Labeling Shifts to Trusted Infrastructure

These two shocks show that human data is now bought like strategic infrastructure, not temp labor. When one of the few vendors trusted by frontier labs becomes conflicted or insecure, budgets can move fast because the work is mostly interchangeable and buyers already multi source across providers. Meta taking a stake in Scale raised neutrality concerns, then Mercor’s March 2026 breach raised security concerns, pushing work toward vendors like Handshake, Surge, and Prolific.1 2 3 4 5 6

• The first reset was about independence. After Meta invested in Scale in June 2025 and brought in Alexandr Wang, OpenAI said it was winding down Scale, and reporting showed Google also pulled back. In this market, labs do not want a rival with visibility into post training workflows, evaluator pools, or model roadmaps.1 2 3 4
• The second reset was about trust and operational security. Mercor disclosed that a late March 2026 supply chain attack hit data tied to projects, workers, and internal systems, and reporting said Meta paused work while it investigated. For buyers handling unreleased models and sensitive evals, security failure can matter as much as price or speed.5 6
• That churn favored vendors with ready pools of credentialed talent and fast fulfillment. Handshake had already repurposed its network of students, PhDs, graduates, and postdocs into an expert labeling supply base. Across the category, labs still use outside vendors even with internal annotator pools, because they need overflow capacity, second opinions, and niche cohorts on short notice.7 8 9 10

The next phase is a shift from simple labor marketplaces toward trusted data infrastructure. Winning vendors will pair expert supply with security, workflow integrations, eval tooling, and specialized pools that are harder to swap out. As labs spend more on safety, red teaming, and domain specific reasoning, vendor concentration will stay low and budget reallocation will remain a major growth driver.8 9 10