With an annual revenue of 100 million US dollars, two post-90s roommates from UC Berkeley have built the most profitable AI business

A company that doesn't develop AI has an annual revenue of $100 million!

What created this business myth is the "Large Model Arena" — Arena, which is being fiercely contested by Silicon Valley AI giants.

Its predecessor was called Chatbot Arena, which was initially just an open - source research project launched by the UC Berkeley team in 2023.

Who could have imagined that in just a short time, it would become the core position that controls the lifeblood of large models.

Just today, only eight months after the commercial service of Arena was launched, the annualized revenue reached $100 million, setting a new milestone.

ChatGPT and Claude Dominate the Ranking, the Large Model Arena

For many people, Arena is not unfamiliar.

What it's most well - known for is the large model ranking list that is completely based on real blind tests by users.

The gameplay is extremely simple yet full of a sense of competition —

Enter a prompt, and the system will blindly select two anonymous models to answer simultaneously; then choose which one is better.

The system aggregates tens of millions of such votes into an Elo - style ranking list.

It's this "arena" mechanism that has made it a holy land for global AI enthusiasts and developers.

To date, the platform has accumulated over 10 million user evaluations, 700 million conversations, 82 million votes, and more than 10 million monthly visitors from over 150 countries around the world.

More importantly, about 80% of users' daily questions are brand new, and no model can "memorize the questions" in advance.

How high is its credibility?

Top companies like OpenAI, Google, Anthropic, and Meta, which usually compete fiercely with each other, all send their flagship models to Arena to be tested by the community.

OpenAI even secretly put its model under the code name "summit" on the Arena for testing before the official release of GPT - 5.

In other words, the most powerful models in Silicon Valley are all waiting for a project by Berkeley students to give them the stamp of approval.

How did it achieve $100 million in revenue?

So the question is — how did a free ranking list become a money - making machine worth $100 million?

In September last year, Arena launched a commercial service called AI Evaluations:

Model manufacturers and large enterprises can pay to let Arena mobilize its community of tens of millions of people to conduct in - depth evaluations of their models and obtain "real - world" performance analyses that can't be obtained through simple benchmarking.

This is a "CI/CD system for the real world."

Once a model is ready for public release, Arena will evaluate it for the community for free;

But if an enterprise wants to know where its model is strong, where it is weak, and where it is spouting nonsense in the hands of real users, it has to pay.

This is a typical "water - seller" business — in a gold rush, gold diggers may not make money, but those who sell water and shovels are guaranteed to profit.

The more fiercely large model manufacturers compete and the more they try to squeeze out the last bit of performance, the greater their appetite for this "post - launch optimization" service.

And Arena happens to be in a position that everyone has to pass through.

Three Berkeley Students

Created the Most Profitable Business

Arena's predecessor was called Chatbot Arena.

Even earlier, it belonged to the well - known LMSYS research group at Berkeley.

Two Berkeley roommates just wanted to do a simple thing — build a neutral arena for large language models so that everyone could compete fairly.

No one expected that this student project would grow into a unicorn.

The timeline is so fast that it takes your breath away: In the spring of 2025, the project was separated from the university and a company was officially established. It secured $100 million in seed funding within a few weeks and was valued at $600 million;

A few months later, the commercial product was launched, and in just four months, the annualized revenue reached $30 million.

Then, in January this year, a $150 million Series A round led by Felicis and UC Investments was completed, and the post - investment valuation was fixed at $1.7 billion.

The three people at the helm are not unknown figures.

CEO Anastasios Angelopoulos is a mathematician at heart.

When he was an undergraduate in electrical engineering at Stanford, he studied under Stephen Boyd, a legend in the field of convex optimization.

When he went to Berkeley for his Ph.D., his supervisors were two godfather - level experts — machine learning master Michael I. Jordan and computer vision master Jitendra Malik.

The main direction of his research in recent years is how to make mathematically rigorous judgments on a black - box model.

CTO Wei - Lin Chiang is a familiar face in the open - source community — the popular open - source chatbot Vicuna was developed by him.

He was a Ph.D. student at Berkeley, studying under Ion Stoica and specializing in distributed systems. He has worked at Google, Amazon, and Microsoft before.

When ChatGPT started its public beta at the end of 2022, he stopped all his previous research and dived into Arena.

His obsession with this project was described by his partner Angelopoulos as "a labor of love."

For this project, the two worked such long hours that they simply moved in together. Two roommates built a company worth $1.7 billion.

The third co - founder is the well - known Berkeley professor and co - founder of Databricks, Ion Stoica. He served as an advisor before the project was incorporated into a company in April 2025.

Being a referee is more important than being a player

Arena's latest move is to launch Agent Mode.

What it evaluates is no longer just "who chats better," but the real tasks that millions of users are using agents for: writing code, debugging, doing research, analyzing documents — those long - term tasks that involve hundreds of tool calls and multiple rounds of interaction.

It starts to score using objective indicators such as task completion rate and hallucination rate, far beyond the initial scope of "human preference voting."

AI is evolving from a "chatbot" to an "agent" that can handle tasks independently. The tasks are getting longer and the stakes are getting higher.

Evaluation is the last probe that humans insert into the interior of AI.

The fact that Arena's business is worth $100 million and $1.7 billion is essentially a bet that this matter will become more and more important and more and more valuable.

But in the end, everyone has to answer the same question — when machines start to set their own questions, who is qualified to grade the papers?

Reference materials:

https://techcrunch.com/2026/06/29/arena-the-ai-leaderboard-everyone-uses-is-now-a-100m-business/

https://x.com/ml_angelopoulos/status/2071629882057228680?s=20

This article is from the WeChat official account "New Intelligence Yuan". Author: ASI Revelation, Editor: Taozi. Republished by 36Kr with permission.