Mercor Raises $750M Series D for Agent Evals and Post-Training at $20B
Mercor says it spends three times as much on inference as on employee salaries. That one line from CEO Brendan Foody's Series D post tells you more about where the company thinks work is going than the round size does.
The round is $750 million at a $20 billion valuation, announced Nov 10. No lead investor is named. Part of it is an employee tender, and Mercor doesn't say how much of the $750 million is new money versus shares bought from staff. Participants 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. The money goes to infrastructure and research for evals and post-training, and to building more environments where agents can practice complex workflows.
The business started as a network of experts paid to train frontier models, now more than five million people. The pitch has moved one step down the stack: take the same expert network, eval methods and post-training setup that the labs buy, and sell it to enterprises that want to start from an open model and make it their own. The Deeptune acquisition earlier this year fits that, since its team spent two years recreating hundreds of enterprise apps as practice grounds for agents. Mercor also runs the APEX-Agents leaderboard and opened offices in New York and London.
The bet is clear and a little self-serving: if every company ends up post-training its own models, someone has to supply the humans who know what good work looks like and the evals that prove it. Mercor sitting on both the lab side and the enterprise side is the real moat, because what it learns grading frontier models feeds what it sells to Fortune 500 buyers, and the other way round. The weak spot is the tender. A $20 billion headline with an undisclosed primary amount is a valuation signal more than a war chest, and the "double-digit GDP growth" extrapolation from Mercor's own inference bill is a CEO's argument, not a forecast anyone should bank.
Source: https://www.mercor.com/blog/series-d/
← Back to all articles
The round is $750 million at a $20 billion valuation, announced Nov 10. No lead investor is named. Part of it is an employee tender, and Mercor doesn't say how much of the $750 million is new money versus shares bought from staff. Participants 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. The money goes to infrastructure and research for evals and post-training, and to building more environments where agents can practice complex workflows.
The business started as a network of experts paid to train frontier models, now more than five million people. The pitch has moved one step down the stack: take the same expert network, eval methods and post-training setup that the labs buy, and sell it to enterprises that want to start from an open model and make it their own. The Deeptune acquisition earlier this year fits that, since its team spent two years recreating hundreds of enterprise apps as practice grounds for agents. Mercor also runs the APEX-Agents leaderboard and opened offices in New York and London.
The bet is clear and a little self-serving: if every company ends up post-training its own models, someone has to supply the humans who know what good work looks like and the evals that prove it. Mercor sitting on both the lab side and the enterprise side is the real moat, because what it learns grading frontier models feeds what it sells to Fortune 500 buyers, and the other way round. The weak spot is the tender. A $20 billion headline with an undisclosed primary amount is a valuation signal more than a war chest, and the "double-digit GDP growth" extrapolation from Mercor's own inference bill is a CEO's argument, not a forecast anyone should bank.
Source: https://www.mercor.com/blog/series-d/
Comments