Mercor's $2B Run Rate Is Gross: Reading Vendor Revenue
The Information's funding materials show Mercor took in $614 million in H1, paid about two-thirds to contractors, and ran gross margins of 27% last year and 33% in Q2. The valuation under discussion: $20 billion, about 10 times gross annualized. Vendor annualized figures cannot be compared directly, nor their multiples with software companies. Align the basis first.
01Two bases, four vendors
| Company | Disclosed figure | Basis | Source |
|---|---|---|---|
| Mercor | Over $2 billion (June this year) | Gross annualized, includes contractor payments | The Information |
| Handshake | Near $1 billion (first four months this year, up 82% over the same period) | Not stated; peers pay 60-70% to experts (TechCrunch) | The Information, TechCrunch |
| Micro1 | $500 million, up from $100 million eight months earlier | Gross annualized; net $150-200 million | TechCrunch |
| Snorkel AI | $375 million, up 18x in twelve months | Expert fees in cost of goods sold | TechCrunch |
The last row is key. Snorkel sells RL environments and finished datasets, not labor hours; expert fees go to the cost line. The other three sell expert hours plus matching, so expert fees sit inside revenue. $375 million and $2 billion on one board compares two accounting methods.
Checked detail: the Micro1 report says the company "keeps 60% to 70%", but its stated net of $150-200 million is 30% to 40% of gross, matching TechCrunch's expert-share figures — a typo. The number is billing flow, not revenue.
02The margin sets the multiple
About one-third stays. My estimate: roughly $670 million net annualized, so the $20 billion valuation is near 30x net — arithmetic on disclosed ratios, magnitude only; Mercor never disclosed net. The margins corroborate — 27% and 33% is roughly "one-third stays." Gross margin here is largely the net share after the expert cut.
Per the report: Anthropic expects 77% gross margin in 2029, OpenAI near 70% by end of 2029, GitLab 86% most recent quarter. Mercor's internal forecast: 46% in 2027, 56% in 2030 — fully delivered, still below software companies today.
Mercor turned its first profit last year with $9 million of operating profit, a 3% margin; it expects $226 million, 12%, this year; 26% in 2027; 45% in 2030 — forecasts shown to investors, not confirmed results. The 3%-to-12% jump comes from diluting costs beyond the expert share, and requires revenue to keep doubling and clients not splitting orders; every line depends on the retained third.
Benchmarked against outsourcing, the multiple should use outsourcing standards. But Mercor's investor story differs: by 2030, foundation model clients fall below one-third of revenue and Fortune 2000 companies take 57%, which holds only if the margin structure changes.
03How the margin mix changes
In July, Mercor acquired Deeptune, an RL environment builder. Researchers in the report argue that in many professional domains, simulations built with Excel or Salesforce train agents better than human annotation. An environment trains repeatedly, so cost stops scaling with hours; Snorkel books environment revenue on a net basis too.
Reselling: Micro1's off-the-shelf data reportedly carries gross margins of 80% to 90% — custom tasks pay expert fees every time; off-the-shelf pays once, sells many times.
Quality control: in January, Handshake acquired Cleanlab; TechCrunch reported the deal was mainly about absorbing nine core staff. Cleanlab flags bad data without a second human reviewer; one fewer review round trims the expert share — revenue unchanged, margin moves. Its strategy lead told TechCrunch the team has long studied how high its own data quality really is. Internalized, this becomes a second basis in price talks: quality guaranteed by its own models.
Reselling runs into exclusivity: critics told TechCrunch that selling off-the-shelf data to Chinese model developers helps them catch up with top US models, and Micro1 founder Ali Ansari said publicly he does not sell to Chinese model developers. A Chinese data industry figure told 智能涌现 that large domestic buyers now set exclusive dataset windows or buy out supplier core staff (anonymous). These margins are paid for with exclusivity and customer scope.
Environment delivery's cost is acceptance: the scorer must resist model gaming; part of the margin saved returns to verification. And Mercor does not claim humans disappear — human review stays irreplaceable in high-risk, complex scenarios such as legal cases.
04How buyers use the basis
My judgments, not disclosed terms — vendor pricing has no public data.
One: ask for net. Every dollar covers expert hours plus vendor retention — retention in the 30-40% range is a variable to probe. Break quotes into expert hourly rate, spot-check ratio, service fee.
Two: ask exclusivity for off-the-shelf data. One dataset sold to several clients pushes vendor margins above 80%; pay an exclusivity premium or accept other buyers hold the same data — ask for the exclusive window and sales scope.
Three: pay for environments on acceptance. Models exploit scorer loopholes, documented in training; independent review and lookback windows beat squeezing the unit price.
Four: prevent single-vendor dependence. Mercor took 91% of first-half revenue from foundation model companies. In March a supply chain attack hit; Meta paused cooperation, and on June 25 the company said all frontier model clients had raised cooperation over recent months — the biggest valuation variable.
$2 billion is gross annualized, about two-thirds to experts. Read vendor revenue in order: basis, what is retained, hours or assets.