Nvidia Reportedly Circling a $20 Billion Stake in AI Data Startup Mercor

[EasternAnvil]

Nvidia is in talks to invest in Mercor, the AI data labeling and expert network startup, at a valuation that could hit $20 billion, according to a new report. That would roughly double Mercor's valuation from its last major round, and it comes less than a year after that round closed.

Mercor's whole business is connecting AI labs with lawyers, doctors, engineers, and other specialists who create and evaluate the training data, benchmarks, and reinforcement learning environments that labs like OpenAI and Anthropic depend on. It is not a flashy consumer product, it is the unglamorous plumbing behind frontier model training, and apparently that plumbing is now worth tens of billions of dollars.

This fits a pattern Nvidia has been running for a while now. Rather than only selling chips, they keep taking equity positions across the entire AI stack, from model developers down to the companies that feed those models data. It effectively means Nvidia benefits twice from the same AI boom, once through GPU sales and again through appreciation of its portfolio.

Worth noting that Mercor's growth has followed the broader surge in demand for reinforcement learning environments specifically, not just plain data labeling. As labs move toward training agents that need realistic tasks to practice on rather than static datasets, that shift seems to be a big part of why a company most people have never heard of is suddenly being talked about in the same breath as a twenty billion dollar valuation.

Still early and unconfirmed by either company, but if it goes through it is another sign of how much capital is flowing into the unseen layers of the AI supply chain

Nvidia in talks to invest in AI data startup Mercor at $20 billion valuation - Tech Startupstechstartups.com

[Matt95]

Never heard of Mercor before this and apparently they are already worth more than most public companies most people could name. The invisible layer of the AI stack keeps getting bigger and richer while nobody outside the industry pays attention to it.

[Jackson79]

The reinforcement learning environment angle is the real story buried here.
Everybody talks about training data like it is still 2023 and it is just text scraping, but the actual bottleneck now is building realistic task environments for agents to practice in, and that apparently takes an army of contracted professionals to build well.