Gokul Rajaram: AI Data Companies Need More Than Revenue to Build Durable Value

Summary

Gokul Rajaram argues that companies supplying data, experts, evaluations, and reinforcement-learning environments to AI labs are growing rapidly, with some reaching tens or hundreds of millions of dollars in annualized revenue within a year. Investors nevertheless face uncertainty because these businesses may depend on a small number of customers, have limited recurring revenue, and must continually produce new data as labs’ needs change. Rajaram compares the sector with early-2000s advertising networks: those companies solved fragmented buying and selling, but ad exchanges and demand-side platforms eventually centralized the durable value, leaving many networks acquired or shut down. He says AI suppliers face a similar risk when labs define the work, own the output, and can move between vendors. His proposed sources of durable equity are differentiated supply, such as an exclusive community or relationship; deep integration into a lab’s workflow for identifying weaknesses, designing tasks, evaluating results, and choosing future data; and self-improving environments that turn model failures into increasingly useful training tasks and verify their outcomes. He expects vertical specialists with these assets to outperform broad horizontal providers as data sourcing becomes easier and more commoditized. The question for founders, he says, is what they will own after the first $100 million in revenue that makes the next $100 million easier to win and harder for competitors to take.