From Annotation To Expert Infrastructure

This market shift means the scarce input is no longer cheap labor, it is proprietary judgment wrapped in workflows that look like real work. Generic labeling vendors mainly sell large pools of annotators. The newer post training stack sells specialists who can solve hard tasks, simulated environments where models practice multi step actions, and eval systems that score whether a model actually performs like a lawyer, doctor, coder, or analyst in production.

• Big incumbents have moved up the stack. Scale now sells RL environments and reports that nearly half of new data training projects involve reinforcement learning environments. Labelbox has added RL data, private AGI benchmarks, arena evals, and an expert network through Alignerr, which broadens competition far beyond annotation marketplaces.
  1 sacra 3 sacra 4 sacra
• Talent marketplaces have also invaded the category. Turing packages frontier data for coding, STEM, multimodal, and domain specific reasoning, alongside RL gyms and human in the loop synthetic pipelines. Handshake bought Cleanlab in January 2026 to expand from expert supply into evaluations and RL environments, following the same path as Mercor and Surge.
  1 sacra 4 sacra 5 sacra
• The revenue pool has followed the shift. Scale reached an estimated $1.5B revenue in 2024, Turing reached an estimated $300M, Handshake reached an estimated $1.1B gross annualized revenue in April 2026, and AfterQuery was at an estimated $100M annualized revenue in April 2026. Buyers are concentrating spend with vendors that combine labor, tooling, and benchmarks in one contract.
  2 sacra 3 sacra 5 sacra

Going forward, the winning vendors will look less like staffing platforms and more like applied AI infrastructure companies. The durable position comes from owning three things at once, expert supply, task environments, and trusted evals. That is where budgets, switching costs, and product differentiation are compounding across the post training market.