Mercor's $20B Valuation Is a 35% Tax on 5,464 Public Profiles

Mercor is reportedly in talks to raise $500M at a $20B valuation, double its September 2025 mark, on the back of a contractor network Bloomberg says it pays more than $2M a day. If you run sourcing at a lab, a scale-up, or a data team buying from Mercor, that $2M is functionally your outbound recruiting budget. The uncomfortable part: most of the people on the other end of the invoice are searchable on LinkedIn.

What Mercor actually sells for 35%

Mercor sells routing latency, not access. The pool of ~300,000 professionals it markets to OpenAI, Anthropic, Meta and Google is directly-sourceable on the open web, and the "vetting" is a 20-minute AI video interview scored by proprietary language models that any competent team can reproduce with a $20/month LLM.

The pitch to buyers is that Mercor stands up a contractor with a domain skill (RLHF, legal review, medical annotation, code eval) in days instead of weeks. The pitch to contractors is that Mercor books the demand. What sits in the middle is a roughly 35% take rate, per Value Add VC's breakdown of the unit economics. Contractors keep 60 to 70% of gross billings per Bloomberg. On the company's disclosed ~$95/hour average rate, that is a $30 to $40 spread per contractor-hour. Annualized on a full-time equivalent, the spread is $62K to $83K per seat.

That number matters because it is the number an in-house sourcing function has to beat. It is not the cost of finding one senior engineer. It is the cost of renting one middle-of-the-market annotator for a year, and it recurs.

$2M

The 300K pool, decomposed

Mercor's headline 300,000-professional pool is a marketing number. The disclosed active roster is closer to 30,000 contractors per Big Think, and the slice doing frontier-lab AI training work is smaller still and openly visible under a handful of title conventions.

In Refolk's index of professional profiles, 25,054 people worldwide currently hold titles like "AI Trainer," "AI Tutor," "Data Annotator," "Prompt Engineer" or "RLHF." That is 83% the size of Mercor's disclosed active contractor base, and none of those profiles cost a 35% margin to contact. The US alone accounts for 5,464. Their top listed employers are the exact competitor set: DataAnnotation, Outlier (Scale AI's contributor arm), Handshake, and Alignerr.

The India line is the arbitrage story. India has more AI-trainer profiles than the US in the index (5,620 vs 5,464), clustered in Hyderabad, Mumbai, Bengaluru and Indore. Mercor's blended $95/hour is a US-anchored number. Buyers sourcing the India cohort directly can plausibly pay contractors $30 to $50/hour and skip the take rate at the same time. That is not a small optimization on a workforce burning $2M a day.

Why the moat is speed, not scarcity

Mercor's moat is a fast turnaround on a vetted contractor with a specific domain. The moat is not the identities of the contractors, which are public, and not the vetting method, which is reproducible with off-the-shelf models.

Consider what the 20-minute AI interview actually does:

1. Screens for English fluency and articulation.
2. Runs a domain probe (a law question, a medical question, a code task).
3. Scores the transcript with an LLM against a rubric.
4. Ranks the candidate against past hires for the same buyer.

Every step is reproducible. Step 1 is a baseline any recruiter already does. Steps 2 and 3 are a prompt and a rubric. Step 4 is a spreadsheet. The reason labs pay 35% for this is not that it is hard. It is that they do not want to staff the sourcing function to hit the underlying pool at their volume.

That is a build-versus-buy call, and it is the exact gap Refolk closes on the "find the people" half. You describe the person in plain English ("US-based RLHF contractor with a physics PhD, currently at Outlier or DataAnnotation, open to 20 hours a week") and get a ranked shortlist across GitHub, LinkedIn and the open web. The vetting layer you still own, but you are no longer paying $30/hour to have someone else run a Boolean query.

You are not paying Mercor for access. You are paying for the org chart you did not build.

The 30,000-person contingent workforce you cannot see

If your company is on Mercor's invoice list, you are already funding a 30,000-person contingent workforce with no candidate submittal, no interview loop, and no right of refusal. That is worse than a contingent search agency on every axis except speed.

A traditional contingent search fee is 20 to 30% of first-year salary, one time, on a person you interviewed and chose. Mercor's 35% is on every hour, forever, on people you never met. The individual contractors are perfectly good. The invisibility is the problem. If a specific annotator is producing 90th-percentile RLHF labels for your safety team, you do not know their name, cannot hire them full-time without going through Mercor's conversion terms, and cannot re-engage them if the platform pauses your account.

Which happens. In March 2026 a supply-chain attack on the open-source LiteLLM library exposed up to four terabytes of Mercor's internal data and contractor records. Meta paused all work with the startup indefinitely. Directly-sourced contractors sit in your ATS. Marketplace contractors sit in someone else's blast radius, and when that blast radius takes a hit, your project timeline takes it too.

Where the poachable bench actually lives

The poachable bench for AI training work is concentrated at four employers in the US and three metros in India, and all of it is public. The US bench sits at DataAnnotation, Outlier, Handshake and Alignerr. The India bench sits in Hyderabad, Mumbai and Bengaluru, with a secondary cluster in Indore.

