Summary: TechCrunch reported on June 29 that Arena reached a $100 million annualized revenue run rate eight months after launching its first commercial product, driven by paid AI Evaluations for labs and enterprises. The article cautions that the figure reflects consumption rather than subscription revenue, and says Arena’s free leaderboard plus commercial analytics create a relationship worth watching.
TechCrunch dropped a figure on June 29 that hardly anyone saw coming: Arena, the company operating the crowdsourced AI leaderboard, now books a $100 million annualized revenue run rate — a mere eight months after its first commercial product shipped. The leaderboard most people recognize costs nothing: you type a prompt, two anonymous models answer side by side, and you vote for the better reply. Those evaluations top 10 million in aggregate, covering text, coding, vision, and image generation. CEO Anastasios Angelopoulos put it bluntly: “A lot of people don’t even understand that our business is making any money at all,” adding that “people still see us as an open source project.”
How does a leaderboard everyone uses at no charge grow into an operation collecting $100M a year? Anyone who follows model evaluation should walk through that transformation.
A Berkeley Research Project, Now Valued at $1.7B
In 2023, Arena was a research project inside UC Berkeley; the company incorporated in April 2025. Co-founders Anastasios Angelopoulos (CEO) and Wei-Lin Chiang (CTO) are both Berkeley postdocs, advised by Berkeley professor and Databricks co-founder Ion Stoica. The leaderboard runs on a simple exchange: users get their hands on new, frequently unreleased models ahead of the public, handing over the pairwise votes that give the rankings statistical weight.
Money arrived faster than recognition: in January 2026, Arena closed a $150 million Series A at a $1.7 billion post-money valuation, annualized revenue then standing at $30 million. Funding across all rounds totals $250 million from Felicis, Andreessen Horowitz, The House Fund, LDVP, Kleiner Perkins, Lightspeed, Laude Ventures, and UC Investments.
Where the Money Comes From: AI Evaluations
Revenue flows from AI Evaluations, introduced in September 2025 to sell in-depth performance analytics to model labs and enterprises. Its buyers are largely the same labs whose models occupy the public leaderboard, so the upsell nearly closes itself: those teams know their ranking — they are paying to understand why. Over the eight months after launch, the analytics line carried annualized revenue from $30 million to $100 million, better than a threefold increase.
One billing aspect deserves a careful look: Angelopoulos says charging is based on consumption, so the $100M annualized figure is not genuine recurring subscription revenue but current usage stretched across a year. In a sector where AI startups routinely dress up their ARR, that is an honest yet necessary footnote: if usage cools, the number will not hold the way subscription revenue does.
The Actual Competition: Data Labeling
Pressed on rivals, Angelopoulos does not soften it: over the data and evaluation work that feeds model post-training, Arena goes “for the same dollar” as Mercor, Surge, and Scale AI. That market keeps swelling. Handshake’s annualized AI training revenue climbed from $550 million to nearly $1 billion, while Mercor has cleared $1 billion — up from $500 million the previous September.
On the leaderboard front, Arena runs effectively unopposed: Yupp, the nearest comparable crowdsourced-evaluation startup, shut down in March 2026. Arena’s leaderboard is broadening too — a new Agent Mode scores long-running agent workflows instead of single-turn question answering, mirroring the industry’s move from chatbots toward tool-using agents working across minutes or hours. The expansion of those labeling-heavy rivals also signals demand for post-training and evaluation is far from saturated — the rising tide behind that $100M figure.
Clout, and the Caveats Attached
Arena’s clout rests on more than 10 million crowd votes, and many voters came for the chance to try new and often unreleased models first. The same arrangement is at once a moat and a built-in conflict of interest: labs buy the deep analytics while counting on the leaderboard for visibility and credibility. Now that “the leaderboard everyone uses” is a $100M operation, expect the boundary between neutral evaluation and commercial interest to be picked apart with each model release.
Model teams should walk away with two things. First, human-preference evaluation now serves as the industry’s pricing-and-positioning infrastructure — academic benchmarks alone no longer cover it. Second, a fast-growing commercial services layer rides on top of those rankings, and any serious reading of a leaderboard position should account for that ecosystem.