Mercor is not a household name, but it is quietly running one of the most consequential operations in AI right now. The company recruits professionals, such as doctors, lawyers, accountants, and software engineers, and deploys them as human reviewers and graders for AI labs. Those professionals judge whether an AI's output is correct, nuanced, and good enough to train on. OpenAI, Anthropic, and most of the major AI developers pay Mercor to supply this expert labor at scale.
This week, the company said its annual revenue has crossed $2 billion, up 100% from four months ago. It is also in talks to raise around $500 million at a $20 billion valuation, double what it was worth in October last year. The deal has not closed, and terms could change, but at least one term sheet has reportedly already been signed.
The revenue figure is real but needs context. Mercor keeps roughly 30 to 40 cents of every dollar it bills. The professionals who do the actual training work take home the rest. So the company's own net revenue sits closer to $600 to $800 million. At a $20 billion valuation, that still implies a very high multiple, but not an extraordinary one for a company growing this fast.
A big part of Mercor's rise traces back to a single event in mid-2025. Meta paid $14.3 billion for a near-half stake in Scale AI, which was then the dominant company in this space. Within days, Google, OpenAI, and other major AI labs began distancing themselves from Scale, worried that a company half-owned by Meta could not be trusted with their proprietary training data. That business walked straight to competitors including Mercor, which went from roughly $75 million in annual revenue in early 2025 to over $1 billion by late in the same year.
On top of the funding news, Mercor also acquired Deeptune, a startup that builds what its founders call training gyms: simulated digital workplaces where AI agents can practice real tasks without touching live business systems. Think of it as a flight simulator for AI. An agent can attempt to process an invoice in a fake accounting system, or respond to a customer ticket in a fake CRM, make mistakes, and learn from them before ever going near a real company's data.
The acquisition is strategically coherent. Mercor already has the human experts who write tasks and grade AI outputs. Deeptune provides the simulated environments where AI agents actually practice. Together, Mercor wants to own the full process: building the practice floor, staffing the graders, and certifying whether an AI is ready to do real work. It is moving from a staffing marketplace to something closer to a readiness platform for enterprise AI.
There is a governance detail worth noting. Mercor's CEO personally invested in Deeptune's $43 million funding round in March, just three months before his company bought it. He has said the investment was made with the acquisition already in mind. That is legal and perhaps strategically clever, but it raises a straightforward question about whether the company's board had full visibility before the personal check was written.
For business operators outside the AI industry, the practical takeaway is this: the professionals being recruited to train AI systems are not interns or general workers. They are the same doctors, lawyers, accountants, and engineers who work in your industry. If AI labs are paying $85 to $95 an hour to have these professionals teach AI to replicate their judgment, it tells you something specific about how far along the automation of knowledge work actually is. It is not theoretical. It is happening, and it is funded at scale.