Mercor's 30,000-Contractor Bench Is the Best Sourcing List in Tech

You are cold-InMailing Google engineers who reply at 3%, while 30,000 vetted PhDs, doctors, and senior engineers are already 1099'ing on nights and weekends for Mercor. They have passed a structured screen, they have ranked performance data attached, and they have told the market in writing that they want outside work. Almost nobody is sourcing them.

Mercor was reported in July 2026 to be raising ~$500M at a ~$20B valuation, roughly doubling its Series C mark from nine months prior. The company now manages 30,000+ weekly active contractors and pays out over $1.5M per day. That is not just a training-data story. It is the most credential-dense passive-candidate pool in tech.

Why the AI-training marketplaces are a sourcing list, not a threat

The training-data marketplaces have pre-vetted a bench of doctors, lawyers, PhDs, and senior engineers who are already contracting on nights and weekends. If you are hiring, Mercor, Handshake AI, Outlier, Surge, and Alignerr have done the top of your funnel for you.

Consider what these platforms actually do. AI labs pay hourly for vetted domain experts to perform post-training work: reasoning-trace generation, rubric creation, evaluation design. Mercor's screening includes automated 20-minute video interviews evaluated by proprietary language models, with performance data attached to every candidate that improves match predictions over time. That is more structured screening than most Series B startups perform on their own hires.

Roughly $10B flows annually into training-data providers. Handshake AI alone hit a $150M run rate by November 2025 and is approaching $1B in gross annualized AI training revenue as of April 2026, up from $550M previously. Micro1 went from $100M to $500M run rate in eight months. Turing reached $300M in annualized revenue by the end of 2024. AfterQuery crossed $100M in revenue at a $300M valuation. Every dollar of that spend produces a paid, ranked, actively-working professional who could be hired full-time.

$10B

The math that proves they are moonlighting

At 30,000 weekly active contractors and $1.5M in daily payouts, each Mercor contractor is grossing roughly $50 per day, or about $18,000 per year on the platform. That is side-gig income, not primary employment, and it is the single strongest buy signal a sourcer will ever get.

Cold LinkedIn InMails to random Google engineers reply at roughly 3%. Someone who has voluntarily signed up for a 20-minute AI-graded video interview, passed it, and is now billing rubric-design hours on Tuesday nights is a fundamentally different lead. They have already told the market: I have capacity, I will do paid work outside my day job, and I am comfortable with a 1099. A full-time offer that pays them for that same skill during business hours is the easiest conversation a recruiter can have.

The rate cards confirm the profile. Generalist AI trainers start at $22 to $30 per hour. Master's and PhD holders earn $30 to $150 per hour. Specialists in medicine, law, and finance can reach $175 to $300+ per hour. Handshake AI pays $30 to $160+ per hour across software engineering, finance, law, pharmaceutical research, and even music theory. Mercor experts earn $85+ per hour on average. These are not gig workers. They are professionals renting their credentials by the hour.

The bench, sized

Here is the moonlighting-expert bench in comparable numbers. The interesting row is the last one.

That 4,877 number is the one to sit with. In Refolk's index of professional profiles, 4,877 US professionals openly self-identify with titles like "AI Trainer," "Data Annotator," or "Prompt Engineer." The top current employers on those profiles: Handshake, DataAnnotation, Outlier, Alignerr. The exact marketplaces this article is about. Top hubs: NYC, Chicago, LA, Seattle, Bay Area.

That is not a hidden bench. That is a bench that has already listed itself. And 4,877 is only the surface layer, the people who put the label on their profile. The full moonlighting pool, using Mercor's 300,000-person network as a ceiling, is at least an order of magnitude larger.

Why traditional sourcers miss this pool

Traditional sourcers miss the training-data bench because it does not live in the places their tools look. LinkedIn Recruiter indexes job titles and companies. GitHub indexes code. The AI-training bench sits between them, in profile bios, personal sites, and Substack footers where an ER doctor mentions she "writes evaluations for Handshake AI on weekends."

A few concrete reasons the pool stays invisible:

• Domain experts are not on GitHub. Mercor's contractors are doctors, lawyers, PhDs, and specialists. Traditional tech sourcers cannot easily reach them because they do not live on the platforms tech sourcers search.
• The signal is a phrase, not a title. Nobody puts "Mercor" in their headline. They put "post-training evaluations" or "RLHF contractor" in a bio line or a project description.
• Boolean search collapses. "AI trainer" OR "prompt engineer" OR "data annotator" OR "RLHF" OR "reasoning traces" OR "rubric author," filtered by domain, filtered by location, filtered by seniority, is not a Boolean string. It is a paragraph.

This is the exact gap Refolk closes: you describe the person in plain English (an "MD who is moonlighting on medical-eval work for an AI lab in the Northeast") and get a ranked shortlist pulled from GitHub, LinkedIn, and the open web. No Boolean, no seat licenses per recruiter, no guessing which of a dozen self-labels the candidate happened to use.

Five non-obvious things about hiring from this bench

The moonlighting-experts pool behaves differently from a standard passive-candidate list. Here is what changes.

