Handshake AI's $1.1B Runs on 500,000 PhDs. Poach the 3,279.

Handshake spent a decade as a campus job board and then, in about 18 months, turned into a $1.1B annualized business paying physics PhDs $100/hr to grade model outputs. That pivot, plus Scale's July 2025 cuts and xAI's September 2025 purge of generalist tutors, has cleanly separated the AI training market into two pools. One pool is being fired. The other is credentialed, remote-ready, and sitting behind a single Boolean query.

The market split you can actually source against

The AI training industry has bifurcated into generalist annotators (being cut) and credentialed domain experts (being onboarded at record pace), and only the second group is worth a recruiter's time. Handshake AI reached roughly $1.1B in annualized gross revenue by April 2026, growing about 349% year over year, by contracting PhDs, MDs, JDs, and quants to frontier labs. On the other side, Scale AI cut 200 FTEs plus 500 contractors on July 16, 2025 after Meta's $14.3B investment, and xAI dropped about 500 generalist "AI tutors" on September 12, 2025, roughly one third of its 1,500-person annotation team.

The two events happened inside eight weeks of each other, and they carried opposite signals:

• Generalists (text cleaning, image labeling, basic QA): oversupplied, being replaced by cheaper models grading themselves.
• Specialists (medicine, law, finance, advanced code, safety evals): 10x expansion promised by xAI in the same September announcement that killed the generalists.

Then xAI paused specialist hiring in 2026, and Meta suspended its Mercor partnership in March 2026 after a data breach. Each pause dumps thousands of pre-vetted PhDs back into the open market for 30 to 60 days before the next platform absorbs them. That window is your job.

What the numbers actually say

The specialist AI training market moves roughly $660M to $770M a year directly into the pockets of individual domain experts, which sets a hard floor on what full-time recruiters have to beat. Contractor payouts run 60 to 70% of gross across the category, per Dealroom, and Handshake alone advertises pay rates up to $80 to $100/hr on its Fellows program, with $100M paid out to date.

Do the math on the individual. A physics PhD moonlighting 15 hours a week at $100/hr clears $75K a year in side income. Any full-time offer under about $220K total comp will lose to "keep the day job and stay on Handshake." That is a defensible number you can quote to clients before you write the first InMail.

3,279

Why "AI Trainer" on a profile is a green flag, not a red one

The AI trainer title is the single highest-signal filter for finding domain experts who have opted into remote contract work, cleared an NDA, and passed a credential check. Most recruiters read the label as "gig worker" and scroll past. That is a misread and it is costing them the pool.

In Refolk's index of professional profiles, only about 3,279 US profiles currently carry "AI Trainer," "AI Tutor," "AI Evaluator," "Model Evaluator," "RLHF," or "Research Contractor" in the title string. Against the base rate of the US professional workforce, that is essentially zero. What it selects for is specific:

1. A domain credential that a frontier lab was willing to pay for (Handshake screens for PhDs; Mercor screens for verified expertise; Outlier screens on domain quizzes).
2. Comfort with remote 1099 work and NDA-heavy environments.
3. A public willingness to say "I do this," which is itself a self-selection into being reachable.
4. Recency: these titles did not exist as a category before 2023.

You are not looking at an unsorted pile. You are looking at a pre-filtered shortlist that the platforms paid to build. The four employers that dominate that shortlist, in Refolk's sample of top companies for these titles, are Outlier, Handshake, Alignerr, and Pareto AI. Those are your farm teams. Scrape them first.

The Boolean string that opens the pool

The right query mixes platform names with credentialing signals, and drops the generic buzzwords that flood the results with people who took one online course. A working starting point:

• Titles: "AI Trainer" OR "AI Tutor" OR "Model Evaluator" OR "RLHF" OR "Research Contractor" OR "MOVE Fellow"
• Employers: Outlier OR Handshake OR Alignerr OR "Pareto AI" OR Mercor OR Surge OR Turing OR AfterQuery OR Cleanlab
• Credential floor: PhD OR MD OR JD OR "PhD candidate"
• Domain narrowing: physics OR biology OR oncology OR "securities law" OR quant OR "medicinal chemistry"

The MOVE Fellowship deserves its own callout. Handshake's Model Validation Expert program is the formal on-ramp to its highest-paying evaluation work, and "MOVE Fellow" in a bio is roughly equivalent to "passed a domain-specific eval that Anthropic and OpenAI trust." It is a credential the market has not learned to search for yet. Add it. Handshake also acqui-hired Cleanlab (a nine-person team plus co-founders) for quality tooling, so "Cleanlab" is now a second-order signal for the same pool.

This is the exact gap Refolk closes for anyone who does not want to hand-tune Boolean strings across three platforms. You describe the person in plain English, and Refolk returns a ranked shortlist pulled from GitHub, LinkedIn, and the open web in one pass.

