Capability Gap

← All companies

Mercor

The fastest-growing broker of expert human labour to AI labs — $2B gross run rate on a 33% gross margin and 91% revenue from foundation-model companies.

high confidence5 minupdated 2026-08-29ai labs · data · expert labour
Vertical
Expert data for frontier labs
Founded
January 2023
Headquarters
San Francisco (New York office added 2026)
Raised
~$484M through Series C
Last valuation
$10B (Oct 2025); ~$20B reported in talks, July 2026 — no confirmed close
Revenue
$2.0B annualised GROSS run rate (June 2026); $614M gross in H1 2026. Net is roughly a third of that.
Status
Active; fastest-growing in the category; hit by an April 2026 data breach that cost it the Meta relationship

Mercor is the archetype of the expert data vertical and the only company in it whose real economics have leaked. Founded January 2023 by three Bellarmine College Prep friends — Brendan Foody, Adarsh Hiremath and Surya Midha, all Thiel Fellows who dropped out in March 2024 (Contrary Research) — it started as an AI-interviewer recruiting marketplace and pivoted into training data across 2024–25.

The growth is not in dispute. Gross run rate went $1M in early 2023 → $75M in February 2025 → ~$450–500M by autumn 2025 → $1B in February 2026 → $2B annualised in June 2026 (Contrary; TechCrunch). Foody frames it as $1M to $2B in 24 months (Fortune).

The gross/net gap

Say it plainly: that $2B is gross payment volume, not revenue. Contractors keep 60–70% (Dealroom).

Internal documents obtained by The Information in July 2026 give the sharpest picture anyone has of this business:

MetricFigure
H1 2026 gross revenue$614M, +70% YoY
Gross margin, 202527%
Gross margin, Q2 202633%
Labour as share of revenue~two-thirds
Revenue from AI foundation-model companies, H1 2026~91%
Internal projection, end-2026 run rate$2.8B
Internal projection, gross margin46% by 2027, 56% by 2030

Source: BigGo Finance relaying The Information; the 91% figure corroborated in AI Weekly's summary.

Three independent readings converge on the same take: an earlier Dealroom snapshot put net revenue at ~$20M against ~$83M gross, and Contrary describes Mercor charging clients roughly 35% above contractor cost (Contrary). So the honest restatement of the $10B Series C mark is not 13x gross — it is closer to 38x net. At $20B on a $2B gross run rate the headline is ~10x and the net figure is ~29x. Both are true; only one is a multiple on money Mercor keeps. See GMV is not revenue and What the public market pays for labour.

Profit: $6M in H1 2025 (Contrary). No later profit figure exists.

Concentration

~91% of H1 2026 revenue came from AI foundation-model companies, with OpenAI and Anthropic dominant. Foody has himself compared the concentration to Nvidia's. That is the single most important risk line on the page, and One customer is a binary event explains why: Appen was a $4.3B company with 80% of revenue in five clients before Google walked.

Mercor claims all top-five labs and six of the Magnificent Seven (TechCrunch). Individual contract sizes have never been disclosed. Roughly 2% of Q2 2026 revenue reportedly came from Chinese labs.

Not verified

The Chinese-revenue figure comes from Forbes reporting that could not be fetched directly; it is second-hand and unverified.

Supply

30,000+ active vetted contractors across 45+ countries as of October 2025, India the largest single source; 468K+ applicants evaluated by February 2025; at $50M ARR the company ran on 30 US FTEs plus 20 contractors in India (Contrary). Average pay crossed $100/hour per Foody in 2026 (Foody on X); TIME reported an $81/hour average with senior domain experts above $200/hour (TIME). Daily payouts exceeded $1.5M in October 2025.

Sourcing runs through a 20-minute structured AI video interview, with explicit targeting of ex-Goldman, JPMorgan, McKinsey, BCG, Latham & Watkins and Mount Sinai people. That is also the model's legal exposure: Mercor is buying proprietary workflows from people who still hold the day jobs. Foody's own comment is that at scale "there are things that happen."

