Mercor CEO Brendan Foody: working with 6 of the Mag 7, 45% month-over-month growth, and why AI training now needs real-world professionals not academics

Summary

Mercor is now working with six of the seven Magnificent 7 companies, including all five of the top-tier hyperscalers, and has averaged 45% month-over-month growth for the past 12 months. CEO Brendan Foody notes the momentum has been amplified by turbulence at competitor Scale AI, which drove a meaningful surge in inbound customer interest.

The Shift From Mechanical Turk to Professional Expertise

The nature of AI training data work has fundamentally changed. Reinforcement learning has become effective enough that models can rapidly saturate any evaluation benchmark once one exists, which means the bottleneck is now building evals and RL environments across new domains. That is pulling demand away from academic or PhD-level annotators and toward working professionals.

• Mercor is actively recruiting consultants, doctors, lawyers, and bankers to evaluate and teach models how to perform real on-the-job tasks.
• The focus is on creating the measurement infrastructure first — defining what success and failure look like — so training data can be generated around it.
• Foody frames this as the primary constraint on capability expansion: anything a model cannot yet do needs an effective way to be measured before it can be improved.

The "Book a Flight" Problem as a Proxy for Agentic Gaps

Despite frontier models demonstrating Olympiad-level math reasoning, basic agentic tasks like booking a flight with personal preferences remain unsolved at a consumer trust level. Foody attributes this to the absence of robust evals for tool-use and computer-use workflows. Hyperscalers are running a broad buildout across use cases ranging from simple transactional actions to high-complexity reasoning over large knowledge bases, and Mercor is embedded in that pipeline.

Video as an Emerging Frontier

Multimodal and video model training is an active and growing workstream. Beyond standard transcription, clients are building rich metadata layers — human-expert tags describing what is physically happening in video content — to enable more accurate generation and retrieval. For long-form content like hour-long lectures, eval work around what elements matter to users is a specific focus area. For generative video models, Foody confirms demand for human annotators who can describe scene-level physics and action detail in plain text, feeding that structured data back into training pipelines to reduce hallucinations in future model versions.