Hugging Face Co-Founder on Open-Source, Microduck, and AI Infrastructure

The Business of Open Source

Thomas Wolf, co-founder and Chief Science Officer of Hugging Face, clarifies that while the company is famous for hosting open-source models, its primary revenue driver is specialized storage infrastructure. Unlike generic cloud storage, Hugging Face provides high-speed, AI-optimized storage that is pre-warmed and located in close proximity to GPUs. This architecture solves the specific bottlenecks of modern AI workloads, such as handling massive multi-terabyte model weights and versioning large datasets without redundant storage costs.

The Microduck Side Quest

Despite being a software-first company, Hugging Face acquired Pollen Robotics to explore the intersection of AI and physical hardware. The result, 'Microduck,' is a $399 open-source robot that sold 16,000 units following a single social media post. Wolf frames this as a 'side quest'—a project driven by passion rather than immediate profit—that serves to build a new community of AI builders interested in robotics. The robot is designed to be accessible, allowing users to train behaviors in simulation and transfer them to the physical unit.

AI Infrastructure and Compute

Wolf explains that Hugging Face intentionally stays one level below the 'token routing' business models seen in other platforms. Instead, they focus on providing the underlying compute and infrastructure for both training and inference. The company has seen a massive surge in demand for these services, particularly as the ecosystem for open-source models has matured significantly over the past year, with models like DeepSeek closing the performance gap with proprietary alternatives.

Security and Industry Responsibility

Addressing the recent security breach at Hugging Face, Wolf emphasizes the importance of basic security hygiene, such as credential management, in an era where AI agents are becoming increasingly autonomous. He draws a parallel between AI safety and aviation safety, suggesting that as models become more powerful, the industry must adopt rigorous, standardized safety protocols rather than relying on reactive measures.