Episode Summary
Executive Summary: Harry Stebbings interviews Osvald Nitsky, CPO at Mercor, about how AI is reshaping enterprise software, data, and product management. Osvald argues open models raise the floor but don’t remove demand for frontier data, that ROI is still exploratory, and that Mercor’s biggest opportunities are specialized eval/training data, AI deployment services, and emerging markets like RL environments, security, and robotics.
Main Topics: Open models vs. Mercor’s data business: Osvald argues that open-source model progress does not cannibalize Mercor; it mainly reduces demand for basic capabilities while increasing interest in frontier gaps, specialized evals, and training data for harder workflows. Enterprise ROI and AI spend: He says there is no broad ROI problem yet because companies are still experimenting, and spend should be judged by use case: growth, engineering leverage, and revenue impact matter more than fixed budget rules. Product management in the AI era: AI reduces tool complexity and execution friction, making judgment, prioritization, experiment design, and business impact the core PM skills. Mercor is simplifying product surface area even as velocity rises. Services as a short-term AI deployment model: Osvald sees AI services/FDE-style deployment as necessary in the near term because expertise is concentrated in a few hubs, but expects software to absorb more of that work over time. Mercor’s market structure and growth strategy: The company is moving downmarket from frontier labs toward broader enterprise self-serve data workflows to reduce concentration risk, while still benefiting from high demand and cash generation. Emerging data markets: RL environments, security, robotics: He highlights RL environments as the fastest-growing data type, security as an adversarial category with uncapped demand, and robotics/physical data as a major three-year opportunity. Hiring, culture, and talent in the AI startup era: Mercor biases toward more senior, high-agency hires who can think in terms of business value. The company values speed, ownership, and a startup culture despite scaling to roughly 500 people.
Key Arguments: Open-source model improvements do not automatically hurt Mercor because customers still need frontier-specific eval and training data for the capabilities they want next. The 90/10 framing is too simplistic; many workflows are latent demand or continuous/improvable tasks rather than binary automations. There is not yet an enterprise-wide AI ROI crisis; spend decisions should be tied to the business function and whether AI creates compounding gains or revenue lift. AI is raising the bar for PM judgment while lowering the need to learn many tools, because coding agents and cloud-based design tools compress execution. Mercor should stay simple operationally; supporting too many workflow variants creates chaos and should be constrained with guardrails. Services are a short-term necessity because AI deployment knowledge is concentrated, but that knowledge will eventually diffuse across enterprises. The fastest-growing opportunity areas are environments for agent training/evaluation, security data, and eventually robotics/physical-world data. Mercor’s long-term strategy is to move from lab-heavy, high-touch projects to more self-serve enterprise workflows, reducing customer concentration. Hiring should prioritize high-agency people who care deeply, can run experiments, and understand business impact rather than just tool fluency. Higher PM-to-engineer ratios make sense as coding agents increase engineering throughput and shift bottlenecks toward product judgment.
Data Points: Mercor headcount growth: 10x+ since Osvald joined - Used to illustrate hypergrowth and demand outpacing the company’s ability to spend. Mercor employee count: ~500 people - Osvald says the company is still culturally a startup despite scaling. Enterprise workflow coverage by open models: ~90% (challenged by Osvald) - Harry raises the common claim; Osvald disputes that this is currently accurate. Apex benchmark coverage for long-horizon workflows: ~50% - Osvald says top models are reaching this level on their internal benchmarks. Example AI spend at Salesforce: $300 million/year on Anthropic - Referenced to discuss how AI spend may compare with salaries and grow over time. AI spend as % of developer salaries: ~3.8% - Harry’s calculation based on the Salesforce figure. Desired PM prime age: 25-35 - Osvald’s estimate of the career stage where candidates are most effective. Framer SLA: 99.99% uptime - Promotional mention in sponsor read, not part of the core discussion. Zero Hash customers cited: Stripe, Morgan Stanley, Gusto, Interactive Brokers - Sponsor read highlighting enterprise usage. Offsite location: Tofino, Canada - Team offsite example used to describe culture-building.
Pivotal Quotes: "We can't spend money fast enough to service all of the demand that we have." — Osvald Nitsky: Describing Mercor’s hypergrowth and why he does not see an ROI problem today. "Get a real internship as soon as possible because whatever you learn in school is probably going to be outdated quickly." — Osvald Nitsky: Advice to computer science students about how fast the AI landscape is changing. "I don't think there's an ROI problem right now." — Osvald Nitsky: His view that enterprises are still in exploration mode and the ROI math is not settled.
Implications: AI infrastructure winners may be those that pair frontier model understanding with specialized data, evaluation, and deployment services. Enterprises will likely shift from generic automation to high-value, domain-specific workflows, while product teams must optimize for judgment, not just speed.