Connaissez-vous Scale AI, l’opérateur discret de la chaîne d’entraînement des modèles d’IA ?

Scale AI: Enhancing AI Model Training with Data Structuring and Validation

Founded in 2016 by Alexandr Wang , Scale AI specializes in data structuring, annotation, and validation for machine learning model training. Unlike companies that focus on model development or hardware infrastructure, Scale AI deals with the critical but often underestimated aspect of data quality and organization.

Scale AI offers services across the entire data processing cycle, from manual annotation to dataset optimization for specific applications. Its clientele includes both civilian and public sectors engaged in large-scale AI system development, such as Microsoft , OpenAI , and the U.S. Department of Defense , particularly in the Task Force Lima program focused on securing military AI applications.

The company customizes datasets for specific contexts, including military simulation, legal or medical model training, and embedded system validation.

Scale AI reportedly generated $870 million in revenue in 2024 , with projections reaching $2 billion for 2025 . This growth reflects increasing demand for annotated data to support the rise of generative AI models and autonomous agents.

In terms of valuation, the company was valued at $14 billion in its last funding round in 2024. Early 2025 discussions suggest a potential valuation of $25 billion in a secondary market operation, although no official comments have been made by involved parties.

Data labeling is a crucial step in AI system development. While open-source models and diverse computing resources are expanding, creating specialized, reliable, and representative data corpora remains a rare skill.

In this environment, Scale AI holds a strategic intermediate position , not as a model developer or infrastructure provider, but as a vital link in the learning chain. This role enhances its visibility among industries and public institutions involved in AI technology development or deployment.

D’après FrenchWeb.