Stripe

3 billion USD at that time. 82 days later, it was acquired for 8 billion USD. The buyer is Stripe, one of the world's largest payment companies, whose core business is handling payment collection for enterprises. Let's first introduce what this company does. OpenRouter does not develop AI models, purchase GPUs, or build self-operated data centers. Its full team of 90 people only focuses on one single thing: packaging the APIs of more than 400 global models into one unified interface. 5% platform fee for each call request. Think about the old days: if you wanted to build an AI application, you had to register accounts for OpenAI, Anthropic, even Google, manage dozens of different keys, and handle separate billing processes.

Now you only need one key and one single bill to get everything done. 5% fee mean? 5 USD is the toll for using the platform service. This revenue is generated in a completely compliant way. It does not record the content of your queries, sells models at the exact same official price set by model vendors, and charges the platform fee separately. For the same open-source model that is supported by multiple service providers, you can fully choose the one with lower price and faster response speed as you like. I counted that they have integrated more than 80 suppliers in total, covering more than 400 different models.

Model vendors are also willing to cooperate with this platform, after all, the existing mature channels are ready to use, and the distribution service is free, so listing their models on OpenRouter adds an extra path to consume their model Tokens. With this business model, can the accounts break even and make profits? Industry institution The Information estimated in July this year that its annualized revenue reached 140 million USD, with a gross margin of 71%. It does not need to stock any physical goods, and its main cost is almost only the server expense.

5 million USD. Calculated based on the 8 billion USD acquisition deal, the per capita value of the company is nearly 100 million USD. Its traffic scale is also very large: 10 million developers and enterprise customers generate 10 trillion Tokens of call volume through it every day, and well-known companies including NVIDIA and Zoom are its clients. I checked its background, and wondered what makes it worth so much? Some people say it is because of the large number of supported models, others say it is because of the unified interface. Both points are correct, but neither hits the core.

A middleman that does not produce any products, its only core product is trust. I put forward a formula: Trust = Transparency × Certainty. Transparency refers to clear pricing, open accounts, and accessible call logs; certainty refers to direct supply from official vendors, automatic fault switching, and support for using your own private keys. OpenRouter has achieved full marks in both two dimensions. Therefore, this toll-collecting business is worth 8 billion USD. However, for the same toll-collecting business, some people make profits in a compliant way, while others do not.

Where is the difference? The first point is the capital source and operation compliance. There are three major barriers for domestic users who want to use overseas large models: poor network connectivity, inconvenient payment methods, and unaffordable price. Official platforms do not support domestic registration, and payment requires overseas credit cards. As long as there are people who want to cross these barriers, this business will exist. Since formal channels cannot provide low prices, some people get model resources from account pools. I specially investigated this operation mode: Spend 200 USD to subscribe to an overseas account, split the account quota and sell it on a pay-as-you-go basis, serve multiple users with one single account, and the monthly subscription cost of 200 USD can generate 2400 USD of revenue.

How is this splitting operation realized? The subscription account is originally designed for web page use. The operator uses reverse engineering to extract the web chat interface, and transforms it into an API that conforms to the OpenAI format. When you call the operator's API, the operator will log in to the web page on your behalf to send the query, and the same account is shared by dozens of users in rotation every day. The profit comes from this price difference: the cost is the fixed subscription fee, and the revenue is charged on a pay-as-you-go basis, the gap between them is the profit.

This is the most legitimate way to get model resources in this gray area. Going further down the gray path, some people use credit cards obtained through fraud, free quota obtained from batch registration wool grabbing, and capabilities extracted from reverse cracking of web interfaces. They pick the cheapest resources to use, and switch to other paths once the current one is blocked. The threshold for getting model resources is not high, and the threshold for opening such a service site is even lower. You only need 2000 to 5000 RMB, one server, and one set of open-source forwarding programs to build a service site.

Some researchers investigated the background of these sites: 11 out of 17 sites use the same open-source framework, and only 1 of the 17 sites has completed the official ICP filing. This research is completed by the Helmholtz Center for Information Security in Germany. When I checked the materials, I also found a real story reported by China Newsweek: A worker who worked in a factory workshop built a transit site according to online tutorials, and his business was so good that he had a daily revenue of thousands of RMB; he did not even know the meaning of the term "vibe coding" until he entered this industry.

