Jul 25, 2026, 10:00 AM ETInterview, Fireside Chat

Mercor Head of Product on Revenue Concentration from Frontier Labs

McCore reports a "hyper-growth" state where weekly cash flow increases by "millions," with the company unable to spend money fast enough to service current demand.
The business has expanded headcount by more than 10x since the CPO joined, driven by insatiable customer demand for human data and eval services.
While revenue is currently concentrated with frontier model labs, the strategic pivot involves moving down-market to serve smaller enterprises via self-serve tools.
McCore is actively reducing product surface area to simplify operations, a necessary counter-measure to the chaotic expansion of features enabled by AI coding agents.

Oswald Nitzke argues that open-source models do not cannibalize McCore's core business because data value is highest on the "frontier" of model performance.
Open models raise the baseline capability, shifting customer demand toward the remaining 10% of enterprise workflows that open models cannot yet handle.
Nitzke disputes the "90/10" split (open vs. closed) often cited in the industry, estimating via APEX benchmarks that top models only handle ~50% of long-horizon workflows effectively.
A significant portion of future demand lies in "latent demand," specifically long-horizon tasks (e.g., fully automated procurement agents) that enterprises have not yet attempted.
The market for "specialized models" per company is growing, driven by the need for continuous improvement in areas like legal advice and medical outcomes where "sufficiency" is not enough.

Enterprises currently show less skepticism toward data sharing for non-core workflows (HR, procurement) compared to core differentiating activities (e.g., proprietary legal memos).
Nitzke states there is no current "ROI problem" in AI, describing the market as being in an "exploration and experimentation" phase with high tolerance for spend.
Salesforce's spend on Anthropic (~$300M/year, ~3.8% of dev salaries) is considered the baseline, with projections suggesting this percentage will rise over time.
Nitzke advises founders to prioritize token spend on growth-driving functions (e.g., coding agents for software engineers) over fixed-cost optimization (e.g., low-margin customer service).
The "services" sector for AI deployment is viewed as a short-term necessity to bridge the current skills gap, expected to evolve into a standard job function within a decade.

Product management roles are shifting from tool proficiency to high-level business judgment, as coding agents and cloud design tools reduce the need for technical "how-to" skills.
McCore is moving away from Figma in favor of cloud-based design tools to reduce license complexity and friction.
The ratio of Product Managers to Engineers is increasing because engineering velocity is no longer the bottleneck; understanding user needs is now the primary constraint.
A key product lesson learned was the need to place "guardrails" on service offerings rather than building a maximally flexible tool for every customer request to avoid operational chaos.
Two main product groups operate independently: the "Marketplace" (matching experts to jobs) and "Studio" (the annotation and eval platform).

McCore biases hiring toward senior candidates (ages 25–35) who can rapidly "grok" business impact, as junior execution skills are increasingly automated by AI.
Interview processes now emphasize running good experiments, systems design, and statistical judgment over traditional take-home assignments, which are replaced by agent-based testing for AI fluency.
The company accepts high attrition to founding teams ("McCall Mafia") as preferable to competitors, valuing the growth of high-agency individuals over retention.
Hiring challenges in San Francisco remain acute, particularly regarding talent retention and assessment of "agency" and ownership during the interview process.
McCore prioritizes "experts who give a shit" over cultural fit smoothing, believing that a lack of ownership is harder to coach than personality clashes.

"RL Environments" (simulations of apps and digital worlds for agent training) are identified as the fastest-growing data type, shifting from supervised fine-tuning to complex interaction-based tasks.
Cybersecurity data is seen as a permanent "uncapped reward" category due to the adversarial nature of the field, where goalposts constantly move between offense and defense.
Robotics data is predicted to see significant growth in the next three years, though scaling may follow a "Waymo" trajectory (geographic expansion) rather than a "ChatGPT" trajectory (instant global scale).
McCore respects customers more than competitors, noting that competitors typically replicate features only a week or month after McCore releases them.
The company plans to scale physical/data supply ahead of demand by retaining exceptional talent and creating off-the-shelf data assets for future resale.

Nitzke recommends CS students secure internships immediately, as university curricula will be outdated before graduation.
He advises joining fast-growing, frontier-focused companies in San Francisco (approx. 50+ employees) to stay relevant in the evolving field.
The company maintains a "cult-like" startup culture despite 500 employees, emphasizing paranoia, office presence, and high-speed iteration.
McCore views itself as a tech-enabled services company where the primary bottleneck for model performance is currently evals and training data.
Future revenue lines will likely include "real-world physical data" for robotics, complementing the current Gen AI focus.