Mercor raises $750M at a $20B valuation to sell AI evals to enterprises
The AI training-data startup doubled its valuation in about a year and wants to sell the evals, environments and post-training it built for frontier labs to everyday enterprises.
Mercor, the San Francisco startup that pays domain experts to train and test AI models, has raised a $750 million Series D at a $20 billion valuation, the company said on Tuesday. The round doubles the $10 billion price it set a little over a year ago and comes with a shift in pitch: Mercor now wants to sell to ordinary enterprises the same evaluation and post-training work it has been doing for frontier AI labs.
Investors in the round include 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. Co-founder and CEO Brendan Foody wrote that the round also included an employee tender, so part of the money buys shares from staff rather than funding the company's balance sheet. Mercor did not disclose how the $750 million splits between new capital and the tender.
From hiring marketplace to evals vendor
Mercor started as an AI-driven hiring platform and pivoted to supplying specialists such as scientists, doctors and lawyers to AI labs for model training, charging an hourly finder's fee, TechCrunch reported when the company raised a $350 million Series C at a $10 billion valuation in October 2025. Felicis led that round and the $100 million Series B before it, which valued Mercor at $2 billion. Mercor's growth picked up after labs including OpenAI and Google DeepMind reportedly cut ties with rival Scale AI following Meta's roughly $14 billion investment in that company.
The company now says more than five million experts are in its network. Its argument for the new round is that every profession has workflows no model has been tested against. Foody pointed to more than 800 distinct occupations in the United States, and said agents need two kinds of knowledge to work in production: how a task is done in general, and how a particular company does it through its own systems, policies and quality bar.
In practice, Mercor builds custom evaluations, or evals, with each customer and experts in its field, based on the company's actual workflows. The results show how models perform against that company's requirements and where more training, or practice in realistic simulated environments, could improve them. That training step, known as post-training, tunes an existing model on the company's own data and expertise. Foody wrote that many enterprises can start from a capable open model and post-train it "with a fraction of the resources required at the frontier."
Environments and benchmarks
The other half of the plan is training environments, simulated software where an agent can practice a job and be scored on the result. In July Mercor agreed to acquire Deeptune, whose team had spent two years recreating hundreds of enterprise applications, from spreadsheets to Salesforce, as reinforcement-learning environments. Andreessen Horowitz had led Deeptune's $43 million Series A earlier in the year, and Mercor was already one of its customers. In Mercor's framing, each environment has three parts: the software, the tasks an agent must finish, and verifiers that judge whether it succeeded. Mercor's experts write the tasks and verifiers, and Deeptune's engineers build the software.
Mercor also publishes APEX, a family of benchmarks for economically valuable work. APEX-Agents, for example, has 240 multi-hour tasks in investment banking, corporate law and management consulting that agents complete using real Google productivity apps. On Tuesday its leaderboard showed Gemini 4 Carbon High on top at 84.2%, while Anthropic's Fable 5.2, released on Monday, scored 80.1% at its Max setting.
A bumpy year, then a doubling
The new valuation was not a surprise. In July, Bloomberg reported that Mercor was in early talks at $20 billion and had already received a term sheet at that price, while Foody said on X that the company's annualized revenue run rate had passed $2 billion, double the level of four months earlier. TechCrunch also noted that Mercor suffered a data breach earlier in 2026 and that several contract workers filed lawsuits, according to Business Insider.
The raise lands in a month of very large private rounds for AI infrastructure companies, including inference provider Fal's $1.3 billion Series F last week. Mercor's bet is different from a compute play: it sells human judgment, packaged as tests and training data, to companies that want agents doing real work. In an interview on CNBC's Squawk Box on Tuesday, Foody said the funding will accelerate that enterprise business.
Mercor says it has opened offices in New York and London this year and will put the new funding into research and infrastructure for evals and post-training, and into building more environments where agents can learn complex workflows. Foody also wrote that the company already spends three times as much on inference as it does on employee salaries.
