Overview of 20VC: Mercor CPO on Revenue Concentration from Frontier Labs with Osvald Nitski
This episode is a wide-ranging conversation between Harry Stebbings and Osvald Nitski, CPO at Mercor, focused on how AI model improvements are reshaping the human data market, enterprise adoption, and product strategy. The core thesis: frontier labs still drive the most valuable demand for eval and training data, but Mercor is increasingly moving downmarket as enterprises build specialized models and AI workflows. The discussion also covers why ROI concerns are still overstated, why services are a short-term bridge for AI deployment, how product teams must evolve in the AI era, and why robotics and physical data may be the next major opportunity.
Main Takeaways
Frontier vs. open source is changing the market, but not killing demand
- Osvald argues that open-source model gains don’t cannibalize Mercor’s business; they mostly raise the floor of what customers expect.
- Mercor’s data products are most valuable at the frontier of performance, where model capabilities are still incomplete.
- The company believes there is still plenty of latent demand that current benchmarks undercount, especially for long-horizon enterprise workflows.
Enterprise AI adoption is still blocked by workflow complexity, not lack of ROI
- Osvald says there is no broad enterprise ROI problem yet.
- Companies are still in an exploration and experimentation phase, with spending tolerated as they learn what works.
- ROI depends heavily on the use case:
- High-growth or engineering-heavy use cases can justify significant spend.
- Customer support or low-value workflows are more sensitive to token costs.
Specialized models for every company is a real future
- He broadly agrees with the idea that companies will increasingly need tailored models for their own objectives.
- Mercor sees itself as benefiting from this trend because each specialized model needs:
- eval data
- training data
- model-specific measurement and feedback loops
- The future is less about one universal model and more about continuous optimization for each company’s goals.
Large enterprises are still cautious with core proprietary data
- Enterprises are more comfortable using AI on general workflows like HR and procurement.
- They remain more hesitant to share or expose core differentiating work, like legal advice, strategic memos, or domain-specific operations.
- Osvald notes that open-weight models offer more control over inference, which can reduce some trust concerns.
Mercor is shifting from pure frontier-lab dependence toward enterprise self-serve
- One of Mercor’s biggest strategic goals is to reduce revenue concentration by serving more enterprises.
- The company wants to make human-data projects easier for smaller customers via:
- better product tooling
- more automation
- AI project managers
- self-serve workflows
- This is harder than the white-glove lab motion, but it’s the path to broader distribution.
Product and Company Strategy
Mercor’s product surface area has to stay simple
- Osvald says the hardest product challenge is resisting feature bloat.
- AI makes it tempting to expand rapidly, but Mercor is actively trying to:
- simplify workflows
- narrow supported use cases
- focus on the highest-value product paths
- Their product team is now more focused on what drives business value than on tool mastery.
PMs need stronger business judgment in the AI era
- The role of product management is changing:
- less time spent learning tools
- more time spent using coding agents
- more emphasis on business impact and judgment
- He believes “skill issues” are increasingly less about execution and more about deciding what matters.
Mercor’s internal structure is small, fast, and pod-based
- The company operates with relatively small product pods across two main areas:
- Marketplace: matching experts to jobs
- Studio / annotation platform: running eval and training data projects
- They keep weekly product-area meetings to align pods and avoid communication breakdowns as the company grows quickly.
Data, Workflows, and the Next Frontier
The fastest-growing data type is “environments”
- Osvald says RL environments / simulated workflows are becoming the fastest-growing category.
- These are high-fidelity simulations of real tools and systems, such as:
- software apps
- enterprise systems like Salesforce
- laptop-like file environments
- The goal is to train and evaluate agents on tasks that look like real deployment conditions, not toy examples.
Security and cyber may become a major AI data market
- He sees cyber as one of the most interesting and fastest-growing categories because it is:
- adversarial
- constantly moving
- hard to “solve” definitively
- That makes it a great fit for AI-driven data generation, evals, and competitive benchmarking.
Robotics is the big under-discussed opportunity
- Osvald is especially bullish on robotics and physical-world data over the next three years.
- He thinks the market is still early relative to genAI and autonomous vehicles.
- His view is that robotics will likely follow a path more like Waymo than ChatGPT:
- useful and real
- but slower to scale due to physical-world constraints
Hiring, Culture, and Talent
Mercor is hiring for seniority, agency, and judgment
- The company now prefers candidates who can:
- understand business impact quickly
- run experiments well
- think through systems and tradeoffs
- Tool fluency matters less than judgment, ownership, and ability to drive outcomes.
They test for AI fluency, then probe deeper
- Their hiring process includes:
- one take-home assignment using an AI agent
- then whiteboarding around experiments, statistics, and systems design
- The goal is to ensure candidates can still think independently and not over-delegate judgment to models.
High-agency people sometimes leave to found companies
- Osvald is proud that many of Mercor’s alumni go on to start companies.
- He views that as better than losing people to random jobs.
- Mercor’s culture is intentionally high-performance and high-agency, even if that makes management harder.
Advice and Notable Remarks
Advice for students and early-career builders
- His advice to a computer science student:
- Get a real internship as soon as possible
- what you learn in school will likely become outdated fast
- join a fast-growing company, ideally near the frontier
- He recommends not going too early-stage unless you can clearly filter for quality.
On product and judgment
- He repeatedly stresses that teams should not delegate:
- decision-making
- judgment
- strategy
- In his view, AI can accelerate execution, but overreliance can weaken thinking muscles.
On scaling Mercor
- Mercor’s bull case is that it becomes a huge business because:
- evals and training data are becoming the bottleneck to better models
- specialized enterprise models will need continuous data support
- the company can expand into agent deployment and robotics data as well
Bottom Line
Mercor’s CPO paints a picture of an AI market that is still early, still messy, and still expanding. Frontier labs remain the biggest customers today, but the long-term play is broader: helping every enterprise build specialized AI workflows, from evals and training data to deployment and environments. The episode’s central message is that the real bottleneck is not model access anymore — it’s high-quality data, workflow design, and operational execution.