20VC: Mercor CPO on Revenue Concentration from Frontier Labs | Why Large Enterprise is Scared to Partner with Frontier Labs | Why Small Specialised Models is the Future with Osvald Nitski

What It Covers

Mercor CPO Osvald Nitski covers how open-source models, enterprise AI skepticism, and revenue concentration from frontier labs shape Mercor's data business. He addresses specialized model demand, the shift toward RL environment data types, product team structure in high-growth AI companies, and why robotics represents the next major data market opportunity.

Key Questions Answered

• •Open Source Model Impact: Open-source model improvements raise the floor of what data buyers need, rather than cannibalizing human data providers. When open models handle commodity tasks, demand shifts toward harder frontier capabilities. Mercor's benchmark data shows top models completing roughly 50% of long-horizon workflows, leaving substantial uncaptured latent demand in areas like fully autonomous procurement agents running unsupervised for weeks.
• •Enterprise AI Concentration Risk: Mercor's primary revenue concentration among frontier labs is a known strategic vulnerability. The mitigation path runs through moving downmarket so individual enterprises can self-serve human data projects for proprietary model eval and training. This requires building AI project managers into the platform to replace the current white-glove operations model that only scales for large lab customers.
• •RL Environment Data as the Next Frontier: The fastest-growing data type at Mercor is reinforcement learning environments — simulated applications with rich start states representing real machine data, used to train agents on realistic deployment conditions. This mirrors how supervised fine-tuning and preference ranking were difficult to operationalize early on, and labs are still in early stages of standardizing environment-based training pipelines.
• •PM-to-Engineer Ratio Shift: As coding agent velocity increases, engineering becomes less of a bottleneck and product judgment becomes the constraint. Mercor is deliberately increasing the ratio of PMs to engineers, hiring more senior product managers who understand business impact over tool proficiency. Interview processes now include one AI fluency take-home task followed by whiteboard sessions testing experiment design, statistics, and systems thinking.
• •Services as Temporary Knowledge Gap: The rise of AI services businesses at companies like Microsoft and Palantir reflects a temporary concentration of deployment expertise in San Francisco rather than a permanent business model. As agent deployment knowledge disseminates across industries over roughly a decade, services will transition into an internal job function similar to software engineering, and product sophistication will reduce the need for external setup teams.
• •Cybersecurity as Uncapped Data Demand: Cybersecurity represents a structurally different data category from standard enterprise workflows because it is permanently adversarial. Unlike CRM updates where sufficiency ends demand, offensive and defensive cyber capabilities create continuous uncapped reward structures where model performance can always improve. This makes cyber one of the fastest-growing data categories at Mercor, with no equivalent saturation ceiling to the 90% enterprise workflow completion framing.