Speaker
Clément Delangue
Appearances over time
1 episodes
Episodes
1
Podcasts
Quotes & moments
Hugging Face crossed $100 million in annual recurring revenue, validating that an open-source AI platform can build a sustainable business without prioritizing monetization.
A Stanford study found that 70% of queries sent to ChatGPT could be accurately answered by smaller, cheaper models running locally on a laptop.
Clément Delangue argued that the most dangerous technologies — like the nuclear bomb — were never developed in open source but in closed, well-funded proprietary environments.
Clément Delangue contends open source models are inherently safer because they are less generalist, less focused on dangerous capabilities like cybersecurity, and benefit from sunlight as a disinfectant.
Clément Delangue said distillation is a universal practice across AI labs, and stopping it wouldn't cause Chinese labs to collapse — it's an accelerant, not the core reason for success.
Restricting open weights is fundamentally different from restricting an API — removing a model from Hugging Face just pushes it to ModelScope or torrent platforms.
Clément Delangue warned that a world where a handful of companies concentrate all AI power, capabilities, and wealth is more dangerous than those companies losing billions in revenue to competition.
Hugging Face's Llama CPP library is cited as the most widely used runtime for running AI workloads locally on consumer hardware.
Clément Delangue believes Europe has the talent, energy infrastructure (especially French nuclear), and existing labs like Mistral to build a world-class frontier AI lab if it focuses its resources.
Clément Delangue observed that young people have rapidly moved past the user phase of AI and are now building models, training datasets, and products themselves across domains like climate and biology.
As routing directs workloads to specialized cheaper models, Clément Delangue predicts a significant redistribution of revenue away from frontier labs toward a long tail of specialized models.
Open source AI isn't just idealism — it's a business. Hugging Face crossing $100M ARR proves the platform model works, even without aggressive monetization. And with surging interest in local models, the growth trajectory is accelerating.
Running AI locally on your phone or laptop is free, private by design, and ideal for sensitive workloads. From personal health conversations to heavy agentic tasks running 24/7, local models solve problems that API-based services structurally cannot.
Routing AI queries to the right specialized model — rather than defaulting every request to a frontier giant — could fundamentally redistribute revenue across the AI stack. A Stanford study found 70% of ChatGPT queries could be answered locally on a laptop. We're still in the first, simplistic phase of AI. The second phase is about routing, open source, and specialization.
Anthropic and OpenAI are the fastest-growing companies in the world and have become trillion-dollar players overnight. Clément Delangue has little sympathy for their competition complaints — and argues that society is better served by more competitors in AI, not fewer.
Young people have already burned through the consumer phase of AI and are moving into builder mode — fine-tuning models, curating datasets, building products in overlooked domains like climate, biology, and chemistry. This generational shift is Clément Delangue's biggest white pill for the industry.
Distillation is a universal practice across every AI lab — calling it theft while being the world's fastest-growing company is hard to take seriously. Clément Delangue argues that stopping distillation won't kneecap Chinese AI, and that Anthropic and OpenAI need more competition, not less protection.
Open source AI models are less generalist, less focused on dangerous capabilities, and benefit from the transparency of public scrutiny. The nuclear bomb — history's most dangerous technology — was never built in open source, and Clément Delangue argues the same structural logic applies to AI.
Europe has the talent, the energy infrastructure — especially France's nuclear power — and existing labs like Mistral operating at the frontier. Building a world-class frontier AI lab is achievable, but it requires fostering an open research ecosystem rather than trying to conjure a champion out of thin air.
Models can be taken away, can be biased, or can refuse to give you correct information. Fusion models that combine the capabilities of multiple models are gaining traction on platforms like OpenRouter as companies realize the danger of single-model dependency.
The US government's move to restrict GPT-4.6's release is unprecedented — but Clément Delangue says frontier labs brought it on themselves with years of doom marketing. He hopes the restrictions stay contained to a few frontier generalist models and don't spill over to startups, academia, or smaller players.
Remove an open-weight model from Hugging Face and it immediately reappears on ModelScope or torrent platforms. Unlike APIs — where access can be cut and data is transmitted — open weights give users full control and transparency by design, making country-of-origin largely irrelevant.
Analysis
What they talk about
- Technology 84%
- Business 8%
- Education 8%
Connections
Shows they appear on and people they share episodes with. Drag to explore.
