Community Announcement:
We are also building DeepVista to help you build a self-improving system for your company. If you are interested, please subscribe at deepvista.substack.com. We will share private beta invitations and regularly publish updates and tactical tips on running agent-native operations there.
In our previous essay, we explained why a true competitive moat lies in compounding “slow assets”—those that require time, friction, and operational grit to mature—and took a deep dive into how Red Bull did just that with its distribution network.
Now, let’s examine our second category of slow assets: ecosystem. There is no better example of this than Hugging Face. As the bedrock of open-source AI, Hugging Face hosts 2.5 million public models, fueling global innovation and democratizing machine learning.
Moreover, they created an ecosystem with a strong moat. Participants in the Hugging Face ecosystem generate value for one another, which led to a powerful “soft lock-in”: the more developers build habits around the platform, the harder it becomes for the industry to migrate.
In this essay, we’ll examine how they build slow assets and use them to create an ecosystem moat.
Hugging Face started with a failed consumer chatbot.
Originally, the team was building a Tamagotchi-like conversational AI for teenagers. For three years, they hit a wall. To make their chatbot hold an intelligent, human-like conversation, they had to coordinate dozens of specialized, narrow models—one to classify emotions, another to process images, and a third to generate text replies.
The early team ran into a brutal operational reality: back then, natural language processing was highly fragmented. Machine learning models were siloed, written in incompatible custom code, and incredibly painful to deploy. There was no standard interface or a simple way for different models to talk to one another.
When Google open-sourced its landmark BERT language model in late 2018, the Hugging Face team saw the real opportunity. They realized the world didn’t need another wrapper application. It needed a unified translation layer and a central repository for the models themselves.
So they abandoned the chatbot and pivoted entirely to a model registry play, betting that as AI matured, developers would need a central marketplace to share, test, and host model weights.
But a marketplace has a classic cold-start problem:
How do you bootstrap a registry when you have zero models and zero users?
In ecosystem economics, the cold start is the moat’s constructor. If bootstrapping were easy, any competitor could replicate it. The operational friction of starting from zero is precisely what makes the resulting network defensible.
In an AI model registry, the ecosystem has two distinct tiers: upstream researchers (the academic creators who experiment and publish weights) and downstream developers (industry engineers who deploy and scale them). Upstream researchers are the core slow asset; downstream developers will only gather where pre-trained models are already abundant.
In late 2018, Google released the state-of-the-art BERT model exclusively in TensorFlow, aiming to lock enterprise developers into Google’s cloud and TPU hardware ecosystem. This created a structural divide: TensorFlow optimized downstream production scalability but sacrificed the flexibility upstream researchers needed for rapid, iterative experimentation. For those using PyTorch for its dynamic iteration, BERT was functionally inaccessible.
Thomas Wolf, co-founder and CSO of Hugging Face, faced this exact friction. As a researcher himself, he needed a PyTorch-compatible BERT to fine-tune models for their chatbot. To bypass TensorFlow’s rigidity, Wolf spent a weekend porting BERT to PyTorch and uploaded the raw code to GitHub.
“I remember really vividly Thomas told us, oh, there’s this new transformer model that came out from Google. It’s amazing, but it sucks because it’s in TensorFlow.”
Solving his own pain point resolved the exact bottleneck paralyzing the upstream research community.
The PyTorch BERT repository became their initial wedge, focusing all their energy on winning over that tiny, elite group of upstream creators. This early win kickstarted Hugging Face’s engine for long-term compounding: once academic researchers started hosting their models on the platform, downstream developers followed naturally.
Once you bootstrap your supply of producers, you have to retain them. In developer tools, the ultimate retention strategy is simple: absorb the friction so they don’t have to.
In 2019, the biggest barrier to AI contribution was framework fragmentation.
The community was split between PyTorch and TensorFlow. If a researcher wrote a model in PyTorch, it was useless to developers using TensorFlow. To reach the whole market, producers faced a massive friction barrier: they had to manually rewrite and maintain two entirely different codebases.
Instead of forcing developers to pick a side or do double the work, Hugging Face built the translation plumbing.
They did the painstaking engineering to make their “Transformers” library cross-compatible. Within a single workflow, a developer could build in PyTorch and use it in TensorFlow with three lines of code.
Hugging Face absorbed the tedious, framework-level engineering friction on their side so developers could have a friction-free experience on theirs. Model producers could now publish once and automatically reach 100% of the market.
By removing the framework tax, Hugging Face made contributing to their registry the absolute path of least resistance.
Once you have high-value producers and zero contribution friction, the final step is to activate the ecosystem’s commercial engine: creating a give-and-take flywheel.
In an ecosystem moat, monetization is never front-loaded. It is a downstream byproduct of absolute community adoption.
The flywheel operates on two distinct gears:
The “Give” (Distribution standard): Hugging Face keeps the entire platform free for individual researchers and developers. Hosting models, indexing datasets, and downloading weights costs nothing. By absorbing these multi-million dollar infrastructure costs, Hugging Face ensures that every AI practitioner’s default habit is rooted in their platform.
The “Take” (Enterprise monetization): Once a company’s engineering team has built their entire workflow around Hugging Face’s free standards, the enterprise monetization triggers naturally. Large organizations pay for secure, private hubs, dedicated GPU compute, and enterprise-grade deployment support. In 2024, this enterprise layer generated $130M ARR.
While the base model registry layer has largely consolidated, the playbook of building a slow-asset ecosystem remains highly relevant. Today, fragmentation has shifted from model hosting to the next layer of the stack: AI agents, tools, and skills.
Currently, the agentic layer is as fragmented as NLP was in 2017. Different agent frameworks have proprietary protocols for calling and executing external tools. If a developer builds a high-value skill—like a secure database write or a web action—they must manually rewrite execution wrappers for multiple frameworks.
If you are building in the agent space, you can apply Hugging Face’s playbook to this fragmentation:
Target the producers: Focus on the developer community creating specialized APIs and discrete agent skills.
Absorb the friction: Build a unified, open protocol that translates any skill into a format any agent framework can execute out-of-the-box, eliminating the framework tax.
Run the flywheel: Keep this registry free for developers (the Give) and charge enterprises for secure execution, credential management, and auditing (the Take).
Instead of competing to build another standalone agent, you build the default infrastructure where agent actions are traded.
Hugging Face proved that an irreplaceable ecosystem moat is built by securing an initial wedge with the highest-leverage producers, absorbing workflow friction for the broader community, and forming a self-reinforcing, give-and-take flywheel.
Next week, we’ll conclude this series by breaking down how the “FDE” (Forward Deployed Engineer) mode captures unwritten operational workflows and build compounding data moats that competitors cannot copy.
Subscribe to our newsletter today to ensure you don’t miss this final deep dive. Stay tuned!
Never miss a FounderCoHo community event, podcast, or newsletter. Join us:
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.