Nvidia May Back Mercor, the Company That Feeds Its Models

The most valuable input into a frontier model is no longer just chips or electricity. It is data, and specifically the kind of data that a domain expert has to sit down and produce. That reality is written all over the news, reported by The Information on August 19, that Nvidia is discussing an investment in Mercor as part of a round that would value the data-labeling startup at $20 billion.

That figure is double the $10 billion Mercor commanded in an October Series C, when it raised $350 million. The size of Nvidia's stake and the total round have not been reported. But the direction is unmistakable: a company barely known outside the industry is now worth as much as some publicly traded chipmakers, on the strength of supplying human expertise to machines.

What Mercor actually sells

Mercor sources specialized human-expert data, pairing companies that need training material with people who have real knowledge in law, finance, medicine and the sciences. Nvidia is already a customer, paying Mercor tens of millions of dollars last quarter for curated data used to develop its Nemotron open-weight models, the same family we covered when Nvidia built a trillion-parameter open model of its own. "Each project expands what models understand," Mercor chief executive Brendan Foody told The Information.

An investor buying into its supplier is a familiar pattern in this cycle, and not always a comfortable one. Nvidia has also backed rival data providers, including a stake in Scale during its 2024 round. Putting money into the firms that both feed and buy from you keeps the ecosystem turning, though it blurs the line between customer, investor and dependency.

The quiet economy behind the models

For a while the story of AI progress was about scale: more parameters, more GPUs, bigger clusters. The Mercor valuation points at the less glamorous bottleneck underneath. Once the easy public text is used up, the gains increasingly come from carefully built, expert-labeled examples that no crawler can scrape. That work is slow, it is human, and it is expensive, which is exactly why a labeling firm can be worth $20 billion.

It also fits the broader money story of the year, in which enormous sums are flowing into the plumbing of AI rather than the models themselves. Nvidia has been at the center of that, from its move to recruit Wall Street for a $500 billion buildout to smaller bets like this one. Whether the returns eventually justify the prices is the question hanging over the entire sector. For now, the people teaching the models have never been in higher demand.