Time your outreach to the vendor news

The best moment to reach a domain expert is the week their platform freezes hiring, cuts a contract, or loses a customer, because that is when the "AI training as a career" story cracks. Four live events since summer 2025 have done exactly that:

• July 16, 2025: Scale AI cut 200 FTEs and 500 contractors after Meta's $14.3B investment triggered a customer revolt from rival labs.
• September 12, 2025: xAI cut about 500 generalist tutors and promised a 10x specialist expansion. The specialists onboarded, then watched the promise fail.
• March 2026: Meta suspended its Mercor partnership indefinitely after a data breach. Mercor had just hit $1B ARR from $500M in September 2025.
• 2026: xAI paused specialist AI tutor hiring, freezing the pool that had just been scaled up.

Each of those events created a 30 to 60 day window where the affected experts were actively receptive to cold outreach. The pattern will repeat. Vendor concentration in this market means single-point-of-failure sourcing events, and monitoring the trade press (TechCrunch, Sacra notes, Dealroom) is now part of the sourcing workflow, not a nice-to-have.

The ideology of AI training as a career has cracked in under twelve months, and that is the psychological precondition for cold outreach to work.

The four vendors where the names actually live

Focus your first sweep on Outlier, Handshake, Alignerr, and Pareto AI, in that order, because Refolk's index shows those four dominate the top-employers distribution for AI trainer titles. The rest of the market matters as second-pass targets: Mercor, Scale AI, Surge AI, Turing, and AfterQuery.

A quick read on each:

• Outlier (Scale AI's specialist arm): high volume, mixed credentialing, best for coding and quant.
• Handshake AI: 500,000 PhDs across 1,600+ universities, MOVE Fellowship as the quality tier, $100M paid out to Fellows to date.
• Alignerr: safety and alignment focus, smaller pool, higher rates per hour.
• Pareto AI: enterprise domain evals, common for finance and legal specialists.
• Mercor: three-year-old, $1B ARR by early 2026, currently shaky after the Meta suspension.
• Surge AI, Scale AI, Turing, AfterQuery: Surge and Scale both crossed $1B gross ARR in 2024; Turing hit $300M ARR by end of 2024; AfterQuery just crossed $100M at a $300M valuation.

Cross-reference by domain. A cardiologist doing evals almost always lives at Handshake or Pareto AI. A Rust systems programmer doing red-team work almost always lives at Outlier or Alignerr. A former SEC lawyer doing legal evals almost always lives at Pareto AI or Mercor. The mapping is tight enough that you can filter queries by the employer field and skip a whole layer of noise.

$75K

What to actually say in the first message

Lead with acknowledgment of the platform work, name the domain specifically, and quote a number the candidate already knows to be true. Generic outreach dies in this pool because these people have been pitched by every AI lab's internal sourcer for 18 months.

A working template shape:

1. Line 1: "I noticed you were a MOVE Fellow at Handshake working on [specific domain]."
2. Line 2: One sentence on why the domain matters to the hiring company, with a named product or team.
3. Line 3: A comp anchor that beats their platform income (calculate at 15 to 20 hrs/week times their public rate, add 40%).
4. Line 4: One question, not a link to a job description.

The comp anchor is where most outreach fails. Sourcers who quote a $180K base to a Handshake Fellow lose to inertia every time, because inertia pays $75K a year for evenings-and-weekends work the candidate can do in slippers. If you cannot get to $220K+ total comp, do not open the conversation with money.

For teams running this play at volume, Refolk will pull the shortlist and dedupe against your ATS in one query, so you spend the day writing the messages that actually convert instead of building the list.

FAQ

Is it legal to poach contractors from Handshake, Outlier, or Mercor for full-time roles?

Yes, in almost every case. AI training contracts are 1099 arrangements with confidentiality clauses about the labs' data, not non-competes against the platforms themselves. The specialist can leave a platform any day and take a full-time role at a lab, a startup, or a bank without breach. The one thing to check is whether the target has a specific client-side NDA with the end lab, which is worth asking about before sending a role at a competing lab. Most experts will tell you upfront.

How do I tell a serious "AI Trainer" from someone who took a weekend course?

Look for three signals in the profile: a named platform employer (Outlier, Handshake, Alignerr, Pareto AI, Mercor), a stated domain (oncology, securities law, Rust systems), and a credential that predates the AI role (PhD, MD, JD, senior IC title at a real employer). Titles like "AI Trainer" with no employer named and no domain named are almost always the weekend-course crowd. The MOVE Fellow tag on Handshake is currently the cleanest single filter for credentialed work.

What total comp do I need to offer to actually convert this pool?

Around $220K total comp is a practical floor for a PhD who is currently earning $75K a year on the side from Handshake or Outlier. Below that, the math favors keeping the day job and the platform income. Above $260K, conversion rates climb sharply because the platform income becomes a rounding error rather than a real alternative. Equity helps but does not substitute, because platform income is cash and cash-vs-equity is the wrong comparison for someone who already has a stable primary role.

When is the best time to run this play?

Immediately after a vendor event: a layoff, a hiring freeze, a customer loss, or a data breach at one of the top four platforms. The Meta-Mercor suspension in March 2026, the xAI specialist freeze in 2026, and the Scale-post-Meta customer revolt each opened 30 to 60 day windows where cold reply rates jumped. Set a Google Alert on "Handshake AI," "Mercor," "Scale AI," "Outlier," and "xAI tutor," and pre-build your shortlist so you can move the day the news hits.