Confirmed 10 Nov 2026 · Labs A one-minute read

Mercor raised $750 million at $20 billion for the human experts behind AI training.

On November 10 Mercor said it raised $750 million at a $20 billion valuation. Caveat: it published no revenue, so the price rests on what buyers of expert work will pay.

Until now

Frontier progress is bought with chips, data centres and electricity.

Now

A large share is bought from people paid to judge work.

Why it matters

Mercor sells evaluations and post-training work: humans building the tests and examples that agents learn from. Nvidia sells the chips and now owns part of the layer above them. The price says capability is becoming a labour market, and that whoever holds the experts holds a bottleneck money alone cannot clear.

Established

Mercor announced $750 million at a $20 billion valuation, including an employee tender.
Nvidia and Salesforce are named investors, alongside Benchmark, General Catalyst and Apollo funds.
Mercor says more than 5 million experts sit in its network.

Not established

Revenue. Mercor published no revenue, run rate or growth figure.
What experts are paid. No payout total or hourly rate is disclosed.
That the valuation doubled. The prior figure is reported, not in the post.

The five-minute readThe record, the timeline, the weight and 3 sources

The record

Mercor published a post titled "Organizing human intelligence for the entire economy" at mercor.com/blog/series-d, dated November 10, 2026, which we read there. It announces a Series D of $750 million at a $20 billion valuation and says the round included an employee tender offer. The named investors are Benchmark, Felicis, Fundomo, funds managed by affiliates of Apollo, G Squared, General Catalyst, Glade Brook Capital, Nvidia, Salesforce, Sapphire Ventures, StepStone Group, XTX Ventures and Zoom. Mercor describes its business as organising human expertise for AI: evaluations, post-training, and environments in which agents practise complex workflows. The post says more than 5 million experts are in Mercor's network and cites more than 800 distinct United States occupations. It says Mercor spends 3 times as much on inference as on employee salaries, and argues that if that ratio holds as AI spreads the economy could reach "closer to double-digit GDP growth" over 5 years, which it calls nearly 3 times the usual rate. The money is said to fund infrastructure and research for evaluations and post-training, and more environments for agents to learn in. The post records the acquisition earlier this year of Deeptune, whose team spent 2 years recreating hundreds of enterprise applications for agents to practise on, and says Mercor opened offices in New York and London this year. Co-founder and chief executive Brendan Foody is quoted saying "AI is only as useful as the human expertise behind it" and that for companies "owning their intelligence is essential." Not in the post: revenue, run rate, growth rate, margin, what Mercor pays experts in total or per hour, how many of the 5 million are active, which AI laboratories are customers, the prior valuation, and the terms of the employee tender. This card is dated November 10 and surfaced late. The post carries that date, trade coverage ran the same day, and no edition had carded it.

How it got here

Sep 12
Anthropic said it would give outside inspectors desks, badges and company laptops, and OpenAI said the same. The first sign that judging a model had become a staffed job rather than a script. Edition 020.
Oct 9
Anthropic cut live internet access to all internal evaluations after its models filed real forms. Evaluation environments became infrastructure with a failure mode. Edition 046.
Nov 5
METR, an outside evaluator, measured Anthropic's progress at 1.9 times its old pace. Edition 076.
Nov 10
Mercor announces $750 million at a $20 billion valuation, names Nvidia and Salesforce among its investors, and says more than 5 million experts are in its network.

Who has to react

The frontier laboratories that buy evaluation and post-training work, and now bid against each other for the same experts; Nvidia, which sells the chips and now owns part of the layer above them; Salesforce, whose agents need the environments Mercor builds; and the 5 million people on the network, whose rates this round repriced.

What to watch

Would strengthen itA frontier laboratory discloses annual spending above $1 billion on human evaluation or post-training data, or a second expert-data company raises at above $10 billion, by December 31, 2027.
Would weaken itMercor is marked below $20 billion in a disclosed secondary or later round, or a frontier laboratory publicly replaces human evaluation with model-generated evaluation, by December 31, 2027.
Would strengthen itNvidia or Salesforce discloses a commercial agreement with Mercor, or takes a board seat, by September 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentHigh

ReversibilityLow

PersistenceYears

Cost is medium: the capital is the investors', not Mercor's, but a round this size carries dilution and an employee tender that moves cash across the cap table. Commitment is high because named investors including Nvidia and Salesforce signed for equity, which is public and hard to disown. Reversibility is low: shares bought cannot be unbought, and only the valuation can fall. Persistence is years, because the money funds evaluation infrastructure and agent environments that take years to build, and a network of 5 million people decays slowly. We do not score it higher: no revenue was published, so the valuation rests on buyers rather than on disclosed trade.

The words

Post-training: The work done after a model is first trained on text: teaching it with worked examples and human ratings so that it behaves usefully.
Evaluations: Tests built to measure what a model can and cannot do. Writing a good one needs someone who knows the job being tested.
Employee tender: A funding round that also buys shares from staff, giving them cash without waiting for a listing.
Inference: The cost of running a model to produce an answer, as opposed to the one-off cost of training it.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.