Nvidia Is Reportedly in Talks to Back AI Data Supplier Mercor at a $20 Billion Valuation

Just weeks after lining up more than $500 billion in third-party financing for AI data centers, Nvidia is now reportedly turning its checkbook to the other side of the training pipeline: the humans who teach frontier models what to know. According to a report from The Information published August 19, Nvidia is in talks to invest in Mercor, the San Francisco-based expert-data marketplace that supplies AI labs with domain specialists for model training. The investment would form part of a new funding round valuing Mercor at roughly $20 billion — double the $10 billion valuation it commanded at its Series C just ten months ago — with existing investor General Catalyst reportedly in talks to lead.

The deal, if it closes, would be the latest and loudest signal that expert training data has become one of the most contested inputs in the AI stack, and that the market for it is consolidating around a handful of well-capitalized platforms.

What Mercor actually does

Mercor started life in 2023 as an AI recruiting startup founded by three high-school friends — Brendan Foody, Adarsh Hiremath, and Surya Midha, all Thiel Fellows in their early twenties — who are now among the world’s youngest self-made billionaires. The original product interviewed candidates with an AI avatar and matched them to employers. When frontier AI labs came asking for the same thing in reverse — a pipeline of vetted human experts to label, evaluate, and correct model outputs — the company pivoted into what has become the fastest-growing corner of the data business.

Today Mercor operates a marketplace that connects AI labs and enterprises with domain experts: PhDs, doctors, lawyers, investment bankers, even poets recruited to grade verse. Experts go through Mercor’s AI-conducted interviews and skill assessments, then work on training and evaluation tasks — reinforcement learning from human feedback (RLHF), red-teaming, benchmark construction, and the increasingly valuable craft of building evaluation frameworks for specialized domains. Forbes described the model bluntly: Mercor recruits people to train more AI, and business has been explosive.

The numbers behind the $20 billion

The valuation talk is underpinned by unusual growth. According to Sacra and documents reviewed by The Information in July, Mercor hit $2 billion in annualized gross revenue in June 2026 — up from $760 million at the end of 2025 — after taking just 17 months to go from zero to a $500 million run rate. Gross revenue for the first half of 2026 came in at roughly $614 million, up about 70% from the company’s entire 2025 total.

Forbes first reported on July 9 that Mercor was in talks to raise around $500 million at a $20 billion valuation. The Nvidia involvement now being reported adds a strategic dimension: Nvidia would not just be a passive investor but a supplier-turned-backer of a company that feeds its most important customers. Notably, The Information’s July reporting also flagged the concentration risk — roughly 90% of Mercor’s H1 revenue came from a small group of the biggest AI companies, a dependency that any investor in the round will have to underwrite.

Why Nvidia wants in

The logic follows the same vertical-integration playbook Nvidia has run across the AI stack. The company has already taken large equity positions in its own customers — OpenAI, xAI, CoreWeave — to guarantee demand for its GPUs. Backing Mercor extends that logic upstream into the data supply chain: better expert data makes frontier models better, better models drive more training and inference compute, and more compute demand sells more Nvidia systems.

It is also a hedge on the direction of model improvement. As labs exhaust high-quality web text, competitive advantage has shifted toward curated, expert-generated data — synthetic data grounded by human specialists, domain-specific RLHF, and human-designed evaluations. The companies that can source and orchestrate that expertise at scale (Mercor, Scale AI, Surge AI) have become chokepoints for the next generation of models. Nvidia’s reported interest suggests it sees them the same way it sees power and advanced packaging: strategic inputs too important to leave to the open market.

A three-way race for the expert-data market

Mercor is not the only player monetizing the shift. Scale AI, the incumbent, saw its fortunes transformed by Meta’s $14.3 billion stake for a 49% nonvoting share at a $29 billion valuation, before pivoting toward enterprise and government work. Surge AI, Scale’s smaller rival, reportedly generated around $1 billion in revenue last year and carries a $15 billion valuation. Mercor’s $20 billion round, if completed, would place it second in the cohort by valuation despite being the youngest of the three — a remarkable position for a company founded in 2023.

The market itself is still early. Projections cited across the industry put data annotation spending at roughly $1.2 billion in 2024, growing toward $10 billion or more by 2034 — figures that look conservative if RLHF and expert-evaluation demand keeps compounding at anything close to Mercor’s pace.

What to watch

Three open questions will determine whether the $20 billion price makes sense. First, concentration: with ~90% of revenue tied to a handful of frontier labs, Mercor’s fortunes rise and fall with a very small customer set, and any lab deciding to internalize its data operations would sting. Second, margin structure: Mercor’s figures are gross revenue, and the marketplace pays out a large share to its expert contractors — net economics are far thinner than the topline suggests. Third, disintermediation: the same labs that are Mercor’s best customers are also its most credible competitors, since building an expert-recruitment pipeline in-house is hard but not impossible.

None of those risks have slowed investors so far. If General Catalyst leads and Nvidia joins, Mercor will have doubled its valuation in under a year while converting from a recruiting-adjacent startup into what may be the purest public bet on the thesis that, in the frontier-model era, the scarcest resource is not compute or capital — it is qualified humans willing to teach the machines.