The Data Bottleneck Behind Frontier AI

Mercor's CPO Osvald Nitski opens with a striking claim: "We end every week with so much more money in the bank." The company, which supplies human-generated evaluation and training data to frontier AI labs, is struggling to spend fast enough to keep up with demand. This is not a business suffering from an ROI problem—it's one where the bottleneck has shifted from capital to execution.

Why Open Source Doesn't Threaten the Data Business

When asked whether open-source models cannibalize Mercor's core business, Nitski pushes back on the premise. Open models raise the floor of what's possible, but data becomes most valuable at the frontier of model performance. "As long as customers still have new capabilities that they want to get better at, our business still continues to grow," he explains. The common framing that 90% of enterprise workflows can already be handled by existing models misses something crucial: latent demand. People aren't even trying to do long-horizon tasks like fully automated procurement agents that run for months with only weekly check-ins. Those use cases aren't captured in current calculations, and the market for data to support them is growing.