Hugging Face Raised $235M. Community Still Is Not a Moat

Hugging Face's Series D landed in late August 2023: $235 million at a $4.5 billion valuation. Nathan Labenz notes that the round is now more than half of all capital the company has ever raised, with total funding still just under $400 million. The investor list reads like a census of the compute stack — Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM, Salesforce, and Sound Ventures. Valuation doubled in a year.

The revenue numbers sit in a different universe. Labenz puts annualized revenue at roughly $30 million, which makes the multiple north of 100 times sales and closer to 150 times if that figure is right. He had seen the deal coming. Starting in the first half of July, expert-network calls about Hugging Face suddenly piled up from multiple firms asking the same questions. This episode is, in effect, the briefing he had already been giving investors, now on the record.

Co-host Erik Torenberg supplies the comparison set. Runway is valued over a billion dollars and is a bet on models. MosaicML sold for more than a billion as training scaffolding plus inference. Hugging Face, feature for feature and dollar for dollar, does not obviously out-execute those companies. What is special is the seat it occupies: the center of activity across the open-source AI ecosystem, more a community business than almost any other high-flying AI startup.

The company started in 2016 as a chatbot business, closer to Replika than to the model hub people know today. The turn that mattered was an open-source library, Transformers, that became canonical execution code, then a Model Hub on top of it where people post the weights that run on that library. Diffusers later covered image generation. Datasets have their own section. The hub even mirrors arXiv papers and hangs the weights on the paper page when authors open-source them.

Spaces host interactive demos, made much easier after Hugging Face acquired Gradio. A Pro account, about $9 a month, buys bigger servers and a little coding environment so a model can get a wrapper of inputs and outputs. Labenz first noticed the platform in early 2022 at Waymark, hunting for the best image-captioning and image-retrieval models. Cool demos kept showing up there first, a pattern that held for eighteen months. He later bought the Expert Acceleration Program: a monthly allotment of questions to a team that could talk through feasibility, fine-tuning, and the state of the art.

The team impressed him even when the answer was often "it depends on your use case." Traffic is not a vanity metric. While browsing for the episode, he found a niche fine-tune of Code Llama 34B with 35,000 downloads in a week. Tens of thousands of models sit on the hub; maybe hundreds do significant inference. His recurring product note is curation. The platform already has about twenty task categories — image captioning, speech-to-text, text-to-speech. Make the task page the front door, rank a top five by likes, downloads, and commentary, and let people try running demos before they commit.

The change in research workflow is the part he keeps coming back to. A few years ago, a new paper meant a week of reading, hunting for code, setting up an environment, maybe retraining, only to learn that academic numbers collapsed on real inputs. Academic labs published code mainly to prove reproducibility, not to hand you a production tool. Community norms flipped: "weights or it didn't happen," plus a working demo. The same evaluation can now take two hours — open a task page, try demos, run a small script, move on.

Torenberg starts the moat discussion with Digg. Before Reddit, it was the front page of the internet. Users rebelled over a redesign, switched almost overnight, and Digg never recovered. Nothing on the site was hard to recreate: a username and some posts. Nightclubs and Clubhouse illustrate the other failure mode. What makes a scene special is often exclusivity; as it scales, the bigger the community, the less special it feels. Negative network effects are real.

LinkedIn is the counterexample. People have spent fifteen years saying they hate it and still cannot leave. Connections, professional history, and recommendations are state on the site. Switching and coordination costs are too high for a mass rebellion. GitHub is the vertical version for engineers: code, commits, and follow graphs live there. Product Hunt and Hacker News never crossed that line. The leaderboard resets every day. Torenberg, who worked at Product Hunt, calls them not venture-scale businesses because they never compounded identity or data. The escape hatch would have been to become more like G2 — a historical review library, not a daily rat race.

G2, Yelp, and Glassdoor work because users arrive with a spending decision and reviews accumulate. Labenz adds that Yelp's market cap may already sit below Hugging Face's valuation: review businesses can go public, but they rarely become giants. Hugging Face's problem is sharper. Weights are downloadable. Demo code is open source. People show up every day, yet almost nothing on the plaza has to stay there. The community is real; the accumulated state is not.

The current commercial thesis is that discovery and compute belong together. Users find a model, click a few times, and Hugging Face serves it as an API via inference endpoints, taking a margin over the underlying clouds. That is the story that has to support $4.5 billion. The investor list — Google, Amazon, Nvidia, and the rest of the silicon and cloud complex — suggests they will not lack for capacity. Labenz likes the vision. He is a paying customer with no financial stake in the company, and he joined the endpoints beta.

The product, he says, is not there yet. OpenAI's GPT-3.5 fine-tuning is the comparison that hurts: fine-tune, call it a hundred times immediately, never call it again, and the system scales and prices around that convenience. Hugging Face endpoints tend to stay on, autoscaling is rougher, and the surface is less mature. Optimizing across tens of thousands of models and wildly different workloads is a hard engineering problem even before competition.

Switching is frictionless by design. The whole premise is that the model is open, so a sophisticated developer can take the same weights to Amazon, Google Cloud, or Azure — Microsoft is notably absent from this round. Demo code is open too. Labenz's own pattern is to pay $9 a month for a Pro account while prototyping, then shop on price once traffic is real. Community affection does not follow the workload. It is a lead-gen function for an inference product that still has to win head to head. Demand for compute is growing so fast that every player may grow for a while. That is not the same as a moat.

Hugging Face has welded its corporate identity to open source, including advocacy in places like the United States Senate. The community that loves it most treats open access as an inherent good. That stance sits in tension with AI-safety hawks, who are less worried about today's models than about the pattern of releasing systems before anyone knows what they do. Hugging Face discourse, Labenz says, does not leave much room for that concern. Egalitarian access is the whole story.

