At hypergrowth companies like Mercor, AI token spend can exceed employee salaries. This is justified because the expenditure directly fuels the ability to service massive, otherwise unserviceable, customer demand, making it a necessary cost for rapid scaling rather than a simple operational expense.
Traditional hiring methods are becoming obsolete. To hire PMs, Mercor uses a single test for AI fluency, then immediately shifts to whiteboarding. This assesses fundamental skills like judgment, experimental design, and statistics—capabilities that cannot be outsourced to AI and reveal a candidate's core thinking.
Data annotation companies face a peculiar competitor: a "cottage industry" of early-stage startups where founders do the annotation themselves. This VC-subsidized labor is often mispriced and attractive to labs for small projects, but it presents a scaling challenge that larger, more systematic providers are built to overcome.
The current boom in AI services is a short-term solution to a knowledge gap. Expertise in deploying AI is currently concentrated in tech hubs. Over the next decade, as this knowledge disseminates and products mature, the need for hands-on services will decline as companies build in-house capabilities.
As AI coding agents make engineers more productive, the development bottleneck eases. The new constraint becomes product management—understanding user needs and business impact. This shift will necessitate a higher ratio of product managers to engineers to effectively guide the accelerated development cycle.
The future of enterprise AI isn't one-size-fits-all. Because performance can always be improved based on unique company goals like growth versus margin, every company will eventually require its own specialized models trained on enterprise-specific evaluation and training data.
The role of a Product Manager is shifting in the AI era. With coding agents handling execution, the need for diverse tools like Figma is diminishing. The PM's core value is now elevated to strategic business judgment, focusing on simplifying product surface area and prioritizing high-impact initiatives.
When critics debate whether Mercor's revenue is from services or software, the CPO offers a practical rebuttal: cash flow. The fact that the company ends every week with significantly more money in the bank demonstrates the business's fundamental health, making semantic arguments about revenue classification moot.
The improvement of open-source models doesn't cannibalize demand for specialized data providers. Instead, it elevates the baseline capability, pushing customers to focus on more complex, frontier problems where high-quality, specialized data is most valuable and commands a premium.
Enterprises are skeptical of sharing core, differentiating data with frontier model providers. They are more comfortable using proprietary models for general functions like HR and procurement, while keeping their most sensitive business data for open-source or in-house models they can control.
The core challenge in making human data projects self-serve isn't the tooling but the nature of the work. These projects solve problems outside current model capabilities, meaning they are defined by a continuous stream of edge cases that require intense, real-time human alignment and paranoia to resolve.
The frontier of AI data is shifting from simple text pairs to complex "environments"—rich simulations of applications or entire user desktops. This data type is growing fastest because it allows models to be trained and evaluated in a context that closely resembles their future deployment.