The chip company that dominates AI hardware is reaching down the supply chain for the raw material its customers cannot get enough of: training data.
Nvidia is in talks to invest in Mercor, an AI data provider that supplies labeling, curation and a specialized talent network for model training, according to The Information. The round is expected to value Mercor at roughly $20 billion. If the deal closes, Nvidia would extend its footprint from the silicon layer to the data layer, locking in access to the high-quality training data that has become one of AI’s scarcest resources.
Mercor’s business sits between the internet and the model. The company assembles large networks of contractors and specialists who label data, verify model outputs and perform the human-in-the-loop work that shapes how models learn. Its platform combines software tooling with a managed workforce, and it has become a critical supplier to the labs that train the world’s largest models. The company’s revenue has historically been concentrated among the closed-model developers: OpenAI, Google and Anthropic.
The Nvidia connection runs through the Nemotron program. Nvidia has pushed a line of open-source models designed to compete with the best open-weight systems available, a strategy aimed at broadening the ecosystem around its hardware. Training those models requires the same data infrastructure the closed labs use, and Mercor’s work for Nvidia has grown as the program has expanded. An investment would formalize a relationship that is already a revenue line.
The strategic logic mirrors Nvidia’s pattern of investing across the AI stack. The company has backed cloud providers that buy its chips, energy companies that power data centers and a widening portfolio of startups that depend on its hardware. A stake in Mercor would add data to that list, giving Nvidia visibility into, and a claim on, the input that determines model quality. Analysts said the move reflects a simple insight: whoever controls scarce inputs controls the industry’s economics.
For Mercor, the deal would bring the most important chip company in the world onto its cap table. That carries commercial advantages: a supplier backed by Nvidia gains credibility with the labs and enterprises that buy Nvidia hardware, and the relationship could open doors across the industry. But it also creates complications. The closed-model labs that provide most of Mercor’s revenue are, in some sense, Nvidia’s customers and competitors, and their willingness to keep buying from a company with the chip giant as an investor is an open question.
The $20 billion valuation would mark one of the fastest ascents in the data industry. Mercor’s growth has been driven by the explosion in demand for data work as models have scaled, and investors have rewarded companies positioned at the bottleneck. The round’s backers, with General Catalyst in discussions to lead, are betting that data will remain a constraint on AI development for years.
The deal would also sharpen the competition among AI’s infrastructure suppliers. Data providers, like chip makers, are consolidating around a few large players, and the companies that control the best data networks will have outsized influence over what models can learn. Nvidia’s entry into that market puts pressure on rivals and on the labs that would rather own their data supply chains.
The unresolved details are substantial. The size of Nvidia’s stake, the final valuation and the terms of any commercial agreement remain in flux, according to people familiar with the discussions. Regulators could also take an interest: an investment that links the dominant AI chip maker to a dominant data supplier touches the same concentration concerns that have drawn scrutiny to the broader AI supply chain.
The data market itself is shifting beneath the deal. Model developers are spending more on data licensing, synthetic data generation and evaluation, moving beyond the simple labeling that defined the industry’s early years. Mercor has positioned itself across those categories, offering evaluation and alignment work alongside annotation, and its growth reflects the industry’s growing conviction that data quality, not just quantity, separates the best models from the rest. That conviction is one reason investors have been willing to pay startup prices for data companies, and it is the same conviction driving Nvidia’s interest.
For the industry, the talks are the latest evidence that the AI value chain is consolidating around a small set of companies that control critical inputs. Nvidia already owns the chips, the software and much of the networking; adding data would complete a portrait of vertical integration that no other company in the industry can match. Whether the deal closes or not, the signal is clear: in the AI economy, the fight for data has moved to the top of the agenda, and the companies at the center of the market intend to win it.