Buyers are attracted by the low price: a developer paid 25 RMB for a monthly subscription, the platform promised to give 120 USD of call quota every day. After 20 days of use, the platform said the ban became stricter and the cost rose, so the price increased by 3 to 4 times. Soon the platform shut down directly, the 25 RMB of cheap benefit he got was eventually lost completely. The second point is the compliance of service content. You may think you are calling GPT, but the model running on the backend may actually be a small open-source model with lower performance.

83% of the sampled sites did not pass the model fingerprint verification, nearly half of the sites were secretly replacing the model you requested with other lower-level models. They charge the price of flagship models, but run low-performance models. 95%. What is even more horrifying is that some transit sites will tamper with the content in your returned responses. The research team from the University of California, Santa Barbara collected 428 APIs from Taobao, Xianyu and public communities for testing. 9 of them secretly injected malicious content into the returned code, and 1 of them directly stole the test cryptocurrency wallet.

Why can't this risk be prevented? Because you set the middleman as the access interface yourself, the middleman is designed to unpack all your requests and check them, so tampering with the content does not require any cracking. In other words, your agent is controlled by other people. There is also the risk of site running away. The life cycle of small transit sites is 1 to 3 months, they will harvest a batch of recharging users then disappear, change a new domain name and reopen the business. According to my observation: More than 70% of the transit sites can not be accessed after half a year.

There is a site operator in Shanghai whose online name is Guapi, he was detained for 37 days for operating this kind of business and obtained bail pending trial. He said in his site closure statement that he will definitely be sentenced in the future. This business also has a reverse operation: domestic models are cheaper, some people package domestic models and sell them to overseas users to make profits from both sides. In the final analysis, the low price is only the marked price, and the hidden cost lies elsewhere. The two dimensions of transparency and certainty are both zero, zero multiplied by zero equals zero, this business is negative from the first day of its operation.

But one phenomenon needs to be clarified: not all operators who get model resources from account pools, fraud and reverse cracking end up with bad results. The operators who build their own model operation facilities have a completely different business logic. Silicon Flow is using this answer to knock on the door of the Hong Kong Stock Exchange. Its gross margin is -24%, the gross margin of public cloud business is -119%, and it lost 345 million RMB in one year. 74 billion RMB. It is running at a loss, why is it valued so high? I think it is absolutely necessary to understand the background first, after all, the background contains the core driving factors.

The founder is Yuan Jinhui, 45 years old, PhD from Tsinghua University, studied under academician Zhang Bo of the Chinese Academy of Sciences. He only did one thing in the first half of his life: make AI models run efficiently. In 2014, he worked at Microsoft Research Asia and developed LightLDA, the world's fastest large-scale topic model training system at that time. In 2017, he left to start a business, planning to develop a deep learning framework that competes with TensorFlow, called OneFlow, and spent 6 years on this project. Six years later, his framework did not occupy the expected market share, then he found another opportunity: there are so many people who can make AI models run, but the people who really make money are those who package and sell computing power.

In August 2023, Silicon Flow was established. What does this company do? It rents computing power resources by itself, develops self-owned inference engine, adapts NVIDIA GPUs, Huawei Ascend chips, and Metaox chips, then installs open-source models on these hardware, and sells services by Token volume. Downstream developers do not need to touch any hardware, they can call DeepSeek, Qwen, GLM models after recharging their accounts. Is its financial performance good? Not good. 2 RMB of cost. 33 million RMB. The average paying user only spends 77 RMB a year, which is the price of a cup of milk tea, to use AI services for a whole year.

For the same Token selling business, OpenRouter has a gross margin of 71%. One business makes more money as it operates, the other loses more money as it sells more. But the capital market does not think it is a bad business. It applies for listing through the Hong Kong Stock Exchange Chapter 18C special technology channel, which is a dedicated channel for unprofitable tech companies. The -P suffix after the company name means that the company is still not making profits. 74 billion RMB, increasing by more than 30 times in three years. Its shareholder list includes Alibaba, Huawei Hubble, Meituan, SenseTime, Zhipu AI, China Unicom, JinkoSolar, all are leading giants in the industrial chain.

Why do these giants invest in a loss-making company? Because it occupies a very special ecological niche. Large model vendors are unwilling to deploy their own inference services on the cloud platforms of other large companies that have self-developed models, for fear that their call data will be used to train competing products. Silicon Flow does not develop any models itself, so it has no position conflict. It has even been listed on the OpenRouter platform as one of the underlying service providers, and it also connects to overseas service providers including Together AI and Fireworks AI to expand overseas markets.

A neutral platform that does not stand on any side is the only third-party computing power platform that all giants can use with full confidence.