Handshake's ARR crosses $1B as AI training revenue surges

Handshake, a startup founded in 2013 as a job site for college graduates, has become a major player in AI training data. Its gross annualised revenue from AI training has surged to nearly $1B — up from $550M in January 2026 and just $5M–$10M shortly after it launched the business.

Combined with its college recruiting business, Handshake's total gross ARR now stands at roughly $1.15B. After contractor payments, net revenue from AI training approaches $300M.

Handshake joins a small club. Surge AI and Scale AI both crossed $1B in gross annualised revenue in 2024. Mercor, a three-year-old rival, hit the same mark in early 2026, up from $500M in September 2025, and is profitable on a free cash flow basis.

These companies contract lawyers, PhD holders, and medical doctors to stress-test AI models and answer niche queries — from doctorate-level physics to specialised medical topics. AI labs including AnthropicDealroom has a profile for this one. Try Dealroom → and OpenAI are their primary customers.

Demand for domain-specific training data is accelerating as AI labs push models into specialised fields. Generic datasets no longer suffice; labs need verified expert knowledge to improve model performance on complex tasks.

Handshake's pivot is striking in speed. It went from near-zero to nearly $1B in gross annualised revenue in AI training in roughly a year — reflecting how urgently AI labs are spending.

Mercor's trajectory offers a cautionary note. A data breach in late March 2026Dealroom has a profile for this one. Try Dealroom → prompted Meta to indefinitely pause its work with the startup and launch an internal investigation. Whether Mercor has sustained its revenue pace since remains unclear.

Concentration risk looms for both companies. Contractor costs run at 60%–70% of gross revenue, so losing even one major client hits hard. Handshake's recruiting business adds roughly $150M in gross annualised revenue as a buffer, but its AI training unit now dominates its financials.

Data labelling — once AI's unglamorous back-office — is now a billion-dollar market, and it is growing fast. Four startups have reached or approached $1B in gross annualised revenue, with demand still expanding.

The shift toward expert-in-the-loop training data reflects a broader trend: as AI models mature, specialised quality matters more than raw scale. The humans grading AI answers may prove as important to the next generation of models as the engineers building them.

Source:
The Information

Image credit:
Handshake

J.V.