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

Summary

Osvald Nitski, Mercor’s CPO, argues that open-source progress does not cannibalize Mercor because the company sells frontier eval and training data tied to capability gaps, not commodity tasks. He says enterprise AI adoption is still in an experimentation phase, with token spend judged by use case, function, and expected business impact rather than a blanket ROI problem. Large enterprises, in his view, are less sensitive about generic workflows like HR or procurement than about core differentiating work, and that sensitivity should support demand for enterprise-specific models and data. Mercor is also trying to reduce concentration by moving downmarket into self-serve human-data workflows, while simultaneously simplifying product surface area, raising the PM-to-engineer ratio, and focusing on higher-value workflows like RL environments, cyber, and eventually robotics.

Notes

• Mercor’s core data business is framed as frontier-only, not commoditized by open-source gains.
• Mercor’s customers buy eval and training datasets to close specific model gaps.
• Open-source models mainly raise the floor of what buyers consider worth paying for.
• Mercor’s view of enterprise AI is still centered on exploration, not a broad ROI panic.
• Token prices and model performance are still moving enough to delay hard ROI conclusions.
• General workflows like HR and procurement are less sensitive for enterprises to put on proprietary models.
• Core differentiating work is where enterprise sensitivity rises sharply.
• Mercor’s benchmark framing is closer to 50% of long-horizon workflows, not a simple 90/10 story.
• Some workflows should be judged as continuous improvement problems, not pass/fail tasks.
• Mercor believes specialized models are likely, but they need enterprise-specific eval and training data.
• The market for specialized models depends on whether customers can justify the ROI.
• Token budgets should vary by function; growth and efficiency uses can justify aggressive spend.
• High token spend can still make sense for coding agents if it compounds engineering output.
• Customer support is a bad unit-economics test if token cost exceeds revenue per customer.
• Mercor wants company-wide accounting that ties AI spend to outcomes by team.
• Different teams should have different tolerance levels for AI spend.
• Mercor says it cannot spend money fast enough because demand is outpacing headcount and operating capacity.
• Mercor’s headcount has grown more than 10x since Osvald joined.
• Product management is getting harder because AI increases engineering speed without increasing product clarity.
• Mercor is fighting to reduce product surface area and simplify the stack.
• AI velocity shifts the bottleneck from engineering to judgment and workflow understanding.
• Mercor is reducing tool sprawl; fewer tools and more coding-agent leverage are enough for PMs.
• Figma is being displaced internally by cloud design workflows.
• Mercor’s annotation platform became too flexible and needed guardrails around what services it supports.
• Mercor’s flexible annotation tooling became unmanageable at hundreds of projects.
• Mercor has two main product areas: the Marketplace and Studio.
• Mercor serves both talent-only placement and managed-service dataset delivery.
• Mercor expects the PM-to-engineer ratio to rise as coding agents increase engineering velocity.
• Weekly product cadence is pod-based, with separate roadmap accountability.
• Mercor’s main scaling issue is internal communication across products.
• Expert supply scales through great worker experience, referrals, and global sourcing.
• Marginal economics are secondary to customer value and expert experience.
• Mercor says project costs are driven mainly by expert pay and LLM spend for QC and synthetic data.
• Mercor’s cash flow is so strong it says weekly bank balance rises by millions.
• Mercor wants to move downmarket so smaller enterprises can self-serve human data projects.
• Downmarket self-serve is positioned as a way to diversify revenue beyond frontier labs.
• Human data projects are operationally intense because edge cases require fast alignment.
• Mercor’s data formats have evolved from SFT to preference ranking, rubric annotation, and RL environments.
• The fastest-growing Mercor data type is environments.
• Mercor defines environments as simulated apps plus a rich start state and agent tasks.
• Environment start states can require hundreds or thousands of files.
• Better evals and training data can directly lift customer revenue.
• Mercor argues spend on data is acceptable when it trails revenue gain.
• Mercor sees a cottage industry of founder-led annotation work that does not scale well.
• Mercor says smaller annotation shops struggle to scale beyond a few projects or data points.
• Cybersecurity data is attractive because offense-defense dynamics create uncapped rewards.
• Security workflows are unlike CRM tasks because the goalposts keep moving.
• Mercor describes itself as high-agency, high-performance, and high-ownership.
• Mercor wants hires who care deeply, not just people with technical skill.
• Osvald says RL environments were initially underestimated but later proved viable.
• Mercor’s robotics thesis is that real-world physical data should grow sharply over three years.
• Robotics is still nascent compared with genAI and AV data markets.