The most valuable thing in AI right now might not be a chip or a model. It might be the person who knows enough about tax law, or radiology, or C++, to teach a model where it is wrong.

That is the business Mercor is in, and according to The Information, Nvidia has been in talks to fund the company's next round at a valuation of around $20 billion. Nothing is closed — neither Nvidia nor Mercor has commented, and the size of any Nvidia cheque and the total round were not reported — so this is a discussion, not a done deal. But the number itself is the story: $20 billion is roughly double the $10 billion Mercor was worth in October, which was itself five times its valuation the round before.

What Mercor actually sells

The Information describes Mercor as a data-labeling supplier, which undersells it a little. "Data labeling" conjures crowd workers drawing boxes around cats. Mercor sits at the expensive end of that market: a marketplace that vets and matches human domain experts — doctors, lawyers, PhDs, senior engineers — to the AI labs that need them to write, correct and grade training and evaluation data. Its co-founder and chief executive, Brendan Foody, frames it as training frontier models by "sharing knowledge, experience and context that can't be captured in code alone."

The short version: once a model has read most of the public internet, the way you make it better is to pay experts to show it the reasoning the internet never wrote down. That expert labour is the scarce input now, and Mercor is one of the companies selling it wholesale.

Why Nvidia, specifically

Nvidia is the obvious backer because it is already a big customer. As Nvidia builds out its Nemotron family of open-weight models, its spending on Mercor has been climbing — Mercor earned tens of millions of dollars from Nvidia in the last quarter alone, per the report. Funding your own supplier at a moment when you depend on it more each quarter is a familiar move; it locks in access and captures some of the upside you are helping to create.

It also is not new for Nvidia. The company already buys from Mercor's rivals Turing and Scale AI, and it invested in Scale's 2024 round at a $14 billion valuation. What has changed is the price of the whole category. When Meta effectively absorbed Scale's leadership last year, it signalled that the labs see the human-data pipeline as strategic infrastructure rather than a line-item cost. A $20 billion tag on Mercor says the same thing louder.

The pattern worth clocking

Step back and the shape is vertical integration. Nvidia makes the chips the models train on, is funding the company that supplies the human data they train on, and ships its own open models on top. Each layer it owns or backs makes it less dependent on any single customer or supplier — and more central to everyone else's roadmap.

It is also of a piece with the rest of this week's money news, where a lot of AI value quietly moved down the stack, toward the plumbing: the routing layer, the data layer, the pipes rather than the models. The models keep getting cheaper and more interchangeable. The stuff that feeds and connects them is where the durable margins are starting to look.

The caveat stands: this is a reported talk, not a signed round, and valuations at this speed can reprice fast in either direction. But even as a rumour, a $20 billion valuation for a three-year-old company whose product is access to smart humans tells you where the AI industry currently thinks the bottleneck is.