Nvidia Negotiates Investment in Mercor at a $20 Billion Valuation

Nvidia is negotiating an investment in Mercor, a company that supplies experts to train AI models, at a $20 billion valuation, double the $10 billion it was worth just nine months ago. General Catalyst, the company's existing investor, is set to lead the round. Nvidia isn't a random player here: it already pays Mercor millions of dollars a quarter for data feeding its own open-source Nemotron models.

What Mercor actually does

Mercor connects AI labs and companies with experts, lawyers, doctors, bankers, scientists, who train frontier models through RLHF, reinforcement learning from human feedback. In practice, it's a company brokering the most labor-intensive, least glamorous part of building large language models: someone has to rate thousands of responses, fix errors, decide which version of a piece of text is better, and do it with specialist knowledge the model itself doesn't have yet.

The company was founded in 2023 by three 22-year-olds who met on their high school debate team: Brendan Foody and Adarsh Hiremath as co-CEOs, and Surya Midha as board chairman. In October they became the youngest self-made billionaires in the world. Mercor crossed $2 billion in annualized revenue in June, doubling just four months after hitting its first billion in February.

Why Nvidia is even investing in a data company

On the surface, Nvidia and Mercor operate in completely different segments: one makes chips, the other organizes people's work. In practice, the two businesses are more tightly linked than the industry labels suggest. Nvidia sells the hardware models get trained on, but that hardware's value only grows if there are models worth training, and the quality of those models increasingly depends on training data quality involving real experts, not just raw compute power.

By investing in Mercor, Nvidia locks in access to the data infrastructure behind its own Nemotron models, while also tying itself to a company that also serves competing AI labs. It's a strategy consistent with Nvidia's other moves in recent years: instead of just selling chips, the company increasingly invests across the rest of the stack, from data centers to training-data providers, building a web of dependencies where its partners' success also drives demand for its own hardware.

Doubling the valuation in nine months

The jump from $10 billion to $20 billion in under a year shows just how much of a bottleneck training data has become for the entire AI industry. For the past few years, investor attention centered mostly on the models themselves and the compute needed to train them, while data quality and availability, especially data requiring real expert knowledge, stayed in the background. Mercor's financials suggest that's shifting: companies like OpenAI are willing to pay increasingly more for access to real specialists who can rate and correct model outputs, because that stage, not the next generation of GPUs, is increasingly what decides the real difference in quality.

No deal terms have been confirmed, and the investment hasn't closed yet, so it should be treated as advanced negotiations, not a done deal. Still, the sheer scale of the talks, a doubled valuation with one of the biggest players in hardware at the table, is enough to show where capital in AI is actually flowing in the second half of 2026: not just into more models, but into the people teaching those models to think like experts.