Clement Delengue - Founder at Hugging Face

Imagine building a $4.5 billion tech empire by, you know, taking your most valuable assets and intentionally giving them away.

Right. Just completely giving them to your competitors for free.

Exactly. I mean, just open sourcing the very thing that other companies are spending billions to lock away in these secret data centers.

It really defies every traditional rule of Silicon Valley economics, doesn't it? I mean, the assumption has always been that if you build a revolutionary piece of artificial intelligence, You, well, you build a massive wall around it.

Yeah. A walled garden.

Exactly. You charge for API access, you restrict the usage, and you treat that underlying model as a highly classified trade secret.

But today we are unpacking a totally different playbook. Welcome to this deep dive into a really fascinating stack of sources, primarily looking at a text called Hugging Face: The Architecture of Open Intelligence.

That's a great read.

It really is.

Yeah.

And our mission today as we explore this with you is to look at the democratization thesis of AI. We want to figure out how this unsexy, like deeply technical infrastructure play became the ultimate competitive moat by building a community rather than hoarding data. Right. And we are going to trace this whole thing through the story of Hugging Face and its founder, Clement Delang.

The sheer scale of disruption here is just staggering. We are looking at a fundamental shift in where the power actually lies in tech right now.

Yeah. On one side, you have those walled gardens we mentioned, right? The corporate labs trying to control the intelligence layer.

And then on the other side, you have this massive decentralized network effect that basically insists intelligence should be a shared utility.

So to understand how a company gets a $4.5 billion valuation by essentially giving things away, we actually need to look at Clemente Lang's background. Because he didn't start out in some traditional Silicon Valley incubator.

No, not at all. He grew up in a small town in northern France, La Bassée.

Right. And by the time he was 17, he was already running this super successful eBay business.

Yeah. Which is telling, right? He goes on to ESCT Business School, does these stints at a machine learning startup called Moodstocks.

Which Google actually acquired later, right?

Exactly. And then he works at a couple other startups like Mention and Videonaut.es. You can really see the through line there.

Yeah. He's steeped in early internet commerce.

Right, and these agile startup environments. He's internalizing how digital ecosystems scale, and, you know, he's watching the internet itself operate on these open protocols like HTTP and TCP/IP while the actual tech power was kind of centralizing into just a few massive corporations.

He sees that contradiction. And so in 2016, he teams up with two co-founders, Julian Shalman, who later founds Hugging Face, and Thomas Wolf, who is this brilliant machine learning researcher with a PhD in physics.

It's a powerhouse team.

For sure. They started hugging face. But wait, here is the wild part. The initial thesis in 2016 was not to build a GitHub of AI.

No, not even close.

It was to build a chatbot for teenagers.

Yeah, an open domain conversational AI. I mean, if you think back to the landscape in 2016, it was totally dominated by Siri and Alexa.

Right, which were super productivity-focused utilitarians.

Exactly. You used them to set timers or, I don't know, check the weather. But Clement and his team wanted to build something closer to an AI Tamagotchi.

An AI Tamagotchi. I love that.

Right. Just a digital companion that teenagers could chat with about literally anything.

Which mechanically is an incredibly hard technical problem. I mean, open domain conversation means the AI has to handle any topic, any weird slang, any totally random tangent a teenager throws at it.

Oh, for sure. So the team is furiously building these natural language processing models, just trying to make their chatbot smarter.

And then October 2018 happens.

It's a big inflection point.

Yeah. Google releases BERT. Let's unpack this for a second because why did that matter so much?

Well, BERT was a seismic shift in natural language understanding. It was this transformer model that could understand the context of a word based on all its surrounding words.

Rather than just reading left to right.

Exactly. It fundamentally changed what AI could actually comprehend.

But, and here's the catch, Google releases BERT in TensorFlow. And if you've ever tried to work with AI models, you know TensorFlow is Google's in-house machine learning framework.

Right. It's incredibly powerful and highly optimized for production environments, but it was notoriously rigid.

Especially for researchers who just wanted to experiment, right?