Gap in the record

Fortune cites a network of "five million domain experts" against Contrary's 30,000 active contractors. The two figures are three orders of magnitude apart and no source states which is registered and which is active.

Trouble

  • Breach, March/April 2026. ~4TB exfiltrated — SSNs, DOBs, passports, biometric face and voice data, recorded interviews. Attributed to a supply-chain attack via LiteLLM; Lapsus$ claimed it. Six class actions; Meta paused work and opened an investigation (Staffing Industry Analysts).
  • Wage cuts, November 2025. Meta's "Musen" project was cancelled weeks after the $10B round; workers moved to "Nova" at $16/hr down from $21/hr. One affected Slack group had 5,000+ members. Mercor called the characterisation inaccurate (Forbes).
  • Misclassification. October 2025 class action alleging unpaid overtime, meal-break violations and unreimbursed monitoring software, seeking up to $25K per violation. Contrary's read is that reclassification would make the ~35% take untenable — see The law is about to arrive.
  • Cheating. Mercor runs a public Kaggle competition to identify interview cheating from behavioural signals and social-graph structure (Kaggle) — an unusually candid admission that Who is actually on the other end is a live operational cost.

M&A and direction

Sepal AI (February 2026) and Deeptune (July 2026, an a16z-backed RL-environment startup that had raised a $43M Series A three months earlier, in which Foody was personally an angel) (Orrick; Fortune). The thesis is full-stack: experts write tasks and rubrics, Deeptune supplies the environment they execute in. That is the right direction — environments carry a 4–5x exclusivity premium and datasets can be resold, while hours cannot.

Linked from
playbook

Building the supply side

Seed supply first, but only as much as the first contract consumes — and answer the utilisation question before anything else, because it is what kills these companies.

vertical

Expert data for frontier labs

Labs buy throughput of credentialled human labour — annotation, preference data, reasoning traces, RL environments — and pay nine figures for it at a staffing margin.

topic

What a rake can actually be

A take rate above ~20% survives only if you sell something other than discovery. Everything else — expert networks at 70%, freight brokerage at 8.5% — is a consequence of that one rule.

page

What the model actually is

Buyers with money but not capability, sellers who are fragmented, and a middleman who organises them and keeps a big cut. Four conditions have to hold at once — and calling the result a marketplace is how it gets mis-priced by a factor of ten.

vertical

Contingency recruiting marketplaces

Paraform has raised $65M and paid $50M to recruiters. Back the split out and the whole lifetime business is roughly $71M of placements and $21M of net revenue — against a perm-placement pool an order of magnitude smaller than the TAM slide says.

topic

GMV is not revenue

The same company is worth 13x or 38x depending on which number you divide by. Nearly every headline in this sector quotes the one that flatters — here is how to tell in ten seconds.

page

How to evaluate one of these

Six scores, one to five, higher always better for the operator — what each is trying to capture, where each misleads, and the three disqualifiers that override any total.

playbook

Pricing and the contract

The buyer's alternative is nearly always to hire someone, and increasingly an offshore someone at half the cost. Price against that — then decide gross or net deliberately, because the same facts that book you gross make you an employer.

company

Surge AI

Bootstrapped to over $1B of revenue with 130 employees and no outside capital — and the one company whose valuation nobody can pin down.

company

Scale AI

Took $14.3B from Meta for 49% of itself and lost most of its frontier-lab book within weeks — the cleanest natural experiment in why neutrality is the product.

page

The capital register

Every company the sweep found, with what it raised, what it was marked at, what it earns and whether that revenue is gross or net. Plus the exits, the failures and the absence of a public bear case.

topic

What the public market pays for labour

Below ~40% gross margin the revenue multiple is capped near 1.6x, permanently. Accenture proves the ceiling; Fiverr proves that clearing it is necessary and not sufficient.

page

Evidence register

Not a bibliography — a graded list of the claims the atlas leans on, what each one holds up, where it came from and how much weight it will take.

company

Handshake AI

A decade-old campus network that already owned the PhDs its competitors were recruiting — zero to $1B gross in fifteen months, on a valuation nobody has re-marked.

