Scale pivoted from AV to LLMs
Scale: the $290M/year Mechanical Turk of machine learning
This decline shows how exposed early data labeling leaders were to one very specific buyer, the self-driving car R&D team. Scale’s original growth came from AV programs that generated huge streams of camera and lidar data, but those budgets were tied to long horizon research spending, not production software revenue. When Ford and Volkswagen shut down Argo AI, Uber sold ATG to Aurora, and Lyft sold Level 5 to Toyota, the raw volume of new labeling work fell with them.
-
AV was unusually label intensive because every mile produced new edge cases, rain, glare, dusk, odd pedestrians, construction zones, that teams had to tag by hand so models could learn from them. That made autonomous driving far more data hungry than narrower document or image classification use cases.
-
Scale was built to monetize that firehose with per task pricing and bundled labor plus software. That model worked well while Cruise, Argo, Lyft, Uber, and others were still funding open ended autonomy programs. Once those programs slowed or disappeared, usage revenue dropped faster than a seat based software contract would.
-
The practical effect was to push Scale toward newer demand from LLM training and RLHF. By 2023, the business had shifted from labeling sensor data for cars to managing human feedback loops for model builders, which is why AV weakness and overall company growth could happen at the same time.
Going forward, the center of gravity in labeling keeps moving away from broad AV annotation and toward higher value human judgment work, model evaluation, and specialist data collection. The winners will be the platforms that can redirect labor and workflow software from declining bulk labeling markets into newer AI training categories without losing speed or margins.