Source linked

Arena's Crowdsourced AI Benchmarking Tops $ 100M Run-Rate in acht Monaten

Das von UC Berkeley geborene AI-Leaderboard, am besten bekannt für seine Crowd-Source-Modell-Rankings aus über 10 Millionen Bewertungen, erreichte nur acht Monate nach dem Start seines kommerziellen Dienstes einen Jahresumsatz von 100 Millionen US-Dollar.

arenauc berkeleyai benchmarkingmodel evaluationpost trainingventure capital

$100 million in annualized run-rate revenue, eight months after launching a commercial product. That’s what Arena, the AI leaderboard spun out of UC Berkeley, just pulled off.

Most people still see Arena as a free open-source project — a place where you type a prompt, get answers from two anonymous models, and pick the better one. Over 10 million evaluations later, that crowdsourced data has become a goldmine for model labs and enterprises hungry for post-training performance analytics.

From Research Project to Revenue Machine

Arena’s commercial offering, AI Evaluations, launched in September 2025. By January 2026, the company was already at $30 million annualized revenue. That same month, it announced a $150 million Series A at a $1.7 billion post-money valuation, led by Felicis and joined by Andreessen Horowitz, Kleiner Perkins, Lightspeed, and others.

Co-founder and CEO Anastasios Angelopoulos told TechCrunch the revenue isn’t recurring — customers pay for consumption. But the growth trajectory is unmistakable: three months later, annualized revenue hit $100 million. Total funding now stands at $250 million.

Why Model Labs Pay for Crowdsourced Opinions

Arena ranks models across text, coding, vision, image generation, and even multi-step agent workflows via its recently introduced Agent Mode. The raw community votes — humans judging model outputs — are messy, high-signal data that labs can’t get from synthetic benchmarks or internal red-teaming.

Angelopoulos frames Arena’s competition not as other leaderboards (Yupp, a similar startup, shut down in March 2026) but as human-labeling giants Mercor, Surge, and Scale AI. All of them help AI providers refine their models after initial training. Arena’s advantage: it already has a built-in community of evaluators who show up for early access to unreleased models.

The Post-Training Gold Rush

Arena’s numbers land in a market that’s exploding. Handshake’s gross annualized revenue from AI training nearly doubled from $550 million in January 2026 to nearly $1 billion by April, per The Information. Mercor’s revenue topped $1 billion in early 2026, up from $500 million in September 2025.

Arena was co-founded by Angelopoulos, fellow UC Berkeley postdoc Wei-Lin Chiang (CTO), and Ion Stoica — the Berkeley professor and Databricks co-founder who advised before the project incorporated in April 2025. The company now has 100+ employees and investors that include The House Fund, LDVP, Laude Ventures, and UC Investments.

As labs race to squeeze every percentage point of performance out of their models, the appetite for real human judgment in post-training will only keep climbing.


Source: Arena, the AI leaderboard everyone uses, is now a $100M business
Domain: techcrunch.com

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.