Clem Delangue on Open AI, Model Choices and Building Hugging Face

Hugging Face co-founder and CEO Clem Delangue tells Harry Stebbings that the company chose its name from a favorite emoji, initially as a temporary idea. The community embraced it. He and his co-founders wanted to work together and were excited by AI before its current mainstream attention.

The first product was an entertainment-oriented AI companion, which the company pursued for almost three years. Delangue says users exchanged billions of messages, but the underlying tools attracted interest when the team shared them. That response led to a pivot toward an AI platform and open-source community; the usage figure is his account.

Delangue views the 2023 surge in public and investor interest partly as a catch-up with AI already used in search, social ranking and video calls. He credits shared scientific research and open-source work for much of the progress, alongside better hardware and model optimization. His counterfactual claims about how many decades openness saved are opinion, not measurable proof.

He acknowledges Silicon Valley’s intense AI activity, recalling a large community gathering in San Francisco. But he says researchers and engineers are distributed globally and founders need not relocate permanently. He describes Hugging Face’s teams across Paris, New York, San Francisco and elsewhere and recommends choosing a place where the founder can work happily.

Stebbings contrasts a single powerful model with many open models. Delangue argues different use cases have different requirements for cost, speed and control, and expects many organizations to build or adapt their own models. He is the CEO of an open AI platform and has a direct commercial interest in this thesis.

Using a model through an API may be the fastest way to begin, he concedes. His concern is that relying on a single provider can limit customization, internal expertise and differentiation over time. He compares a hosted website builder with writing a tailored application; the analogy is illustrative, not proof that every company must train its own model.

Stebbings suggests enterprises may prefer a bundled, easy-to-buy service for years. Delangue accepts that possibility but predicts AI-native startups that train or optimize models could create differentiated products. His assessment of named startups and incumbents is a forecast and reflects Hugging Face’s business model.

Asked about rights in training data, Delangue says uncertainty affects both open and proprietary models. He refers to a reported threat by Elon Musk to sue OpenAI over use of tweets and expects greater legal clarity. The discussion does not establish that a lawsuit was filed, what data were used or the legal outcome.

He says the future relationship between content providers and AI companies is uncertain. He hopes for arrangements that sustain creators and enable AI development, and describes an opt-out mechanism in a code-model data project. The interview does not resolve consent, copyright or fair-use questions.

Pressed on Musk and OpenAI, Delangue says he cannot know what happens behind the scenes, but appreciates attention to openness and transparency. This is a view of a public dispute at the time, not a finding about either party’s legal obligations.

He describes Hugging Face’s freemium model: basic platform use is free, while some companies pay for enterprise features, support and compute. He cites roughly 15,000 company users and 3,000 paying companies in 2023, including named customers. These are the CEO’s figures, not an independently audited count.

Pricing varied and Delangue says the company had prioritized adoption and usage over maximizing immediate revenue. He expects monetization to evolve as the technology changes. His belief that platform usage will eventually translate into substantial revenue is an assumption, not a demonstrated future result.

Stebbings argues incumbents may capture most value because they have distribution and are adopting AI quickly. Delangue distinguishes API-based applications from startups that build and optimize models, arguing the latter may have opportunities incumbents find difficult to pursue. Both are making competitive predictions, not establishing who will win.

He identifies recruiting people who combine scientific and engineering skills as a major challenge for AI-first startups. Compute and specialized talent can make them more expensive to build than conventional software companies. He is less certain that every company needs the very large funding rounds then appearing in the market.

Delangue questions whether simply spending more on ever-larger training runs will always produce sufficient returns. He welcomes experimentation and investment while acknowledging that not every bet will succeed. These are 2023 assessments of costs and returns, not a universal scaling law.

On regulation, Delangue says AI can amplify bias, misrepresentation and misinformation and needs appropriate safeguards. He disagrees with calls for an immediate broad pause based on existential fears, arguing it could slow benefits without addressing present problems. The disagreement is about risk priorities and policy, not proof that long-term risks are absent.

He objects to treating current models as human-like or assuming autonomous takeover is imminent. He says long-term research should continue but should not crowd out attention to current harms. His view that such scenarios are distant is a contested prediction, not an established timeline or consensus.

Delangue says he generally avoids talking to outside investors between rounds so he can focus on building. Once he has enough term sheets, he stops new conversations. He recalls an investor sending a term sheet before they had met; the anecdote describes his own fundraising process.

Stebbings pushes back that a short fundraising window may not build the trust needed for a long relationship. Delangue responds that companies pivot, investor interests change, and he spends several intensive days and checks references before choosing. They disagree about whether that is enough; neither approach guarantees a good partnership.

Delangue sees investors’ primary role as supporting financing and capitalization, rather than acting as operators for many companies at once. Stebbings agrees that founders should not build solely to satisfy investors. This is a founder’s view of the relationship and does not preclude useful operational help.

He says early backers included French angels and that later venture investors were primarily US-based, so he does not claim broad experience with European venture firms. His general lesson is that company building does not necessarily become easier as the business grows; the nature of the challenges changes.

He urges founders to enjoy the process rather than postpone satisfaction until the next fundraising milestone or IPO. On fundraising, he says difficulty depends more on traction, momentum and negotiating position than on the label of the round. This is his experience, not a guarantee of a quick transaction.

In quick questions, Delangue predicts that companies will have their own AI models and identifies an AI market that fails to deliver as a major risk for Hugging Face. He says an open community approach is intended to help the ecosystem succeed. His prediction and commercial incentives are inseparable from his role.

He names Richard Socher as an influential angel investor, citing experience in research, Salesforce and entrepreneurship. That is a disclosed investment relationship. Delangue says a painful but useful lesson was accepting that each stage brings new difficulties and choosing to build a company he enjoys.

He considers people who can create and train state-of-the-art model architectures especially difficult to hire, offering a rough estimate of the talent pool. The interview provides no independent basis for that estimate. He hopes more people will develop these skills.

Looking ten years ahead, Delangue hopes Hugging Face becomes an impactful AI organization; he treats company size as a possible consequence rather than his stated primary goal. This is an aspiration, not an achieved outcome. Stebbings thanks him and the interview ends.