Nvidia pursues $20B stake in Mercor, its AI training-data supplier
Nvidia is moving to take a direct financial stake in the company that already feeds its own models — a signal that the chip giant now treats training data as strategically captive as silicon.
Nvidia is in talks to lead a round that would value Mercor, an AI training-data supplier, at roughly $20 billion — double the $10 billion it commanded just weeks ago. According to a Bloomberg report, the chipmaker is weighing an investment that would make it both a major backer and a key customer of the startup, whose specially labeled data is already used to train Nvidia's open-source Nemotron models. Mercor posted about $614 million in gross revenue in the first half of 2026 and was running at an annualized rate above $2 billion by mid-year — a pace that helps justify a valuation that looked frothy even by this market's standards. General Catalyst is reportedly shepherding the financing, and Mercor's client list already includes OpenAI and Google.
The more interesting story is what the deal says about where the AI stack is consolidating. For two years the public narrative has been that Nvidia owns the compute layer and everyone else fights over models and applications. But models are only as good as the data they learn from, and high-quality, domain-specific human-labeled data has become a genuine bottleneck as labs exhaust the easy text sitting on the open web. By putting capital — not just purchase orders — behind a data supplier it already depends on, Nvidia is quietly verticalizing into the input side of the equation. It is the same logic that has pushed the company into networking, custom silicon, and software: capture the layers that decide whether its GPUs actually get used.
There is real risk in the intimacy. A supplier that is also a strategic investee of its largest buyer invites questions about pricing, exclusivity, and whether rivals get equal access to the same labeling pools. Antitrust scrutiny of Nvidia is already intense; folding a fast-rising data intermediary into its orbit will not quiet it. And Mercor's own economics are unproven at this scale — a $20 billion tag rests on a revenue curve that, while steep, is a bet on continued enterprise spending on model training rather than a settled business.
What to watch: whether the round closes at the full $20 billion, and whether OpenAI or Google move to lock in their own data supply before Nvidia does.
Is owning your data supplier the natural end state for an AI superpower, or a line Nvidia shouldn't cross? Tell us in the comments.
