50 employees, with a valuation of 47.1 billion yuan

A new wealth creation myth has emerged in the AI circle.

On August 16, multiple media reports stated that U.S. payment giant Stripe has reached a final agreement to acquire OpenRouter, with a transaction valuation exceeding 7 billion U.S. dollars (approximately 47.1 billion yuan). If the transaction is successfully completed, it will become one of the largest mergers and acquisitions in the AI infrastructure sector in 2026.

OpenRouter is a household name in the AI circle, and it is the world's largest large model aggregation platform. Whenever an open-source model is released, it almost always takes "No.1 in global token usage" on OpenRouter as a promotion highlight. While large model companies are competing for higher benchmark scores and lower prices, OpenRouter, the "water seller" in the AI industry, has already reaped huge profits: a "small company" founded in 2023 with only about 50 employees so far has been sold at a sky-high price of 7 billion U.S. dollars; VCs that invested in it just three months ago have already obtained an investment return of more than 5 times.

The "intermediary" that global large models cannot bypass

On November 30, 2022, OpenAI officially launched ChatGPT, which quickly became popular worldwide. Only one month later, the number of monthly active users of ChatGPT exceeded 100 million, making it the fastest-growing consumer app in history. Pandora's box was opened, and the world entered the AI era from then on.

At almost the same time, a far less well-known company also came into the world, that is OpenRouter.

Alex Atallah, the main founder of OpenRouter, used to work in Web3, and he was the co-founder and CTO of OpenSea, the world's largest NFT trading platform. The business of OpenSea is aggregated token trading on the blockchain, no matter which chain or wallet it is, users can complete one-click trading on the OpenSea platform.

Seeing the explosive popularity of large models, Alex Atallah keenly realized that large models might also need a "token aggregated trading" platform. Although the token of large models is not the same as the token on the blockchain, they have one thing in common: the standards on the supply side are very chaotic. Every model manufacturer has its own API format, billing rules and traffic limiting strategies. If users want to access multiple models at the same time, the integration work will drive people crazy.

Therefore, in early 2023, Alex Atallah officially founded OpenRouter with two old partners. Its business model is extremely simple: it is an "AI model aggregator". Users only need to call a unified interface to access all mainstream large models in the world. No need to modify codes, no need to register accounts separately, no need to study the billing rules of each vendor, OpenRouter handles everything for you in the background.

However, in the first year of its establishment, OpenRouter was not very successful. At that time, OpenAI dominated the large model track. The standard-bearer of open-source large models was Meta's LLaMA, which was almost powerless to fight back against GPT. Under such circumstances, the "aggregation of AI models" became water without a source.

All OpenRouter could do was to be fully compatible with OpenAI's API format, so that developers who were already using OpenAI could seamlessly switch to OpenRouter. Although it had the selling point of "obtaining access to hundreds of models at the same time", these "hundreds of models" had almost no practical value compared with GPT. Therefore, for a long time, OpenRouter was only a niche tool in the developer circle, with daily calls hovering at the million level.

It was not until the second half of 2024, when the lineup of open-source models gradually grew stronger, that OpenRouter transformed into a key infrastructure of the global AI industry.

OpenRouter uses the indicator of "annualized inference expenditure" when raising financing to show the growth of its business. This indicator is obtained by multiplying the API transaction volume on the OpenRouter platform in the most recent month by 12. According to data from official press releases and investor disclosures, OpenRouter's annualized inference expenditure in October 2024 was 10 million U.S. dollars, and it rose to 19 million U.S. dollars in December. After entering 2025, the growth further accelerated, reaching 100 million U.S. dollars by May 2025. By March 2026, the figure estimated by third-party reports was 1 billion U.S. dollars.

Such a steep growth curve is the core reason why OpenRouter can obtain a valuation of 7 billion U.S. dollars.

Nowadays, OpenRouter has become the reasoning computing power traffic hub of global large models, and can be called the "water seller" with the most comfortable ecological niche in the AI industry. Not only open-source models, but even closed-source giants such as OpenAI and Anthropic now connect their own models to the OpenRouter platform.