Yeah, that rigidity is exactly where the friction lived. The broader machine learning research community was rapidly migrating toward PyTorch at the time.

Because it was way more flexible for prototyping.

Exactly. So you have this massive breakthrough model from Google, but it's built in a framework that researchers are actively trying to move away from.

And this is where Thomas Wolf steps in. Over the course of a single week, he basically manually ports Google's BERT model from TensorFlow over to PyTorch.

A one-week sprint. It's incredible.

He just translates the architecture, packages it up in a library called PyTorch Transformers, and drops it on GitHub.

And the reaction was explosive. I mean, it went incredibly viral.

Which is wild, 'cause it wasn't some calculated pivot strategy hatched in a boardroom.

No, it was just a developer scratching their own itch.

Right. They needed PyTorch BERT for their teenage chatbot, so they built it, shared it, and suddenly the entire world wanted it.

That one week was the clearest market signal a startup could possibly ask for. Developers didn't care about the Hugging Face chatbot.

Not at all.

They cared about the open-source tooling that made these complex state-of-the-art models actually usable. Hugging Face realized, wow, we are sitting on a goldmine of utility here.

Because they took the friction of translating massive complex architectures and reduced it to what, a few lines of code?

Yeah, pretty much.

Let's actually look at the mechanics of that friction for you listening, because I think it's easy to underestimate how hard this was. Before Hugging Face, if a lab released a new model, You had to read the dense research paper, decipher all the math, figure out the specific neural network architecture.

And write hundreds of lines of code just to load the thing.

Exactly. Hugging Face abstracted all of that away.

Uh-huh.

By September 2019, they renamed their library to simply Transformers. They released version 2.0.0, and they added support backend for TensorFlow 2.0.

Which was brilliant. That interoperability was really the killing blow to the walled gardens. By supporting both PyTorch and TensorFlow, Hugging Face became the Switzerland of AI.

I like that. The Switzerland of AI.

Right. You could train a model in one framework and seamlessly run it in the other. It just eliminated the vendor lock-in that the major tech giants were relying on.

And that creates this massive behavioral shift. The sources point out that by 2020, researchers publishing papers on Arxiv stopped linking to their own messy GitHub repositories.

Yeah, they started linking directly to Hugging Face model cards instead.

Which a model card is essentially a standardized resume for an AI model.

Exactly. It tells you what it does, what data it was trained on, its limitations, its biases, all of that.

But I've always wondered about that transition. Why couldn't GitHub just handle this? Why did Hugging Face become the de facto platform when GitHub already existed for sharing code?

Because a large language model isn't just code. The code to run a transformer is actually quite short. The real bulk of the model is the weights.

The weights. So the gigabytes or terabytes of mathematical parameters.

Right. The stuff the model actually learns during training. GitHub was built to version control lines of text. It absolutely choked when people tried to upload 50-gigabyte tensor files.

Oh, wow. Yeah, that makes total sense.

Yeah.

So Hugging Face built an infrastructure specifically designed to host version control, and instantly deploy massive binary weight files alongside the code.

They completely solved the storage and distribution bottleneck.

And the venture capital world definitely notices. They secure seed funding from Betaworks early on, then a $15 million Series A from Lux Capital in December 2019, followed by a $40 million Series B in March 2021. And the community momentum just keeps snowballing, which leads us to 2021 and a project called Big Science.

Oh, big science. This is arguably one of the most fascinating experiments in the history of open source software.

Right, because at this point, OpenAI had released GPT-3, which was proprietary and totally locked behind an API.

And the community's response, which was orchestrated heavily through the Hugging Face ecosystem, was to build an open source equivalent.

A completely volunteer-led effort. I mean, you had hundreds of researchers from all over the world pooling their computing resources and expertise to build an incredibly powerful 176 billion parameter their model called Bloom.

The goal was simple, really. Match the performance of proprietary models, but give it away for free.

It proved that decentralized, community-driven development could compete directly with the wealthiest corporate labs in the world.

