Mercor Raises $750M at $20B Valuation to Expand AI Training

Mercor announced a $750 million Series D on November 10, 2026, at a $20 billion valuation, one of the largest financings this year for a company built on training and evaluating AI models with human expertise. NVIDIA, Salesforce, Zoom and other investors participated in the round, which also included an employee tender.

Benchmark, Felicis, Fundomo, funds managed by affiliates of Apollo, G Squared, General Catalyst, Glade Brook Capital, Sapphire Ventures, StepStone Group and XTX Ventures also joined. The capital will accelerate infrastructure and research for model evaluations and post-training (additional training applied to an existing model), while funding more environments where AI agents can learn complex workflows.

The round lands amid a broader shift in how enterprises adopt AI. Buyers no longer ask only whether a model is capable in the abstract; they ask what it can do inside their own systems, what it costs and how its performance improves. Mercor is positioning itself as the measurement layer for exactly that question.

Why does a $20 billion valuation hinge on evaluations?

Mercor has spent the past two years building infrastructure to train and evaluate frontier models on real-world work using human expertise. The company is now extending those capabilities to enterprises, with evaluations designed to establish three things: what agents can do, what they cost and how their performance can improve.

Its custom evaluations use an employer's actual workflows to measure models against company requirements and to identify where additional training, or practice in realistic environments, could help. The approach draws a sharp line between two kinds of knowledge: knowing how a task is generally performed, and knowing how a particular employer handles it under its own systems, policies and quality standards.

Mercor's enterprise playbook follows a distinct sequence:

• Start with a capable open model
• Build public evaluations
• Use post-training to reduce costs
• Train models for customers

The company describes post-training as requiring a fraction of the resources needed at the frontier. That cost asymmetry is the commercial argument: enterprises get models tuned to their workflows without paying frontier-scale compute bills.

How large is the expert network behind it?

Mercor reports that its network includes more than five million experts paid to train frontier AI models. Those experts generate the tasks, judgments and learning signals that both frontier labs and enterprise customers rely on.

The company also frames the market opportunity in occupational terms. The United States has more than 800 distinct occupations containing workflows against which models have yet to be evaluated, according to Mercor. That figure defines the remaining whitespace: a vast inventory of professional work that no one has systematically measured AI against.

Mercor's frontier-lab work and its enterprise business draw on the same expert network, evaluation methods and post-training infrastructure. Lessons from each side inform the other, which lets the company reuse what it learns from frontier labs inside customer deployments.

What does the Deeptune acquisition add?

Earlier in 2026, Mercor acquired Deeptune, a reinforcement-learning (a training method where models learn from reward signals) environment company. The July 9, 2026 announcement outlined the planned acquisition and Deeptune's two years of work recreating hundreds of enterprise applications, from spreadsheets to Salesforce, for agent-training environments.

Those environments let agents practice professional workflows before deployment. Each environment has three components:

• The software where work occurs
• Tasks defining what the agent must accomplish
• Verifiers determining whether it succeeded

Mercor's experts create the tasks and verifiers that translate human judgment into learning signals. The July announcement says its APEX system measures model performance on economically valuable tasks across thousands of real-world workflows, while Deeptune supplies the software layer needed to turn that expertise into training environments.

The combination matters because evaluation without a realistic environment produces abstract scores, not operational readiness. Deeptune supplies the simulated workplace; APEX supplies the measurement; the expert network supplies the judgment.

Where is Mercor expanding?

Mercor opened New York and London offices in 2026, expanding beyond its San Francisco base. The company said the move places it closer to enterprise and laboratory customers and supports employee recruitment.

The geographic expansion tracks the commercial logic of the business. Custom evaluations require deep engagement with a customer's actual systems and policies, work that benefits from proximity to enterprise buyers concentrated in those financial capitals.

What does Mercor's own AI spending signal?

Mercor reports that it spends three times as much on inference (the computing used to run models) as on employee salaries. The company says this spending pays for itself by making employees better at their jobs.

Co-founder and chief executive Brendan Foody took that internal ratio and scaled it to the macroeconomy. He said that if the ratio holds as AI spreads through the economy, it points to something closer to double-digit growth in gross domestic product over the next five years, nearly three times the rate he described as customary.

The claim is a projection, not a forecast with a published model behind it, but it signals how Mercor's leadership frames the stakes. If inference spending becomes the dominant labor-adjacent cost across the economy, the productivity effects would be structural rather than incremental.

What is the actual test for enterprise agents?

The funding supports a single enterprise proposition: build the environment, supply the experts and measure whether an agent is ready. For Mercor, the practical test is reliable performance inside a customer's own workflows, where general task knowledge alone is insufficient.

That standard separates Mercor's pitch from benchmark-leaderboard marketing. A model that scores well on public evaluations can still fail inside a specific company's systems, policies and quality standards. Mercor's bet, now backed by $750 million at a $20 billion valuation, is that enterprises will pay for the infrastructure that closes that gap, and that the five-million-expert network plus Deeptune's simulated environments constitute that infrastructure.

The next signal to watch is whether enterprise customers adopt Mercor's playbook of open models plus post-training at scale, which would validate the cost thesis that frontier training is unnecessary for most corporate workloads.

By Marcus Bennett