Mercor's $750M Round Turns AI Benchmarks Into a Business

Mercor has closed a $750 million Series D at a $20 billion valuation, doubling the $10 billion mark from October 2025. Corporate money from Nvidia, Salesforce, and Zoom sits alongside Benchmark, Felicis, General Catalyst, Sapphire Ventures, StepStone, and funds tied to Apollo. No lead investor was named, which at this size reads as a deliberate choice: an alliance assembled, not a single fund writing a conviction check.

What Mercor Actually Sells

On paper Mercor is a marketplace: a network of more than 5 million domain experts whose output, annotations, gradings, domain reasoning, gets sold to frontier labs. Accurate, and increasingly incomplete. The product that raised this round is narrower and more technical, custom evaluation harnesses plus the post-training that follows.

For each enterprise customer, Mercor builds a benchmark that scores models against that company's own systems, policies, and edge cases rather than a public leaderboard. Co-CEO Adarsh Hiremath's argument: a model can top MMLU and still fail on the payroll exception, approval chain, or legacy ERP quirk that defines a real job. If that holds, the scarce input is not tokens or GPU hours. It is a reproducible representation of how a business operates. The benchmark becomes the spec, and the post-training run becomes the build.

Private evals also dodge contamination. Public benchmarks leak into training corpora within weeks, so high scores increasingly measure memorization. Private evals cannot leak, because the tests never appear online. They introduce a different failure mode, reward hacking, where a model satisfies the grader instead of the task. Mercor's answer is to sell eval design and post-training together. Coherent as a product, awkward as an incentive, and something enterprise buyers will eventually price.

The substrate came from Deeptune, the New York startup Mercor bought in July. Its team spent two years rebuilding hundreds of enterprise applications as sandboxes where agents practice and are scored. Mercor says the new capital funds evaluation infrastructure, which in practice means more of those environments. Agent eval quality tracks environment fidelity, and rebuilding a claims system is closer to capex than to a software cost curve.

The Numbers Behind the Multiple

The revenue curve explains the pricing. Mercor went from roughly $1 million to a $500 million run rate in about 17 months, crossed $2 billion gross annualized by June 2026, and is reported near $3 billion now. First-half 2026 gross revenue landed around $614 million.

At $20 billion against roughly $3 billion annualized, the multiple sits near 6.7x, rich for a labor marketplace and cheap against high-growth software. That tension is the underwriting.

The word gross deserves scrutiny. Mercor pays experts out of the revenue it books, so gross run rate overstates software margins. Dealroom's framing of the debate is the right one, GMV, ARR, or net, and which of the three a buyer should trust. On net revenue the multiple is materially higher and the comparison set shifts from software to staffing.

Foody confirmed a tender offer for employee shares without disclosing the split between secondary and primary capital. That matters twice: how much cash reaches the balance sheet for infrastructure, and how much of the round was retention rather than growth financing.

Three diligence questions decide whether $20 billion holds:

• Take rate durability. If labs push annotation work in-house or negotiate volume pricing, the spread between gross bookings and net revenue compresses, and the multiple resets against a services baseline.
• Benchmark lock-in. A custom eval is sticky only if it is refreshed as the customer's systems change. Recurring eval maintenance is the real recurring revenue; one-off dataset builds are projects.
• Sandbox coverage. Environment fidelity sets the quality ceiling on agent evaluation. Every recreated application is an asset that takes quarters to build and is hard to copy quickly.

The competitive frame is tight. Scale AI reported just under $1 billion in 2025 revenue, while Handshake's expert network, Surge, and in-house lab data teams chase the same procurement budget. Mercor's claim is that two years inside the labs produced something harder to copy than a dataset: a repeatable process for turning a job into a test.

Sapphire's Jai Das expects enterprise demand to inflect as open models get good enough to build on. If so, the buyer base widens from a handful of labs to every company deploying agents, and Mercor looks like a toll on verification. If not, it is an expensive staffing firm with excellent revenue and no software margin. Both readings survive contact with today's numbers, which is why the round closed without a lead.