Meta stake triggers Scale client exodus

The Meta deal turned Scale from a neutral arms dealer into infrastructure partly controlled by one of the labs it serves. That is especially toxic in AI training, where customers hand over unreleased model outputs, safety data, eval rubrics, and roadmaps. Scale had ridden frontier lab demand to an estimated $1.5B ARR by the end of 2024, but once Meta bought 49% in June 2025, major labs had a clear reason to move sensitive work elsewhere.

• This is not just normal vendor churn. Google had planned to spend about $200M on Scale in 2025 before pulling back, and OpenAI began winding down its work shortly after the Meta transaction. In this market, trust and data separation matter as much as price or labeling quality.
  1 cnbc 4 washingtonpost 5 apnews
• The work itself is also changing in a way that helps rivals. Early Scale won by bundling software with huge generalist labor pools for image and text labeling. Frontier labs now need smaller groups of verified experts for reasoning, safety, red teaming, and domain specific evals, which favors players like Invisible, Mercor, Handshake, Prolific, and expert networks.
  2 sacra 6 sacra 7 sacra 8 sacra
• Scale still has real advantages, including a broad product suite, government contracts, synthetic data tools, and a long history managing massive annotation workflows. But the Meta tie up makes its best frontier lab accounts structurally harder to keep, because customers do not want a direct rival sitting beside the data exhaust from their model training loops.
  2 sacra 3 sacra 5 apnews 8 sacra

The market is heading toward a split. Neutral providers will take the most sensitive frontier lab work, while Scale leans harder into Meta aligned workloads, defense, enterprise, and productized data infrastructure. Over time, human data vendors that combine trusted separation with specialist talent and fast workflow software will capture the highest value layer of post training and model evaluation.