# Mercor raises $750M at a $20B valuation to sell its AI training and evaluation playbook to companies

> The company behind a network of more than five million paid experts says it now spends three times as much on inference as on salaries.

- URL: https://devisdead.ai/news/mercor-raises-750m-at-a-20b-valuation-to-sell-its-ai-training-and-evaluation-pla
- Published: 2026-11-10T16:16:55Z
- Author: DevIsDead Newsroom
- Section: Startups & Money
- Tags: mercor, funding, evaluations, post-training, agents
- Originally reported by Unite.AI (https://www.unite.ai/mercor-raises-750m-series-d-at-20b-valuation-to-expand-ai-training/).

Mercor, which pays human experts to train and test frontier AI models, has raised a $750 million Series D at a $20 billion valuation, according to [Unite.AI](https://www.unite.ai/mercor-raises-750m-series-d-at-20b-valuation-to-expand-ai-training/). The round includes an employee tender, and NVIDIA, Salesforce and Zoom are among the investors, along with Benchmark, General Catalyst, Felicis and others.

## Where the money goes

Mercor says the round pays for infrastructure and research in two areas: evaluating models and post-training them, meaning extra training on top of an existing model. It also funds more environments where agents can practice complicated workflows. This year Mercor opened offices in New York and London, its first outside San Francisco, to be nearer to the labs and enterprises it sells to.

## From the labs to the enterprise

For two years Mercor has built tooling to train and grade frontier models on real work, using people with domain expertise. It's now pointing that at businesses. Its custom evaluations use a company's own workflows to show what an agent can do, what it costs and where more training would help.

The pitch is that knowing how a task is done in general isn't the same as knowing how one employer does it, with its systems, rules and quality bar. Mercor's enterprise recipe starts from a capable open model, builds public evaluations, then uses post-training to bring costs down before training models for a specific customer. It says post-training takes a small fraction of what frontier work does.

A few numbers from the company:

- More than five million experts in its network, paid to train frontier models.
- More than 800 US occupations with workflows nobody has evaluated models against yet.

## The Deeptune piece

Mercor bought Deeptune, a reinforcement-learning environment company, earlier this year; the deal was announced on July 9. Deeptune spent two years recreating hundreds of business apps, from spreadsheets to Salesforce, as places where agents can train. Each environment has the software, a set of tasks and verifiers that decide whether the agent succeeded. Mercor's experts write the tasks and verifiers.

## Its own AI bill

Mercor says it spends three times more on inference than on employee salaries, and that the spending pays off by making staff better at their jobs. CEO Brendan Foody argued that if that ratio spreads across the economy, GDP growth over the next five years could approach double digits, nearly three times what he called the usual rate.

**[Scoreboard](/scoreboard) update:** Draw. Humans train the models, and the models get the bigger budget line.

Stay undead.