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
This episode focuses on how frontier AI models, enterprise adoption, and product strategy are evolving together. Osvald Nitski argues that open-source improvements do not eliminate the need for high-quality data, especially at the frontier where model performance is still improving. The conversation also explores why many enterprises are still experimenting with AI rather than demanding immediate ROI, and how that changes buying behavior and workflow adoption. A major theme is how AI is reshaping product management, shifting emphasis away from tool fluency and toward judgment, simplification, and business impact. The episode closes with a discussion of robotics, where the speaker is optimistic but sees physical-world scaling as much harder than software and expects narrower service-led breakthroughs before a broad consumer robotics moment.
Key Takeaways
- 1Open-source models are unlikely to destroy the data-provider business because frontier data remains valuable where model performance is still being pushed forward.
- 2Enterprise AI adoption is still driven more by experimentation than by a strict ROI calculation.
- 3AI is changing product management by making judgment and simplification more valuable than raw tool proficiency.
- 4AI deployment services and forward-deployed engineers are likely to remain in demand in the near term because expertise is still concentrated.
- 5Mercor’s growth strategy depends on reducing customer concentration by moving down-market with self-serve products and AI project managers.
- 6Robotics is promising, but its path to scale will likely look more like constrained service deployment than a sudden consumer breakthrough.
Notable Quotes
""We can't spend money fast enough to serve us all of the demand that we have.""
""In our ATX benchmarks, we're getting closer to around 50% of long horizon workflows.""
""I don't think there's an ROI problem right now. I think we're in a period of exploration and experimentation where there's more tolerance, more patience to get that ROI calculation right now.""
""It's a lot better to lose someone to starting a company than to, you know, taking another job.""
""It can't deal with like very ambiguous data. It's still like pretty irrelevant.""
""I think it might play out similar to driverless cars where it's really hard to scale physical things as opposed to software.""
""It might be more of a, more of like a Waymo, robotaxi, Cruise type moment than a, than a ChatGPT moment.""
""Most companies have tried AI, most on seeing results, not because AI doesn't work is because AI hasn't reached the workflow's yet.""
Episode questions
Why doesn’t open source kill Mercor’s data business?
Because Mercor sells data where model performance is still on the frontier, not where models are already good enough. Open models raise the baseline, but they also expand the set of tasks customers want to improve.
What kind of AI workflows are still hardest to automate?
Long-horizon, high-judgment workflows like procurement agents, legal reasoning, and other tasks where performance can always improve. Osvald says these use cases are often not captured in simplistic adoption percentages.
How should founders think about token spend versus performance?
It depends on the use case: for growth or engineering efficiency, higher spend can be justified if it compounds output. But if token cost exceeds the value delivered, especially in customer-facing workflows, the economics break down.
What changed in product management at Mercor because of AI?
Product teams are expected to simplify rather than expand, and PMs are judged more on business impact than tool fluency. AI reduces bottlenecks in execution, so the real constraint becomes judgment and prioritization.