topic

How much money is actually in the buyer pool

There are two buyers, not one: about 10–20 labs signing six-to-nine-figure contracts, and several thousand startups buying at roughly $19K. Quoting the $510B headline as a TAM confuses them.

topic

Getting cut out

Leakage ≈ value per relationship ÷ (frequency × switching friction). Nobody has ever measured it — the best evidence in the sector is Upwork admitting, twenty years in, that it cannot.

company

micro1

5x'd in eight months to a $500M gross run rate, and is the only vendor claiming 80–90% gross margin on anything — by reselling the same dataset more than once.

vertical

Robotics teleoperation and physical-world data

Buyers pay $50–200/hour for demonstrations, operators get $25–50. The rig and the floor space are what make this defensible — and what the commodity end lost.

page

What we could not establish

The questions that would most change a decision, the thirty-three contradictions the research left unresolved, and the structural holes where no evidence exists anywhere — not just where we failed to find it.

playbook

When to stop

Five numbers and one calendar. Write them down before you need them, because every one of them will arrive attached to a reason it does not count this quarter.

company

Turing

A remote-developer staffing firm that repositioned as a lab data provider — profitable, $2.2B marked, and with no hard revenue number since 2024.

topic

Which side you build first

Single-player mode is worth roughly ten times the capital efficiency of subsidising both sides — and supply-side utilisation kills more of these businesses than demand ever does.

vertical

Expert networks

$200 to the expert, $800–900 to the client, held for four decades. The oldest version of this model has the most durable rake in the atlas — and may not sell to AI labs at all.

company

Invisible Technologies

An operations-as-a-service business that discloses a real profit figure — $134M revenue, $15M EBITDA — and is marked at 15x for it.

topic

One customer is a binary event

Above 15% of revenue a customer is a coin-flip, above 25% you are a division of that customer — and when your buyers compete with each other, neutrality is the product you are actually selling.

company

Appen

The only pure-play with audited numbers, and a 97% drawdown from peak — the base rate for what a concentrated data vendor is worth when one hyperscaler leaves.

topic

You pay weekly, they pay in sixty days

At a 25% gross margin on net-60 terms, roughly 11% of annual revenue is permanently trapped in the gap — and it has to be funded again every time you grow. The tempting fix is to fund it out of the crowd, which is how you acquire a docket.

topic

Marketplace, staffing firm, BPO or agency

Almost every company in this atlas is sold as a marketplace and operated as an agency. Two questions separate them, and the answers set the gross margin, the multiple and whether there is a moat at all.

company

Mechanize

$9.1M of seed capital, ~35 people, no disclosed revenue, and reported Google talks at $1.5B+ — the licence-and-hire template for selling environments instead of hours.

company

Prolific

The only company in the vertical that publishes its take rate — 42.8% on top of participant pay — and the best public anchor for what this spread actually is.

topic

What better models do to each layer

Three different things get called 'AI will eat this': models doing the crowd's work, models doing the middleman's work, and models creating the budget in the first place. They point in opposite directions and every vertical in the atlas sits in a different one.

company

Paraform

The archetype of the recruiter-side marketplace: pays independent recruiters ~70% of a placement fee and keeps the rest. $65M raised, $50M paid out, and no independent evidence of any of it.

topic

The law is about to arrive

The EU Platform Work Directive transposition deadline is 2 December 2026. It flips the burden of proof, catches you on where the worker sits rather than where you are incorporated, and closes the BPO workaround with joint-and-several liability.

topic

Who is actually on the other end

State-sponsored infiltration, real-time deepfake interviews, multi-accounting and LLM-assisted cheating are all live and documented. The best control against them — 1:N face matching against your enrolled gallery — becomes illegal for EU workers on 2 December 2026.

topic

Where the supply can legally live

A map of which supply geographies are cheap and safe and which are traps. The EU becomes expensive on 2 December 2026, one global IP clause silently fails in Germany and India, and your real competitor is an offshore FTE at half the cost.

company

GLG

The largest expert network, a 70–80% take held for decades, and a market share that halved anyway. Its 2021 S-1 is the highest-value document in the atlas — quoted in one research file and recorded as unopened in another.