Scale AI

Alex Wang founded Scale AI in 2016 with a prescient insight: the bottleneck for AI model improvement wasn't compute or algorithm design, but data quality. Starting with autonomous vehicle labeling—where annotated driving data directly improved model performance—Scale identified a fundamental truth that would define the company's trajectory for nearly a decade.

Scale began with a straightforward insight but faced a market that didn't yet understand the problem. The company started in the autonomous vehicle space, providing labeled training data that helped self-driving car models recognize objects and understand road scenes. As computer vision matured, Scale adapted, moving into Department of Defense work around 2020 and positioning itself as the infrastructure layer for any organization that needed to label data at scale.

The early customer base came from autonomous vehicle companies and computer vision teams that immediately recognized the value of high-quality labeled data. As the generative AI wave hit and frontier labs began building large language models, Scale found itself perfectly positioned. Companies like Meta, OpenAI, and others discovered they needed massive amounts of carefully curated training data—and specialized expert feedback to improve model outputs.

Scale's breakthrough came from understanding that AI models' needs evolve constantly. Eighteen months ago, the tasks were simple: comparing two short stories or choosing between options. By 2024, the work had transformed dramatically. Now, a single task might involve a top web developer building an entire website as an example for the model, or a PhD cancer researcher explaining nuanced medical concepts. Eighty percent of Scale's expert network holds bachelor's degrees or higher; about 15% hold PhDs, many earning significant sums for their specialized contributions.

The company operates two distinct business models: selling training data directly to model builders, and building end-to-end AI solutions for enterprises and government agencies. The enterprise side revealed a critical insight—generic data has diminishing returns. A healthcare system's internal processes differ from another's, so enterprises now must do their own labeling to capture institutional knowledge and judgment. This shift from low-cost generalist labor to high-expertise specialist work completely changed the unit economics and positioning of the industry.

Jason Droge, who joined as CEO in 2023 (13 months prior to this interview), emphasized that the narrative positioning from competitors claiming to have pivoted entirely to experts was misleading. Scale had been running expert programs since models needed them, continuously adapting alongside model capabilities.

In 2024, Meta invested over $14 billion to acquire 49% non-voting stock in Scale AI—a transaction that kept the company independent, maintained its board structure, and affected only about 15 of its 1,100 employees. Scale continues to grow every month since the deal closed, operating two separate business units that each generate hundreds of millions in annual revenue.

The company now tackles increasingly complex challenges: training AI agents to navigate enterprise software like Salesforce, helping healthcare systems reduce diagnostic backlogs by having AI pre-screen medical records, and building reinforcement learning environments where agents learn through interaction. Droge sees the next two to three years as a crucial inflection point where AI moves from "knowing things" to "doing things"—from knowledge tasks to agentic decision-making. However, he remains realistic about timelines: enterprises typically need six to 12 months of careful work—legal approval, regulatory compliance, change management, accuracy validation—before deploying AI on important processes, not the weeks-to-months timeframes often marketed.