Executive Summary

Key Numbers

Four of them. The first three show what this round bought, and the last one tells you how ready the companies doing the buying are.

Sources: Mercor, "Series D" · SiliconANGLE · Cisco 2026 Data and Privacy Benchmark Study

Mercor said on November 10 that it had raised $750 million in a Series D, and the valuation attached to it is $20 billion. The Series C in October 2025 was $350 million at $10 billion, so the price has doubled in a little over a year. Existing investors came back in, Felicis among them, which led that earlier round, and no lead was named this time. The announcement states that the round included an employee tender alongside the new money. How much of the total that accounts for was not disclosed.

The names that stand out are the new ones. Nvidia, Salesforce and Zoom appear on the investor list, next to funds managed by affiliates of Apollo, Sapphire Ventures, StepStone Group, G Squared, Glade Brook Capital, XTX Ventures and Fundomo. Benchmark and General Catalyst, both returning, are on the list as well. Salesforce and Zoom do not build models; they sell the software companies work inside. They are also the owners of the screens where an agent will be expected to do the work.

What the company puts forward as its asset is people. The announcement puts the network at more than five million experts. In the same post the company notes that the United States alone holds more than 800 distinct occupations, each carrying workflows that no model has been evaluated against yet. Mercor started in 2023 by connecting AI labs with specialists in fields such as law, finance and medicine. In July, Foody said annualized revenue had passed $2 billion as of June.

The parts for this shift came in four months ago. Deeptune was a New York company led by Tim Lupo, and what it built were reinforcement learning environments that copy the software knowledge workers use every day, which the company called training gyms. An agent practices real tasks inside mock spreadsheets and Salesforce queues. Two years went into rebuilding hundreds of enterprise applications that way.

The sequence does not look like coincidence. Deeptune raised a $43 million Series A in March 2026 led by Andreessen Horowitz, and Foody came into that round as an angel with his own money. Four months later, on July 9, Mercor acquired the company. The price was not disclosed. Asked by Fortune whether an acquisition was on his mind when he wrote the angel check, Foody answered that in a lot of ways it was the main motivation.

Deeptune was not the first purchase of the year either. Five months before it, on February 6, 2026, Mercor acquired Sepal AI. That price went undisclosed as well, and the record of the deal is an announcement from Orrick, the law firm that advised it. Mercor described Sepal as a team working at the intersection of human data, reinforcement learning environments and specialized research projects. February and July, twice in a single year, and in both cases the target was a company building the places where agents practice.

Put the two pieces together and the job changes. An expert network on its own lets people grade what a model answers. A work environment on its own gives an agent somewhere to practice. Overlay them and you can run an agent inside a copy of one firm's operations and have someone from that trade grade the output, all in one package. What this round bought is the business of selling that package one company at a time.

In the announcement the company writes that professional work depends on "two kinds of knowledge that no model inherently has: how a task gets done and how a specific company does it according to its systems, policies, or quality bar." The first one lives with the experts in that trade. The second exists nowhere but inside the company.

A sentence from co-founder Adarsh Hiremath points at the second. A model can score well on public benchmarks and still fail on "the systems, policies and edge cases that define how a specific business runs." The sentence stops short of calling the benchmark wrong. It names the place the benchmark never asked about.

Foody's sentence takes that diagnosis and pulls a market out of it: "Every job in the economy will need its own AI agents, and every company will need a way to train and evaluate them on its own work." Read backwards, it declares that there have to be as many evaluation standards as there are companies, and $20 billion is the price now attached to it.

What a public benchmark is sits in the name. The questions are public. Whoever builds a model can read them and tune against them, and while the score rises, that score is not what confirms the underlying work ability rose by the same amount. That is less a reason to distrust the number than a boundary on it: what the score measures does not widen past that test.

Company work lies outside that boundary. Take the same invoice. Which amount triggers whose approval, how far an exception gets handled on the floor before it is escalated, which output is considered good enough to send — all of it differs from firm to firm. A fair share of those rules is not written down anywhere. It survives as the judgment of people who have been there a long time. Knowing whether an agent is doing well means checking against exactly that judgment, and a public benchmark holds no such answer key.