The Open LLM Leaderboard makes the tradeoff concrete. GPT-4 is not on it. GPT-3.5 is not on it. Anthropic's models are not on it. The board is honest about being an *open* language-model ranking, and still fails the job Labenz wants: show the best models, and show how open weights compare with the commercial frontier. Sophisticated buyers already know closed models are winning. Pretending that slice of the market does not exist shrinks the curation value of the hub.

From an app developer's chair, the full menu is the product: every model, how well it performs, what it costs, what the rate limits look like. Commercial APIs are appealing precisely because someone else handles the messy parts. Hugging Face tries to bridge that by running open-source models for you, but Labenz still wants the complete set. Advertising on the G2 model is a weak analogy anyway — nobody buys ads for a fully open-source checkpoint. Revenue share on hosted models, or training a closed model of their own, would collide with a community trained to expect everything on the hub to be open and free.

GitHub sits one product decision away. Model files are downloadable; scraping similar public data has survived court challenges in the LinkedIn cases. If GitHub mirrored the hub, added live demos, and put commercial models on the menu, Hugging Face would have to keep an ideological community happy while competing on a complete catalog. Torenberg's curation question is whether Hugging Face can build a compounding review layer with high spend and some turnover, more G2 than Product Hunt. Labenz's worry is the opposite of evergreen: anything uploaded before ChatGPT already feels dated. Lose focal status for a month and the archive ages out.

Asked what Hugging Face looks like as a $50 billion company, Labenz reaches for GitHub. Microsoft paid $7.5 billion in 2018, and he does not sense regret. Developers pay a few dollars a month mainly to keep code private; Hugging Face's $9 Pro tier is the same shape. What Microsoft actually bought was hooks into tens of millions of developers. At a two-trillion-dollar market cap, the deal did not need a clean ROI. It was a way to stay unavoidable in the developer stack, later filled in with Codespaces and Copilot.

Hugging Face could sell a similar story. Labenz's base case is not that revenue catches the valuation. If the company sold for $20 billion in two or three years, revenue would still likely trail the price. The buyer would be paying for position. "ML is the next programming in general. It's so massive and the people that do it are so important, owning that platform becomes like a strategic priority." Google and Amazon do not own GitHub and want to compete with Microsoft everywhere. An exit to one of them, he says, would be "supernatural."

That buyer, in this telling, would not care how many people pay $9 a month, and would not even need the inference product to be truly top-notch. They would want a community layer on a mega-stack. Anyone buying in at $4.5 billion, Labenz argues, is underwriting that story more than an independent path to matching revenue. World-class product execution on the Replit pattern is the other candidate, and he does not think Hugging Face's community-driven messiness is aimed at that kind of bite out of cloud.

Labenz's phrase for Replit is world-class execution. Write code, hit publish, the app is live and autoscales. The CEO had just posted tests sending a quarter million requests at a personal site hosted on Replit; commenters were uniformly stunned. Labenz puts Replit in the same tier as Cursor and Copilot as one of the best ways to create, then adds the easiest path from that file to a running app. Existing large engineering projects will not migrate. The long tail of new small apps might be born there, and the open question is how much of that growth Replit can keep as those apps mature.

If a product that starts on Replit can scale to millions of users without a cost or reliability cliff, the ceiling is high. If it cannot, the business caps at small-to-medium apps. Replit is also building an AI developer as a pair programmer, plus a bounty economy in which users pay users, users pay AIs, and eventually AIs delegate to AIs. That stack is coherent: be the easiest way to build and deploy, then invent whatever commercial machinery supports it. Hugging Face, by contrast, is diffuse. Labenz would be shocked if a single Copilot-like experience dropped across the hub. Talk-to-Hugging-Face helpers exist, but the product surface keeps sprawling.

BLOOM is the cautionary Hugging Face story. The open-science project trained a 175-billion-parameter model, analogous to GPT-3, to prove an open community could do it in the open. They did: a huge dataset, consensus decisions, hundreds of people. It was a genuine coordination achievement. It was also not nearly as good as GPT-3, was not iterated, and is now mostly a footnote with little usage. Replit fine-tunes code models on how its own platform is actually used. One team works under a belief system; the other aims at a product goal. The tempo difference shows.

Runway and Character generate more wow. Labenz's teammates at Waymark are making films with Runway's latest model that they could not have afforded to shoot. Character has obvious product-market fit around creative companions. Both sit around a billion-dollar valuation. If he were allocating purely for financial return, he would still rather own Hugging Face. Community, discovery, and trusted curation are real. A model company has to stay on the frontier, because a strictly better video model tomorrow could crater usage. Switching costs are close to zero.

Image generation is the exhibit. The original Stable Diffusion went open source about a year before this conversation. Almost nobody uses that checkpoint now. The field has been through something like ten generations, with leapfrogs among DALL-E, Adobe Firefly, Midjourney, and others. Midjourney has shown it can hold a frontier position, but the month Adobe ships something strictly better, Runway's users are already Adobe customers with a credit card on file. The new feature may even arrive as part of an existing license.

Zooming out, neither host thinks it is obvious where in the AI stack massive value will accrue. Incumbents that own infrastructure — the clouds Hugging Face and Runway are built on — still look likely to capture a lot of it. Hugging Face's relative valuation is not crazy given the hub it actually is. Playing that hand into an independent giant, Torenberg says, is strategically very hard. Labenz's close is that we will know in a few years, and that even a strong community is a difficult position to monetize on the terms this round implies.