News

Mercor raises 750 million dollars at a 20 billion dollar valuation to evaluate AI on real work

The Series D includes an employee share tender. Mercor is expanding enterprise evaluations and practice environments, without disclosing revenue or the split between new capital and share sales.

Illustration of documents and work platforms along a practice path leading to a glowing cube
Original AI-generated illustration of practice and evaluation environments. It does not depict Mercor or Deeptune interfaces.

Mercor announced a 750 million dollar Series D on November 10, 2026, valuing the company at 20 billion dollars. The business, which connects human expertise with AI training and evaluation, plans to extend its infrastructure to enterprises that want to test agents against their own work, beyond general benchmarks.

According to Mercor, the round also included an employee share tender. The company did not disclose the split between purchases of existing shares and new capital entering its balance sheet. The full 750 million dollars should therefore not automatically be treated as a new research or expansion budget.

Reuters reports that the valuation doubles Mercor's approximately 10 billion dollar valuation in October 2025. That change describes the pricing of investment rounds. It does not itself demonstrate a doubling of revenue, customer numbers or model quality. The current funding announcement does not disclose revenue or an annualized revenue run rate.

Testing the work a company actually does

Mercor describes infrastructure for training and evaluating models on professional tasks with help from people who know the field. Its enterprise offering includes custom evaluations built around customers' workflows. The question goes beyond a model producing a good answer: can it finish the task within the company's systems, requirements and quality standards?

A conceptual example illustrates the distinction. An ability to draft a document does not establish that an agent can select the correct record, use the current version and send the result to the appropriate approver. This is an illustration of the gap between general skill and a defined business process, rather than a result from a Mercor experiment.

Custom evaluation therefore begins by defining success. Teams must decide what counts as a correct completion, what information may be used and which errors require a stop. Without those definitions, a high score may be difficult to interpret: does it measure what the company needs, or simply what was easiest to test?

Experts contribute more than answer sets

Mercor says its expert network includes more than 5 million people. This is the company's description of its network, rather than a count of salaried employees or people working on a particular project at any given time. The distinction matters when the network's size is used to explain the scale of the business.

Domain expertise involves more than knowing the right answer. Professional work includes sequences of actions, exceptions, judgment and company-specific requirements. In Mercor's account, evaluations translate those requirements into testable tasks and identify where a model needs additional training or practice.

For a customer, network size cannot replace questions about the experts' suitability. Who sets the criteria, how are disagreements between evaluators resolved, and how does the test change when the workflow changes? These aspects of quality cannot be inferred from participant numbers alone, even with a very large network.

Deeptune adds somewhere for agents to practise

Mercor says it acquired Deeptune this year. Its team recreated hundreds of enterprise applications as environments where agents can practise workflows. The combination is intended to connect professional knowledge and evaluations with a setting that reflects where the work actually takes place.

This introduces the distinction between evaluation and training. Evaluation measures how a system performs against a requirement; further training seeks to change its capabilities. A practice environment allows a task to be run again, its outcome assessed and feedback provided. That does not mean every repetition improves performance, or that success in a simulation guarantees success in the real system.

Customers therefore need to examine how closely a training environment resembles daily work. Differences in permissions, data or interface behavior can change the task. When assessing such a system, organizations should define which conditions must remain consistent and which gaps require additional testing before an agent receives broader responsibility.

Many investors, without an announced lead

Mercor's list of participants includes 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. The company does not name a lead investor in its announcement.

The presence of technology companies on that list does not announce a new commercial agreement, product integration or distribution partnership. The announcement identifies them as investors. Separating those roles avoids reading too much into familiar names, particularly in a round combining investment firms with businesses active in AI themselves.

Mercor says the financing will accelerate investment in evaluation and post-training infrastructure and research, alongside more practice environments. The company has also opened offices in New York and London. These are its stated expansion directions, but the announcement provides no budget breakdown by location or development area, nor a detailed delivery schedule for each capability.

Running costs become part of the quality question

In the funding announcement, founder Brendan Foody writes that Mercor spends 3 times as much on model inference as on employee salaries. This is a comparison made by the company itself. It does not disclose the absolute amounts alongside that ratio, so the figure alone cannot be used to calculate its expenses or profitability.

In practical terms, the comparison draws attention to AI costs after a model has been trained. An organization evaluating an agent also needs to ask what completing a task costs, how many attempts it takes and how much human work remains afterward. A good result requiring extensive corrections can be very different from the same result delivered consistently.

Mercor's funding supports an effort to turn those questions into infrastructure companies can use. Managers still have an initial job of their own: select a defined process, establish success and cost measures, and test the agent against real work. The new valuation describes investor expectations; the evaluations will establish what the system can do for a customer.