The firm making this argument about public tests has been setting public tests itself. Mercor released APEX on October 1, 2025, describing it as "a first-of-its-kind benchmark that evaluates AI models based on their ability to perform economically valuable knowledge work." Roughly 100 working experts across four professions — investment banking, management consulting, big law and primary care — wrote the questions and the rubrics, and those rubrics carry an average of 29.09 separate criteria. At release the company wrote that none of the models met the production bar for automating real-world tasks in the four professions it looked at, and the scores are still posted on a leaderboard on its own site.

So this announcement is not a claim that public tests are useless. It is a claim that even with the public test it built itself included, one box remains that nothing there reaches. From the first APEX release in October 2025 to this announcement, the exam has moved like this.

▲ Sources: Mercor, "APEX Benchmarks" · Mercor, "Introducing APEX" (2025-10-01) · Mercor, "Series D"

That missing answer key is the thing Mercor has written down as its product. It will work with enterprises and their experts to build custom evals based on a company's actual workflows, and keep alongside them the infrastructure to post-train models until they pass. That the party writing the evaluation and the party getting models through it are one firm deserves its own note. What happens on a public benchmark once the test and the candidate sit close together is something this arrangement can repeat somewhere new.

This name has passed through the blog before. Once in the story of the labeling industry moving its contest to reinforcement learning environments, and again in the story of AI training-data supply pooling into four companies. What those two pieces looked at was the shape of a supply chain. What differs here is that the customer changes. Once the customer comes down from a few labs to individual firms, the number of answer keys that have to be written rises to the number of firms.

Writing an answer key takes material. Records of work that was actually processed, outputs that were sent back and the reasons given, traces of how an exception was judged. These sit among the innermost holdings a company has, and they are also the ones a competitor would be glad to see. The more accurate the evaluation, the deeper that material has to go. Validity and sensitivity move in the same direction, and that is the structural tension in this business.

Readiness among the companies that would commission this work is not yet equal to that tension. In a study Cisco published in 2026, built on answers from 5,200-odd IT and security workers across 12 markets, 12% said their AI oversight body was mature and proactive. Seventy percent admitted they were exposed to risk by putting their own or their customers' data into AI training, and 55% pin ownership, liability and indemnity for that data into contract terms. More than four in ten, in other words, hand it over without writing any of it down.

So before evaluation is outsourced, three questions come ahead of any performance metric.

• • Who writes the answers? Having our own people write the standard that decides what counts as correct in our work is one thing; having an outside expert write it after reading our material is another. In the second case, that person's judgment becomes our quality bar.
• • Where does the finished answer key sit? An evaluation set and the answers attached to it are an asset you keep using once it exists. Where it is stored, the conditions under which it is deleted, and what becomes of it when the contract ends all have to be on paper.
• • How far does reuse extend? Whether the environment and the answer key built out of our work go on to sharpen another customer's model, and how thoroughly they are stripped if they do, is a line the contract has to draw.

Seen from where AI-Ready Data is the daily job, this round reads as a signal that evaluation data has moved out of the cost column and into the asset column. Once something is an asset, its ownership has to be written down. A model can be swapped out. An answer key with our own work recorded in it cannot.

Thank you for reading this far. The full announcement is on Mercor's blog. At your own company, whose hand holds the pen on the standard that decides whether an agent did the job well?

References

Primary sources

• 1.Mercor. (2026). "Series D." Mercor Blog, 2026-11-10.
• 2.Mercor. (2025). "Introducing APEX: The AI Productivity Index." Mercor Blog, 2025-10-01. On benchmark composition and rubrics, see the APEX Benchmarks methodology.
• 3.Orrick. (2026). "Mercor Acquires Sepal AI." 2026-02-06. Announcement by the law firm that advised the deal (price undisclosed).
• 4.Cisco. (2026). "A Shifting Paradigm: Governance in the Age of AI — Cisco 2026 Data and Privacy Benchmark Study." Survey of 5,200-odd respondents across 12 markets (fielded September 2025).

Press

• 5.Pulse 2.0. (2026). "Mercor: $750 Million Series D Funding At $20 Billion Valuation To Expand Enterprise AI Training And Evaluation."
• 6.SiliconANGLE. (2026). "Mercor raises $750M at $20B valuation to take AI agent evaluation to enterprises." 2026-11-10.
• 7.SiliconANGLE. (2026). "Mercor buys Deeptune to build training environments for AI agents." 2026-07-09.
• 8.FinSMEs. (2025). "Mercor Raises $350M in Series C; Valued at $10 Billion."