RLHF Growth Threatens Scale's Future
Scale at $760M ARR
Scale’s best growth engine is also teaching customers how to need Scale less. The 2023 surge came from RLHF and post training work for frontier model labs, but that work gets cheaper every year as models get better at ranking outputs, spotting bad answers, and generating synthetic training data. That pushes human labor out of the broad, repetitive part of labeling and leaves only the hardest work, expert review, safety, and regulated edge cases, where scale of contractors matters less than quality and trust.
-
Scale started as a usage based labeling business for massive AV datasets, then pivoted as autonomous driving demand weakened. That made LLM post training a lifesaver near term, but it also moved Scale into a category where automation improves much faster than in visual labeling for cars or defense imagery.
-
The competitive center of gravity is shifting from cheap crowd labor to smaller pools of verified experts. Newer providers are built around PhDs, coders, doctors, and linguists, not giant anonymous workforces. That weakens the advantage of being the biggest human labeling marketplace and rewards tighter matching, credentialing, and workflow software.
-
The durable pockets look more like audits than bulk annotation. Customers still buy external human validation when they need traceability, second opinions, safety testing, cultural nuance, or work that cannot be handed to a rival tied too closely to one model lab. That is closer to premium services than commodity labeling.
The market is heading toward a split. Commodity RLHF will be absorbed by models and self serve tooling, while high value human data will concentrate in expert evals, compliance, and mission critical workflows. Scale’s long term outcome depends on moving up that stack fast enough that it is selling orchestration, security, and auditability, not hours of labor.