1. They are pre-screened harder than your own loop. Mercor's 20-minute AI-graded interview plus ranked performance data means a contractor at the top of their vertical has more objective signal than a candidate who just aced your take-home. Ask them for their Mercor tier.
2. Moonlighting income is the tell. ~$18K/year in side income says they have capacity and will read a recruiter message. A cold Google engineer will not.
3. You can target by domain, not by employer. The old sourcing move was "ex-Stripe engineer." The new move is "person who writes financial-reasoning rubrics on Handshake AI." Domain-first beats employer-first when you are hiring for AI product roles.
4. The Meta/Mercor breach opened a defection window. In March 2026, a supply-chain attack on LiteLLM exposed up to four terabytes of Mercor's internal data and contractor records, and Meta paused all work with the startup indefinitely. Contractors whose Meta projects vanished overnight are actively looking. That is a time-limited arbitrage.
5. Vendor churn creates quiet weeks. The training-data business has limited moats. Top labs switch vendors easily and some build in-house. Every time a project pulls, thousands of contractor hours evaporate. Full-time offers land best in those weeks.

A cold Google engineer replies at 3%. A moonlighting PhD who already passed a 20-minute AI interview reads every message.

How to actually run this play

Run this as a four-step sourcing motion, not a one-time list pull. The bench refreshes weekly.

1. Define the domain, not the title

Do not search for "AI trainer." Search for the underlying credential you actually want to hire: "radiologist writing eval sets," "securities lawyer doing rubric work," "distributed-systems engineer contributing reasoning traces." The marketplaces are just the delivery mechanism. The domain is the asset.

2. Pull the profile-text signal

Look for phrases people actually write in bios: "Handshake AI expert," "Mercor contractor," "Outlier reviewer," "post-training data," "RLHF contributor," "evaluation design," "reasoning traces." The 4,877 self-identified profiles in Refolk's index are the visible tip; the phrase-level searches surface the rest. Asking Refolk for "senior backend engineers who mention doing paid AI evaluation work on the side" returns the shape of candidate that a title-based search would miss entirely.

3. Message the moonlight, not the day job

The best opening line acknowledges what they are already doing: "Saw you've been doing eval work for Handshake AI on cardiology reasoning. I'm building the in-house version of that as a full-time role." That reads as informed and specific. It also implicitly promises the same intellectual work they are choosing to do on weekends, minus the 1099.

4. Time the outreach to vendor news

When a lab pauses a project (see: Meta and Mercor in March 2026), a Reddit thread, a Blind post, or a subreddit dedicated to the platform will surface it within 48 hours. Those are the weeks to hit the list hardest. Contractors whose income just dried up will convert.

Named marketplaces to map

Here are the platforms whose contractors show up in Refolk's US index and are worth building a persistent search for.

Handshake, DataAnnotation, Outlier, and Alignerr specifically show up as the top four current employers on self-identified "AI Trainer" profiles in Refolk's index. That is the shortest path to a live sourcing list: mirror those four employer names, filter by the domain you actually want, and message people who have already told you they will do the work.

What this means for the next 12 months

The moonlighting bench is not a fad tied to one lab's training budget. Even if OpenAI, Anthropic, Google DeepMind, Meta, or xAI rebalance vendors, roughly $10B a year keeps flowing through these platforms, which means the bench keeps growing and, more usefully, keeps churning. Churn is the sourcer's friend. Every project pause, every rate cut, every vendor switch produces a wave of experienced contractors who suddenly have Tuesday nights free again and a fresh reason to consider a full-time offer.

The recruiters who win the next hiring cycle for AI-adjacent roles will not be the ones sending more InMails. They will be the ones who treat Mercor's 30,000-contractor bench, Handshake AI's expert pool, and the 4,877 self-listed US profiles as what they actually are: a pre-screened, pre-warmed, pre-consented sourcing list that renews itself weekly.

FAQ

Is it legal or ethical to source contractors off these marketplaces?

Yes. You are not scraping the marketplaces or breaching any terms; you are searching public profiles where people voluntarily disclose their contract work. The candidates listed "Handshake AI" or "Outlier" in their LinkedIn or personal bios themselves. Reaching out for a full-time role is the same as reaching out to anyone else who lists a current employer.

Won't the marketplaces just pay them more to stay?

The rate cards top out at $175 to $300+/hr for the rarest specialists, but the median contractor is grossing roughly $18K/year on Mercor, which is clearly moonlighting income. Full-time roles compete on total comp, equity, benefits, and coherent problem ownership, not on hourly rate. The bench exists because these are people with day jobs; a better day job wins.

How do I find these people without paying for another sourcing tool?

You can do it manually: search LinkedIn and Google for phrases like "Handshake AI expert," "Mercor," "RLHF contractor," and "reasoning-trace author," then filter by domain and location. It is slow because the signal lives in bio text rather than title fields, so most sourcers use a natural-language tool like Refolk to describe the candidate in one sentence and get the ranked list back.

What about signal quality? Aren't a lot of these people junk annotators?

The generalist tier at $22 to $30/hr is noisy, yes. The domain-expert tier is not. Mercor's ranked performance data and Handshake AI's credentialed pool (drawn from 1,600 schools) mean the top decile of contractors has more objective screening attached than most in-house loops produce. Filter on domain and seniority; the noise falls away.