Summary
In this episode of 20VC, host Harry Stebbings sits down with Oswald Nitski, Chief Product Officer at Mercor, for an unusually candid look at how the rise of open-source AI models affects companies that sell training data to frontier labs. Nitski argues that open-weight models don’t cannibalize Mercor’s core business — instead, they raise the floor of what’s already commoditized, pushing demand toward the frontier where capabilities remain underdeveloped.
The conversation covers enterprise skepticism around sharing sensitive data with closed AI providers, Mercor’s thesis on “latent demand” for long-horizon agentic tasks like fully autonomous procurement agents, and how the company thinks about revenue concentration risk from a small number of frontier lab customers. Nitski offers a nuanced take on synthetic data — acknowledging it as a long-term threat while arguing that human-generated eval data remains essential at the capability frontier.
The latter portion of the discussion turns to product leadership and team structure at Mercor, including how Nitski balances experiment design with the risk that over-relying on AI tools atrophies human judgment. Stebbings draws a parallel from his own content operation, and the two explore where the boundary between AI-assisted execution and human decision-making should sit. It’s a rare inside look at the business dynamics of the AI data supply chain from a company working directly with labs like Anthropic, OpenAI, and Google.
📺 Source: 20VC with Harry Stebbings · Published July 25, 2026
🏷️ Format: Podcast