OpenRouter's profit model is very simple, which is to charge users a service fee, and the currently public rate is 5.5%. That means for every 100 U.S. dollars of tokens users consume on the OpenRouter platform, they need to pay 5.5 U.S. dollars. The reason why users are willing to pay this premium is that they can easily call almost all models on the market here.

Therefore, the more intense the competition in the large model track, the better the business of OpenRouter. And this business hardly requires capital investment.

Nowadays, OpenRouter processes 25 trillion tokens per week, with more than 400 large models on the shelves and more than 8 million registered developers. How many people are needed to maintain such a large platform? Menlo Ventures, an investor of OpenRouter, stated in an article at the end of May this year that OpenRouter is a team of about 50 people, praising its extremely high per capita efficiency. In addition, OpenRouter's official page on LinkedIn also marks the number of employees as 11-50. Therefore, around 50 employees should be an accurate number.

50 employees with a valuation of 7 billion U.S. dollars, that is 140 million U.S. dollars per capita. In contrast, Zhipu currently has a market value of 560 billion Hong Kong dollars, with about 1000 employees, equivalent to 560 million Hong Kong dollars (about 71 million U.S. dollars) per capita. OpenAI's valuation in the financing in March this year was 852 billion U.S. dollars, and reports say that the number of employees is expected to reach 8000 by the end of the year, with a per capita value of just over 100 million U.S. dollars. A token porter that does not produce large models has a per capita efficiency exceeding that of the top large model companies.

Chinese models have dominated the ranking list

Of course, OpenRouter can develop so well, and Chinese large models have made extremely great contributions.

In early 2025, Andrej Karpathy, the founding employee of OpenAI known as "K God", publicly stated that OpenRouter's model ranking list is "the most valuable real workload evaluation standard". He meant that the scores of large models can be influenced and manipulated, but the token usage on OpenRouter cannot be faked, because it reflects the real needs of users. This evaluation from Karpathy made OpenRouter's reputation in the open source community soar.

Interestingly, after entering 2026, OpenRouter's model ranking list has been completely dominated by Chinese models. Before February 2026, Chinese models accounted for less than 2% of the total token consumption on OpenRouter, belonging to marginal players. But starting from February, the situation was rewritten at an astonishing speed.

The first "game breaker" was MiniMax. In mid-February, the MiniMax M2.5 model topped OpenRouter's weekly call volume ranking for the first time, with a weekly call volume of 2.45 trillion tokens. This "first shot" proved one thing to the global market: Chinese models are not only usable, but also can be used on a large scale with huge throughput.

Shortly after that, Xiaomi entered the track.

In late March, a mysterious model named "Hunter Alpha" was launched on OpenRouter anonymously, quickly became popular in the community with its excellent programming capability, and topped the daily ranking for many consecutive days. When its identity was revealed, everyone was shocked — it was Xiaomi's self-developed MiMo V2 Pro.

When MiMo V2 Pro officially topped the weekly ranking, the data was quite shocking: the weekly call volume reached 3.28 trillion tokens, accounting for as high as 40% of the total weekly call volume, becoming the first model in OpenRouter's history to break through 3 trillion tokens in a single week. Lei Jun personally reposted the report from China Central Television and publicly praised it.

In April, the ranking list entered a stage where top players competed fiercely. Alibaba's Qwen3.6-Plus topped the weekly ranking, becoming the first model on OpenRouter to break through 1 trillion tokens in single-day call volume. Shortly afterwards, Tencent's Hunyuan Hy3 preview rushed to the first place on the daily ranking just a few days after its release. Then Kimi K2.6 from Moonshot AI won the weekly championship. Chinese models took turns to occupy the top position.

After that, DeepSeek V4 Flash dominated the ranking for six consecutive weeks starting from late May, showing amazing staying power. In late July, Xiaomi's MiMo V2.5 returned to the peak with an astonishing weekly data of 10.5 trillion tokens. In the first week of August, DeepSeek V4 Flash took back the first place with 8.83 trillion tokens.

The result after this round of fierce competition is that now Chinese models account for more than 60% of the total token consumption on OpenRouter, and the top 5 models in the world are all Chinese models.

