Turn a well-timed open release into the neutral commons your whole field publishes to, then let the community's contributions compound faster than any closed competitor can ever catch up.
Hugging Face
Hugging Face began in 2016 as a consumer chatbot for teenagers, then pivoted into the open-source infrastructure that most of modern AI is built on top of, often called the "GitHub of machine learning." Founded in New York by Clément Delangue, Julien Chaumond, and Thomas Wolf, it turned a single well-timed code release into a platform, a self-reinforcing community, and eventually an entire category. The through-line across its stories: become the neutral commons a whole field standardizes on, then let the network compound faster than any competitor can catch.
Platform & Ecosystem: from a library into the hub every model passes through
The problem. After Google released BERT in late 2018, Hugging Face's PyTorch reimplementation (first pytorch-pretrained-bert, then Transformers) became the way engineers ran state-of-the-art models. But a popular library alone is just code you install. The trained models, the datasets, and the demos all lived scattered across university servers, personal Google Drives, and academic repos, with no common place to publish or discover them.
The approach. In 2020 Hugging Face launched the Hub: a hosted repository where anyone could upload, version, download, and share models and datasets, wired directly into the Transformers library so a model was one line of code to load. It layered a full ecosystem on top, adding Datasets, and later Spaces for hosting live demos, turning the library into a two-sided platform of producers and consumers.
How it solved it. The Hub grew into over a million models and roughly 250,000 datasets, and its Inference API alone handles on the order of a billion requests a month. Rather than competing with the frameworks below it, Transformers became the pivot layer that PyTorch, TensorFlow, and downstream inference engines all target, making Hugging Face the connective tissue of the ecosystem rather than one tool in it.
Network Effect: every uploaded model makes the next researcher's choice obvious
The problem. A model repository is only as useful as what is already in it. Early on, Hugging Face faced the classic cold-start bind: researchers would not publish to a hub nobody used, and nobody would use a hub with nothing published. Meanwhile large labs had every incentive to keep releasing models on their own fragmented terms.
The approach. Hugging Face made publishing frictionless and free, gave contributors distribution and credit, and standardized the format so any uploaded model was immediately usable by everyone else through the same Transformers API. Each new model, dataset, and Space raised the value of the platform for the next user, who then had more reason to contribute back, a direct network-effect flywheel built on an open commons.
How it solved it. The compounding is visible in the numbers: the platform now serves roughly 13 million users across some 500,000 organizations, with more than 30% of the Fortune 500 holding accounts. New fields adopt it by default too, with robotics datasets on the Hub jumping from about 1,145 in 2024 to nearly 27,000 in 2025. Investors priced the moat accordingly, backing a September 2025 round of about $2 billion at a $13.5 billion valuation.
Category Creation: defining "the GitHub of machine learning"
The problem. Before Hugging Face, there was no accepted category for a neutral, central home for open machine-learning models. Frameworks existed (PyTorch, TensorFlow) and model zoos existed inside individual labs, but no one owned the idea of a shared, version-controlled commons where the whole field publishes and collaborates. A consumer chatbot startup had no obvious claim to that space.
The approach. Hugging Face reframed itself around a borrowed but sticky mental model: be to machine-learning models what GitHub is to code. The 2019 rebrand of the library to Transformers, the deliberately open and non-proprietary stance, and the 2020 Hub launch all served to define and occupy that category before anyone named it, positioning the company as infrastructure rather than as any single product.
How it solved it. The category label stuck, and Hugging Face became its default answer: by 2023 the company had raised $235 million from a roster of would-be rivals including Google, Amazon, Nvidia, IBM, and Salesforce, an unusual signal that the incumbents preferred to invest in the neutral commons rather than fight it. Owning the category, not just a product, is what let a former chatbot maker sit at the center of open AI.