It did, but it also highlighted a massive operational challenge. Because a model like Bloom is huge. You can't just download it onto your laptop and run it.

Which brings us to arguably the most critical question of this entire deep dive. If you're Hugging Face and you are hosting petabytes of data, serving millions of downloads a day, and giving away all these models for free, how do you avoid going bankrupt? Yes, the server costs alone must have been astronomical. You cannot pay AWS bills with GitHub stars.

No, you definitely can't. That was the existential threat for them, and their solution is honestly a masterclass in infrastructure monetization. Hugging Face realized that while downloading a model is easy, deploying it in production is a complete nightmare.

Okay, let's dig into that nightmare. Say I download a massive open-source model. I've got the weights on my machine. Why can't I just put it on a server and start a startup?

Because the raw physics of compute. A massive model requires specialized GPUs, which are incredibly expensive, and you run into VRAM limits.

The memory on the graphics card itself.

Right. So you have to shard the model, meaning you split it across multiple GPUs. Then you have to manage the latency between those GPUs.

That's awful.

It is. And if your app suddenly gets a spike in traffic, you have to dynamically spin up new clusters, synchronize them, manage the load. I mean, for a typical startup, that requires an entire team of highly specialized machine learning operations engineers.

So Hugging Face looks at this nightmare and says, we will solve that headache.

Exactly. They launch inference endpoints. Models as a service.

It's a brilliant pivot. They essentially tell developers, hey, the model is free, the code is free, but if you want to deploy this model into a live application with one click and let us handle all the GPU provisioning and the auto-scaling, you pay us a fraction of a cent per request.

It's pure convenience layer.

It reminds me of AWS. Amazon didn't invent websites, right? And they didn't write the code for your startup. They just realized that hosting the infrastructure and managing the server racks was where the real reliable money lived.

And that convenience layer scales beautifully. Look at their revenue trajectory. In 2022, Hugging Face did about $15 million in revenue. By the end of 2023, their estimated annual recurring revenue shot up to $70 million.

That is a 367% year-over-year growth rate. Just massive.

And that kind of growth attracts serious capital.

Oh yeah. In April 2022, they raise a $100 million Series C from Sequoia, hitting a $2 billion valuation. And then just 15 months later, in August 2023, they raise a massive $235 million Series D.

And look at who the investors in that round are. It's incredibly telling. Salesforce, Google, Amazon, NVIDIA, the absolute giants of the industry. Yeah.

And the valuation doubles to $4.5 billion.

What's really remarkable about those investors is that many of them are the exact same companies is building the proprietary models. Google and Amazon invest heavily in their own walled gardens, yet they still poured money into Hugging Face.

Okay, wait, let me push back on that. Because why? If Google is trying to lock people into its own AI ecosystem, why fund the company that is actively tearing those walls down?

Because the network effect became too powerful to ignore. By April 2026, Hugging Face had over 13 million users, 500,000 organizations, and over 2 million models on the hub.

Plus over a million repositories.

Right. And over 30% of the Fortune 500 had verified accounts. The Transformers library had over 121,000 GitHub stars. Hugging Face just became the gravity well of the AI ecosystem.

So if you are NVIDIA, you want to partner with them because they provide the training clusters as a service.

Exactly. And if you are Google, you want your open source models distributed there because that is where every single developer actually lives.

So the community itself became the moat.

Yes. It's not about having the smartest algorithm anymore. If you build a slightly better proprietary model, sure, someone might use it. But if you build the platform where every researcher on Earth comes to collaborate, you become indispensable.

You become the infrastructure of innovation itself. And this infrastructure moat allows them to expand really aggressively into new frontiers.

Definitely. Look at what happened in 2025. There were two massive shifts. First, we see the release of models from companies like Deepseek.

Right. These are open-source models that are suddenly benchmarking completely competitively with OpenAI's GPT-4.

That was a huge watershed moment. For years, the argument was always that open source would lag, you know, a generation behind the proprietary labs.

It was almost taken as a given.

