Scale AI's flexible human data engine

Scale’s rebound shows that its core asset is not autonomous vehicle expertise, but a flexible human data engine that can be redirected to wherever AI spending is hottest. When self driving programs cut back in 2022, Scale lost high volume sensor labeling work from customers like Cruise and Argo. When LLM labs began paying for RLHF, output ranking, and safety review, the same workflow stack and contractor base could be repointed at text and reasoning tasks, helping ARR jump from about $215M in 2022 to $760M in 2023.1 2 3 4

• The AV business was attractive because it produced endless edge cases. Self driving teams sent in camera, video, and LiDAR data that humans had to mark up so models could tell a pedestrian from a pole. Scale sold this work on usage, which fit fast growing AV R&D budgets until those budgets cracked in 2022 and 2023.2 3
• LLM labeling is different work, but the same business shape. Instead of drawing boxes around objects, raters compare model answers, rank outputs, flag harmful responses, and sometimes review step by step reasoning. That let Scale reuse its labor marketplace and quality control software for a new buyer set, OpenAI, Anthropic, Cohere, and other model builders.1 4 5
• The rebound also changed the competitive set. In AV, Scale mainly competed with data labeling vendors like Appen, Labelbox, and TELUS. In LLMs, it also competes with newer labor networks like Mercor and Invisible that supply doctors, lawyers, and PhDs for higher skill RLHF tasks, pushing the market up from cheap crowd labor toward expert labor marketplaces.2 5 6

The next phase is a move from generic labeling toward harder to automate training and evaluation work. As base models absorb simpler annotation tasks, the durable part of the market shifts to expert judgment, regulated workflows, and full stack tooling around data, evaluation, and deployment. That favors companies like Scale that can bundle labor with software and sell into labs, enterprises, and government.2 5 6