Those are not obscure lists. They are the top current employers and top current cities on the 25,054-profile global set, pulled directly from title strings. If you want an RLHF contractor with a physics background, the population is finite and the population is indexed. The reason a team still ends up on Mercor is that "finite and indexed" does not mean "easy to reach at volume on a Tuesday morning."

That is where a plain-English search beats a Boolean string. In Refolk you can ask for "prompt engineers who left Scale AI or Outlier in the last 12 months, US-based, with published work on RLHF" and get a list without spending an afternoon in LinkedIn Recruiter building a nine-clause filter. For a lab or a data team already spending mid-six-figures a year through a Mercor-style marketplace, replacing even 20% of the volume with direct hires clears the cost of the sourcing tool many times over.

The insourcing argument the labs are already making

Frontier labs have started building their own data labeling and evaluation teams, which is the strongest available signal that the work is insourceable. If OpenAI, Anthropic and Meta believed the vetting layer was defensible, they would not be staffing internal equivalents.

Mercor's own risk narrative acknowledges this. Analysts note that if the top customers pull external procurement in-house, revenue faces a cliff-like decline overnight, because the top handful of labs are the revenue. Mercor claims to work with all of the top five AI labs and six of the Magnificent Seven. That concentration is the tell. When your five biggest buyers all think a function is worth building internally, the function is worth building internally. The people who staff it are already visible.

Two operational moves follow for a recruiting leader:

• Audit the invoice. Ask finance for the Mercor line item and divide by $95/hr. That is the FTE-equivalent contractor count you are renting. Anything above 20 FTE-equivalents is worth a direct-sourcing motion.
• Name the top 100. Pull the 100 highest-utilization contractors on the platform (Mercor will provide this in a QBR if you ask) and check how many are named individuals versus anonymized IDs. The anonymized ones are the ones you cannot hire away. Push for names on renewal.

The counter-argument is real. Marketplaces exist because vetting at volume is annoying, and Forbes reported that former Mercor employees suspected North Korean operatives had infiltrated the contractor network using stolen credentials. That is the sharpest argument against "just outsource the vetting" and it cuts both ways. It is also the sharpest argument against handing your candidate identities to a third party in the first place.

What a $20B valuation means for your 2026 sourcing plan

The valuation is a forecast that AI labs will keep paying a 35% tax on public profiles for another few years. Your 2026 sourcing plan should assume the opposite for at least part of the volume.

A defensible plan looks like this:

1. Keep the marketplace for surge and for domains you do not staff internally (niche medical, niche legal).
2. Direct-source the top three domains by hours consumed. For most buyers that is code eval, RLHF, and generic annotation.
3. Build the vetting rubric once, in-house, on the same LLM stack the marketplace uses.
4. Route direct-sourced contractors through your own ATS, not a third-party Slack.
5. Reprice the remaining marketplace spend against the direct-hire benchmark every quarter.

The number to beat is $62K to $83K per contractor per year. Every direct hire clears that spread.

FAQ

Is Mercor's 300,000-profile pool actually 300,000 unique contractors?

No. Roughly 30,000 is the disclosed active contractor count per Big Think, and 300,000 is the total registered talent pool disclosed in the July 2026 fundraise coverage. The distinction matters for buyers because the active number is closer to what you are paying against, and it is roughly six times the openly-sourceable US pool of 5,464 AI-trainer profiles in Refolk's index. Any sourcing plan should benchmark against the active number, not the marketing number.

What is a realistic alternative to Mercor for AI training and evaluation work?

The realistic alternative is direct sourcing plus an in-house vetting rubric, not a competing marketplace. The competing marketplaces (Scale AI's Outlier, Alignerr, DataAnnotation, CrowdGen, Handshake) draw from the same public talent pool and charge similar take rates. If you build the 20-minute LLM-scored interview in-house and pull candidates directly from the ~25,054-profile global pool, you keep the margin and you keep the identities. Vetted engineer marketplaces are a routing convenience, not a proprietary supply.

How big is the security risk of using a contractor marketplace?

Non-trivial and documented. The March 2026 LiteLLM supply-chain attack exposed up to four terabytes of Mercor's internal data and contractor records, Meta paused work indefinitely, and Forbes reported that former employees suspected North Korean operatives had infiltrated the contractor network using stolen credentials. That is a specific reason to keep sensitive evaluation and RLHF work on directly-hired contractors whose identities you know and whose devices you provision.

If we still use Mercor, what should we insource first?

Insource the domains with the highest hours-per-month and the lowest domain specificity. For most buyers that is code evaluation, generic RLHF, and standard data annotation, because the population is large, public, and reproducibly vettable. Keep the marketplace for surge capacity and for narrow specialties (rare languages, licensed professions) where the routing genuinely saves weeks. Reprice both categories quarterly against the $62K to $83K per-seat spread you are paying today.