Right. But the Deepseek models proved that a decentralized community iterating rapidly on open architectures could totally catch up to the frontier. And of course, the primary distribution channel for those models and all the fine-tuned variants the community spawned was Hugging Face.

But 2025 also saw Hugging Face step out of the digital realm and into the physical world. They acquired a company called PaLM Robotics, which is a super fascinating strategic move.

PaLM Robotics focuses on open-source robotics, so Hugging Face is taking their democratization thesis and applying it to embodied AI.

They want the Hugging Face Hub to be the brain center not just for chatbots but for physical machines.

Exactly. And they're partnering with Meta and UNESCO on multilingual AI too. They really wanna be the central nervous system for everything.

Which brings us to kind of the contradiction at the heart of this entire movement. We've talked about the democratization of intelligence, like it's this unalloyed good.

Mostly, yeah.

Because lowering the barrier to entry so anyone can build an AI app is amazing for innovation. Mm-hmm. But open source fundamentally means zero control.

It is the defining tension of the platform. Hugging Face is hosting over 2 million models. They provide the infrastructure, but they do not actively police every single output of those models.

Right. If you democratize access to frontier-level intelligence, you are also democratizing access to highly capable tools that can absolutely be misused.

And this is where the philosophy gets really messy. The walled gardens argue that they need to keep models closed for safety.

To prevent, like, industrial-scale disinformation or deepfakes or automating cyberattacks.

Right.

So if an open-source model capable of doing those things is freely available on Hugging Face, doesn't that shift the liability? I mean, it's one thing to democratize a printing press, but isn't democratizing GPT-4-level intelligence more like distributing a loaded weapon to anyone who asks?

It's a heavy question.

If someone downloads a model and uses it to generate a bioweapon recipe, Hugging Face didn't build the model, sure, but they provided the one-click deployment. Can they really wash their hands of the responsibility just because They are the middleman.

You're pointing to the absolute crux of the AI alignment debate right now. The Hugging Face philosophy pushes back against the idea that 5 CEOs in Silicon Valley should get to decide what is safe for the entire human race.

Right. They prefer decentralization.

Exactly. Their stance is that transparency and open community review are the only real ways to build robust, secure systems. Security through obscurity.

Like hiding the models.

Yeah, that has historically failed in software engineering, but you are absolutely right that AI is not just normal software.

It's closer to infrastructure. And if the infrastructure is truly open, then the responsibility of alignment, of making sure these models are actually beneficial and not harmful, is suddenly decentralized too.

It's no longer happening behind closed doors at OpenAI or Anthropic. It's happening in the public square, and the public square is inherently chaotic.

It forces a massive societal reckoning. By making the deployment of these models frictionless, Hugging Face has ensured that the impact of AI, both the good and the bad, will be felt instantly across the globe. They removed the bottleneck of distribution entirely.

Completely.

So to synthesize all of this for you listening, we are looking at a trajectory from a teenager selling on eBay in a small French town to a 2016 chatbot side project to a platform hosting 2 million models with a $4.5 billion valuation.

Mm-hmm.

And the entire empire is built on the realization that in the age of artificial intelligence, the actual models are destined to become commodities.

The proprietary edge is fleeting. A breakthrough algorithm today is just replicated by the open-source community tomorrow.

Right.

The enduring value lies in the ecosystem that connects the researchers, the data, and the compute power. Hugging Face won because they recognized that reducing friction for developers is just a much stronger business model than hoarding data.

They built the toll road instead of the car, and the entire industry decided to drive on it.

Well put.

But this leaves us with a final, slightly unnerving thought for you to mull over. Right now, Hugging Face is the central library where human developers go to download, fine-tune, and deploy AI models. It's a tool built by humans for humans. For now. Exactly. But as these open-source models become increasingly autonomous and capable of writing their own code, what happens on the day an AI agent autonomously logs into Hugging Face, analyzes the latest open-source architectures, and downloads a newer, better model to upgrade itself?

Completely bypassing humans.

Right. When the library starts reading its own books and deploying its own improvements without human intervention, who's really in control of the open-source revolution?