一家未直接开发AI技术 Thus, the company has achieved an annual revenue of $100 million.

The entity behind this remarkable business achievement is Arena, a platform that has become a focal point for major AI firms competing in the large model arena.

Arena originated as Chatbot Arena, initially launched in 2023 as an open-source research initiative by a team from the University of California, Berkeley.

It is surprising how rapidly this platform has evolved into a critical hub for assessing the capabilities of large language models.

In a significant milestone, Arena's commercial services, launched just eight months ago, have now reached an annualized revenue of $100 million.

ChatGPT and Claude Feature Prominently on the Large Model Arena Leaderboard

Many individuals are likely familiar with Arena.

Its most lauded feature is its large model leaderboard, compiled entirely through genuine, blind user evaluations.

The operational mechanism is exceptionally straightforward yet inherently competitive.

Users input a prompt, and the system anonymously presents responses from two different models. Participants then select the superior response.

This platform aggregates millions of such votes to generate an Elo-style ranking system.

This "arena" format has transformed Arena into a vital resource for AI enthusiasts and developers globally.

Currently, the platform has accumulated over 10 million user evaluations, 700 million conversations, and 82 million votes. It attracts over 10 million monthly visitors from more than 150 countries.

Crucially, approximately 80% of daily user queries are novel, meaning no model can pre-memorize answers.

The significance of these rankings is substantial.

Leading AI companies such as OpenAI, Google, Anthropic, and Meta, who are fierce competitors, all submit their flagship models to Arena for community assessment.

OpenAI even discreetly tested a model under the codename "summit" on Arena prior to the official release of GPT-5.

In essence, the most advanced models in Silicon Valley await validation from a project initiated by Berkeley students.

How Did Arena Generate $100 Million in Revenue?

The pertinent question arises: how has a free ranking platform become such a significant revenue generator?

Last September, Arena introduced a commercial service called AI Evaluations.

This service allows model developers and large corporations to pay Arena to leverage its community of tens of millions of users for in-depth evaluations of their models. This provides performance analyses grounded in "real-world" scenarios that are unattainable through benchmark scores alone.

This system functions as a "real-world CI/CD system."

When a model is nearing public release, Arena conducts a free evaluation for the community.

However, companies seeking to understand their model's strengths, weaknesses, and potential inaccuracies with real users must pay for this service.

This business model mirrors the "picks and shovels" approach during a gold rush: while gold miners may not always profit, those selling essential supplies invariably do.

As large model developers engage in increasingly intense competition and strive for incremental performance gains, their demand for such "post-deployment optimization" services grows.

Arena is strategically positioned to serve this essential need.

Three Berkeley Individuals Created the Most Lucrative AI Venture

Arena was formerly known as Chatbot Arena.

Prior to that, it was part of the well-regarded LMSYS research group at UC Berkeley.

Two roommates from Berkeley embarked on a straightforward mission: to establish a neutral platform for large language models to compete fairly.

The rapid trajectory of this student project into a unicorn company was unforeseen.

The timeline has been astonishingly swift: in the spring of 2025, the project transitioned from academia to become a formal company, securing $100 million in seed funding within weeks at a $600 million valuation.

Several months after the commercial product launch, Arena achieved an annualized revenue of $30 million within just four months.

Subsequently, in January of this year, a $150 million Series A funding round, led by Felicis and UC Investments, closed, valuing the company at $1.7 billion post-investment.

The three principal figures leading this venture are all accomplished individuals.

CEO Anastasios Angelopoulos possesses a strong mathematical background.

During his undergraduate studies in Electrical Engineering at Stanford, he was mentored by Stephen Boyd, a renowned figure in convex optimization.

At Berkeley for his doctoral studies, his advisors were Michael I. Jordan, a luminary in machine learning, and Jitendra Malik, a pioneer in computer vision.

His research has primarily focused on developing mathematically rigorous methods for evaluating black-box models.

CTO Wei-Lin Chiang is a recognized name in the open-source community, having developed the widely popular open-source chatbot, Vicuna.

He pursued his Ph.D. at Berkeley under the guidance of Ion Stoica, specializing in distributed systems, and previously held positions at Google, Amazon, and Microsoft.

Upon the public beta release of ChatGPT in late 2022, he redirected his research efforts entirely towards Arena.

His dedication to this project was described by his colleague Angelopoulos as "a labor of love."

Their intense commitment to the project led them to share living quarters, enabling them to collaborate more closely. These two roommates have built a company valued at $1.7 billion.

The third co-founder is the esteemed Berkeley professor and co-founder of Databricks, Ion Stoica, who served as an advisor prior to the company's formal establishment in April 2025.

Being the Judge is More Important Than Being the Competitor

Arena's latest development is the introduction of Agent Mode.

The platform's evaluation scope has expanded beyond simply determining who is a better conversationalist to assessing the actual tasks millions of users are accomplishing with agents: writing code, debugging, conducting research, and analyzing documents—complex, multi-turn tasks often involving hundreds of tool calls.

Arena now employs objective metrics such as task completion rate and hallucination rate for scoring, moving beyond its initial focus on human preference voting.

Artificial intelligence is transitioning from being mere "chatbots" to becoming capable "agents" that can independently handle tasks, with increasingly complex and high-stakes objectives.

Evaluation serves as the ultimate probe into the inner workings of AI.

The success and valuation of Arena, reaching $100 million and $1.7 billion respectively, are fundamentally based on the premise that this function will become increasingly crucial and valuable.

However, all participants ultimately face the same fundamental question: as machines begin to set their own challenges, who will possess the authority to grade their performance?