The most expensive input in artificial intelligence is no longer a processor. It is a person who can tell a model, in detail, why its answer is wrong.
On Nov. 10, Mercor said it had raised $750 million in a Series D that values the company at $20 billion, double the $10 billion it carried in October of last year. The round included a sale of shares by employees, a structure that lets early staff take money off the table without waiting for an exit. Nvidia, Salesforce and Zoom were among the investors.
Mercor’s business is the unglamorous half of model development. It recruits people with domain training and organizes their work into annotations, comparisons and graded responses that labs use to improve a model after its initial training and to measure whether it can be trusted on a specific task. The output is not a dataset in the ordinary sense. It is a record of human judgment, produced at industrial scale.
That business has become more valuable as the easy gains from raw text have narrowed. Model developers have scraped most of what is publicly available and are now competing for material that is scarce, licensed or created on demand. Synthetic data, generated by models to train other models, helps at the margin but tends to reproduce the weaknesses of the model that produced it. Human expertise is the corrective.
The second shift is in who buys. For most of the past three years, the customers were the labs themselves. Now the buyers include large companies deploying agents — software that completes a task rather than answering a question — inside their own operations. Those deployments need something a benchmark cannot provide: a definition of what a correct answer looks like for this company’s invoices, this company’s contracts, this company’s customer records.
That is closer to consulting than to labeling, and it is priced accordingly. It also changes the sales motion. A lab buys capacity by the task. An enterprise buys an outcome, which means the vendor has to understand the workflow before it can build the test.
The mechanics of that sale are unglamorous. A deployment team writes down the tasks an agent will handle, runs the model against them, marks the failures and pushes the fixes back to the vendor. Multiply that by a dozen workflows and the internal cost of quality control starts to approach the cost of the software itself, which is why many companies would rather rent the capability than staff it.
Supplying it is harder than it sounds. The people whose judgment is worth paying for tend to have other jobs, and the ones who sign up want work that fits around existing commitments. Recruiting them at volume is a logistics business as much as a technology one, and it is the part of the model that cannot be copied from a paper.
Mercor’s valuation implies investors expect the category to consolidate around a few vendors with enough expert supply to serve the largest buyers. The best-known competitor, Scale AI, built its business on a similar premise and has spent years expanding from autonomous-vehicle data into models and government work. Smaller firms compete on price and on the speed with which they can recruit.
The risk sits on both sides of the transaction. Expert labor is a market, and markets tighten. If the going rate for a qualified annotator rises faster than what customers will pay per task, the margin compresses. If customers decide the work is strategic enough to bring in-house, the vendor loses the account it spent a year training.
There is also the question of what happens when models get good enough to grade themselves. Some of that already happens: one model reviews another’s output, and a person reviews the disagreements. The human share of the work shrinks in volume but rises in difficulty, which favors firms that can find genuinely rare expertise rather than firms that can find a lot of people.
Mercor has been moving in that direction, from general annotation toward specialists whose credentials let them judge work in technical and professional fields. The pitch to an enterprise buyer is that the same pipeline used to test a frontier model can be pointed at a claims-processing agent or a coding assistant, with the same discipline about what counts as correct.
Whether that is a durable business or a service with a spreadsheet depends on retention. No revenue figures were disclosed, so the $20 billion is a bet on growth rates that outside investors cannot verify. Doubling a valuation in twelve months is what happens when a category is being repriced, not when a company is being measured.
For Nvidia, Salesforce and Zoom, the investment is also a hedge. Each of them sells into the same enterprise accounts that are now buying agents, and each has an interest in the supply of evaluation work that determines whether those agents get deployed or shelved. A stake in the vendor is cheaper than building the capability.
The comparison that matters is not to other startups. It is to the cost of a bad deployment: an agent that answers confidently and wrongly, in front of a customer, at scale. Companies will pay to avoid that, at least once.
Mercor’s next test is whether they pay twice.