That is to say, most of OpenRouter's current traffic is supported by Chinese models. Without the rise of Chinese models, OpenRouter would hardly have a valuation of 7 billion U.S. dollars today.

Objectively speaking, the ranking value of OpenRouter is not as great as many people imagine. The calls of top closed-source model giants such as OpenAI, Google, and Anthropic mainly come from their own interfaces. OpenAI's annualized revenue is said to have exceeded 40 billion U.S. dollars, while the total API revenue of all models on the OpenRouter platform is only 1 billion U.S. dollars, which is just a fraction of OpenAI's revenue. Therefore, topping the ranking on OpenRouter can never be interpreted as the model is the "most popular" in the world.

However, there is indeed no larger third-party large model calling platform than OpenRouter all over the world, so its ranking list still has reference value.

For Chinese large model companies, the OpenRouter platform is very important. Statistics show that 47% of OpenRouter's users come from the United States, and Chinese developers only account for 6%. That means the dominance of Chinese models on OpenRouter is mainly achieved by U.S. developers paying real money to use them. OpenRouter has become a key fulcrum for Chinese large models to go global.

Valuation surged 5 times in 3 months, VCs have earned huge profits

In February 2025, OpenRouter, which was in a stage of soaring traffic, began to arouse the interest of capital. In February 2025, a16z led OpenRouter's 12.5 million U.S. dollar seed round of financing. The valuation of this round of financing is unknown.

In June 2025, OpenRouter announced a 40 million U.S. dollar Series A financing with a valuation of 500 million U.S. dollars. The lead investor of this round was Menlo Ventures, and the follower was HSG.

Then came the Series B round in May 2026, with a financing amount of 113 million U.S. dollars, and the valuation continued to rise to 1.3 billion U.S. dollars. CapitalG, the independent growth fund under Google, led the investment, and a number of CVCs including Nvidia, Snowflake Ventures, and Databricks Ventures followed suit.

Only 3 months after the Series B financing, Stripe launched an acquisition at a valuation of more than 7 billion U.S. dollars. Series B investors have obtained a return of more than 5 times, and Series A investors have obtained a return of more than 10 times. The IPO processes of the two AI giants OpenAI and Anthropic have been delayed again and again, and it is OpenRouter, this "intermediary", that let VCs earn their first bucket of gold in the AI industry first.

The reason why Stripe is willing to offer a high price of 7 billion U.S. dollars is also related to its own anxiety. As one of the world's largest payment giants, Stripe is very worried that the form of commercial transactions will undergo fundamental changes in the AI era. Imagine that in the future, AI Agents will help you book flights, compare prices, select hotels and place orders, the payment scenarios will change dramatically.

Therefore, Stripe is very active in its AI layout. At the Stripe Sessions 2026 conference, Stripe intensively released commercial suites for AI Agents, including agent settlement protocols deeply integrated with OpenAI, Google Gemini, Microsoft and Meta, as well as Link Agent Delegated Wallet evolved from 250 million consumer-side wallets. John Collison, co-founder of Stripe, publicly stated that the traditional e-commerce model based on keyword search will eventually be subverted by "Agentic Commerce" driven by AI Agents.

Interestingly, Alex Atallah, the founder of OpenRouter, also appreciates Stripe very much. He once compared his company to Stripe in the AI field. The only difference is that Stripe settles funds, while OpenRouter settles tokens. As mentioned earlier, Alex Atallah used to work in Web3, and his understanding of tokens seems to have a certain flavor of "cryptocurrency tokens".

If you only look at the P/S multiple, the valuation offered by Stripe is quite exaggerated.

Its annualized revenue at the end of 2025 was only 19 million U.S. dollars, rose to about 50 million U.S. dollars at the time of Series B financing, and the latest data in July was 140 million U.S. dollars. Obviously, the valuation of 7 billion U.S. dollars implies extremely high growth expectations and strategic value. What kind of chemical reaction can be generated between AI infrastructure and payment infrastructure, the answer remains to be seen.

This article is from the WeChat official account "ChinaVenture", author: Tao Huidong, published with authorization from 36Kr.