in addition to directly selling API quotas, some intermediaries have introduced another model that more closely resembles account trading: bundling verified accounts with subscription packages for sale. For example, one platform offers an old Claude account along with a Pro subscription package for 150 yuan per month, promising guarantees for the account’s initial login and subscription service, but providing no assurance against subsequent account bans.
“There are too many suspended Claude accounts in China,” said the platform operator. This risk did not arise out of nowhere. Transparency data disclosed by Anthropic shows that from January to June 2026, the company cumulatively banned 11.4 million accounts; during the same period, it received approximately 398,000 appeals, of which about 42,000 were overturned.
In the lucrative world of tokens, there lies a hidden danger.
“There’s no such thing as a free lunch,” is an unspoken consensus among many API gateway users.
For low-cost resources, most intermediaries attribute their sources to a wide array of upstream channels: some leverage purchasing volume to secure discounts, others take advantage of promotional credits offered by model providers, and still others buy subscription packages and then resell the available call quotas, pocketing the price difference. Yet these publicly known upstream channels are already considered “legitimate” avenues; what truly warrants vigilance may be the models themselves behind those low prices.
In the API gateway, the model names displayed on the platform are those you’ve entered yourself. In other words, the model actually invoked in the backend may not be the full‑featured, “fully loaded” version. Many users have reported that, when used for complex tasks, certain gateway models often show a noticeable performance gap.
This concern is not entirely based on users’ subjective speculation. In March this year, a research team from Germany’s CISPA Helmholtz Center for Information Security 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 output performance of some services differed significantly from that of the official APIs, with the largest performance discrepancy reaching 47.21%; in model fingerprint tests, 45.83% of the tests failed to pass authentication.
Meanwhile, within the user community of API relay services, the practice of “substituting” models has long been an open secret. As a result, model‑inspection tools have emerged, along with various specialized prompts designed to detect water‑leaking models used by these relays. Before subscribing to a recharge‑relay service, many users choose to first test the API using relay‑evaluation websites.
Image source: A certain transit station review website.
“The intermediary water is very deep; you have to find someone you can trust,” one user shared their experience. After discovering that the model provided by a certain intermediary station had obvious performance issues, they did not receive an appropriate resolution. The platform operator initially said a refund was possible, but after the user provided their email address, their account was simply deleted.
Moreover, low prices also come with compliance and data security risks. Since April this year, regulatory authorities have been issuing frequent risk warnings on the issue. On September 15, the Cyberspace Administration of China (CAC) released a list of typical enforcement cases, publicly naming API transit stations for the first time. Two websites operated by a Jiangsu-based technology company were found to be invoking interfaces of multiple large-scale model products through an “API transit station” setup, thereby providing conversational and Q&A services to external users. Upon inspection, the cyberspace regulator determined that the company had failed to conduct a required security assessment and posed content‑security risks. The authority subsequently ordered the company to make corrections in accordance with the law and imposed strict disciplinary measures, including warnings, on the individuals responsible.
Even domestic models have become embroiled in this wave of controversy. Recently, Stars 404, a partner at hvoy.ai, publicly shared that, based on his observations of intermediary services, the issue of “watering down” is even more pronounced among Chinese models, specifically naming Kimi and GLM as “severely affected areas.”
Today, domestically developed models are also becoming a key choice for developers worldwide. OpenRouter’s weekly data highlights the significant share of Chinese models in the global inference market. During the week of October 5–11, 2026, global token consumption totaled 172 trillion, with Chinese models accounting for 86.74 trillion—over 50%.
The greater the demand, the more market potential there should be for intermediary services centered around model invocation. However, even with business compliance and increased technological investment, profitability is far from guaranteed. For example, in China, “Token Factory,” which both distributes models and performs inference optimization, reported revenue of RMB 55.33 million in 2025, while its cost of sales reached RMB 68.632 million, resulting in an overall gross margin of -24% and a full-year net loss of approximately RMB 345 million.
Large models are all the rage, but tokenization isn’t a business that can make money simply by generating traffic.
How many viable options are left for AI hubs?
The AI middleware business is becoming increasingly challenging, and the token market has also found new avenues and opportunities.
Following the closure of a number of transit hubs, platforms seeking to transition to compliance have emerged in the market. Unlike the past, when they relied on opaque upstream channels and lured users with price differentials, these platforms aim to rebuild stable supply and service systems by obtaining official authorization from model providers. Accordingly, authorized procurement, technical services, and compliant operations may all entail additional costs, resulting in pricing that is higher than that of gray‑market transit channels.
Another approach is to integrate the token into a larger service ecosystem and sell it there.
Cloud providers such as Alibaba Cloud, Volcengine, and Tencent Cloud possess resource advantages that are difficult for intermediary platforms to replicate. Model invocation can be bundled with computing power, storage, data management, security capabilities, and enterprise applications, so customers are not just purchasing an API; they’re acquiring a comprehensive suite of services that can integrate with their existing workflows. For enterprises, while the per‑token price is certainly important, deployment costs, service stability, technical support, and data management also play critical roles in the final decision.
Operators, meanwhile, are trying to turn Tokens into another kind of “data package.” Recently, the three major carriers have successively launched Token‑related plans, packaging model‑calling allowances as monthly‑subscription products similar to mobile data plans, and seeking to lower the barrier to entry through methods such as payment via phone bills and cross‑platform usage, thereby covering domestic models like DeepSeek, Kimi, and Qwen.
Compared with the approach of having to independently locate transit stations, top up accounts, and configure interfaces, this product form is more akin to the communication services that ordinary users are familiar with, and it holds promise for expanding its market by leveraging operators’ existing user‑reach and billing systems. For operators, the challenge moving forward will be how to make it easier for users to find the entry point, understand the plan rules, and continue using the service in everyday scenarios.
However, there is still a gap between the launch of these plans and their widespread adoption. In May this year, China Telecom took the lead in rolling out a nationwide trial commercial Token plan at the group level, with a dedicated entry point in its official app where users can search for corresponding tiers using keywords. By contrast, China Mobile and China Unicom’s related services remain largely at the provincial pilot stage, with localized plans currently available only in a handful of regions, such as Shanghai, Beijing, and Hubei.
More notably, competition in the token business is shifting from “how much to sell” to “how to deliver higher-quality model intelligence.”
For example, Nexus, launched by the AI inference platform Fireworks, focuses on helping customers select the most suitable model based on their task requirements—through automated model routing and other methods—to reduce AI development costs. According to its marketing materials, costs can be cut by as much as 50% to 75%, with the actual savings depending on the task type, model selection, and evaluation criteria. Its target is to unlock efficiency gains in the model‑invocation process.
Demand for tokens continues to grow, but as more and more players start selling their tokens, the real deciding factor in who stays in the game will be who can deliver, beyond price, the value that users are willing to pay for.
For more exciting content, follow TMTPost on WeChat (ID: taimeiti), or download the TMTPost App.