Mercor, the San Francisco startup that pays experts to train and test AI models, has raised a $750 million Series D at a $20 billion valuation, TNW reported on Nov 10. That doubles the $10 billion valuation Mercor reached when it raised $350 million in October 2025, and the round also included an employee tender offer, according to a blog post by co-founder Brendan Foody.
For the Bay Area, this round matters for 2 groups at once. Engineers, lawyers, doctors and other specialists are the supply side of Mercor's business, and the companies building AI agents are the demand side Mercor now wants to reach. Below is what the company announced, how its model works, and what the deal suggests for tech workers and AI teams here.
Mercor Series D: the round in numbers
Per TNW, Benchmark, Felicis, Fundomo, General Catalyst, Nvidia, Salesforce, Sapphire Ventures and Zoom joined the round, along with funds managed by affiliates of Apollo, G Squared, Glade Brook Capital, StepStone Group and XTX Ventures. Mercor did not say how much of the $750 million went to the employee tender.
The new valuation was signaled months ago. In July, TNW reported that Mercor had told investors it could raise at $20 billion, citing Bloomberg, on the same day it announced it was buying Deeptune, a startup that builds simulated software where AI agents practice real work. Terms of that deal were not disclosed.
How Mercor's expert network works
Mercor started as an AI hiring platform. It now runs a network of more than 5 million experts who are paid to train and evaluate AI models using their professional knowledge, according to TNW. The company describes itself on its website as organizing human intelligence to power the AI economy, and so far its main buyers have been frontier AI labs.
The product behind the new pitch is the eval. Mercor builds custom tests from each company's real workflows, then uses the results to post-train models on that company's work, often starting from an open model. Deeptune adds practice environments: it recreated hundreds of enterprise applications so agents can rehearse tasks before they touch production systems.
Co-CEO Adarsh Hiremath put the reasoning plainly, saying a model "can score well on a public benchmark and still fail" on the systems and edge cases of a specific business. If you've watched an agent ace a demo and stumble on your actual CRM, you already know the problem he is describing.
Why the new money targets every company
The stated plan is to sell the same training and evaluation service beyond the labs. "Every job in the economy will need its own AI agents," Foody said, and every company will need a way to train and evaluate them on its own work. Foody also shared one internal figure: Mercor already spends 3 times as much on inference, the cost of running AI models, as it does on employee salaries.
Mercor opened offices in New York and London this year, according to Foody's post. That expansion, combined with the enterprise pitch, points to a sales motion aimed at large companies with complex internal tools. For wider context on where AI spending is heading, see our roundup of the biggest AI trends of 2026.
What it means for Bay Area tech workers and AI teams
For specialists, the round is a signal that paid expert work for AI training is a growing category, with a well-funded buyer behind it. Senior engineers, product people and domain experts in Palo Alto, Mountain View and San Jose are exactly the kind of knowledge these networks recruit. If you're considering that kind of side work, read the terms carefully, ask how your contributions are used, and treat it as consulting rather than a job.
For founders building agents, the message is that customers will increasingly ask how you test your product on their workflows, and a public benchmark score will not close that question. Building your own evals early, from real customer tasks, is cheaper than retrofitting them after a failed pilot. Our AI product trust and design checklist covers the user-facing side of the same problem, and our post on product-market fit red flags for AI startups explains why pilot results matter more than demo metrics.
For operators, the 3-to-1 inference-to-salary figure is worth sitting with. If compute keeps growing faster than headcount at AI-native companies, budgets, pricing and even org design change with it, and it connects directly to the power and data center buildout we covered in our jet engines and AI power post.
Questions like these are a good fit for small in-person rooms. Tribe hosted an AI Infra Dinner in Palo Alto on Sep 30 with Prateek Joshi, a table of 12 to 15 seats. Tribe also organizes South Bay Renaissance to rebuild the startup and social scene of the South Bay and Peninsula in person, one gathering at a time, so the people doing this work can meet without driving to the city.
What to watch next
3 things will show whether the enterprise pitch lands. First, watch for named enterprise customers outside the AI labs, since that is the market the round is meant to open. Second, watch how Deeptune's practice environments show up in the product. Third, watch the rest of the AI data and evaluation category, where TNW has pointed to rivals such as Scale AI and Surge AI, since each new round adds a data point on how much companies will pay to make agents reliable.
This post is market context only, with no investment view on Mercor. The round is a useful marker for where AI spending is heading and for the kind of skills Bay Area workers can expect to be in demand. For the wider picture of who else reached a $1 billion valuation last month, see our roundup of October 2026 new unicorns in the Bay Area.
Talk AI with people who build it
If you work on agents, evals or AI infrastructure, the Tribe Events calendar on Luma is the easiest way to find the next room worth being in, including South Bay Renaissance: Designing AI for humans in Redwood City on Nov 19. Follow the calendar to see new events first, and join the AI Trends Tribe to keep the conversation going between meetups.
