Hugging Face

#1In August 2026, a strange acquisition story appeared.

#2Nvidia had reportedly agreed to acquire Hugging Face for about $12.9 billion.

#3The deal had not been publicly confirmed as completed. But the reported price was only the first odd thing about it.

#4Three years earlier, Hugging Face had been valued at $4.5 billion.

#5And among the companies that had invested in it were Google, Amazon, Nvidia, AMD, Intel, Qualcomm, IBM, and Salesforce.

#6They competed in chips. They competed in cloud computing.

#7They competed in AI. Yet they had all put money into the same startup.

#8Hugging Face had even reportedly turned down an earlier proposal for Nvidia to invest another $500 million, in part because it wanted to preserve its independence rather than let one investor become too influential.

#9So what exactly had Hugging Face built that so many rivals wanted to be part of it?

#10The answer begins far from Silicon Valley, with a teenager storing imported vehicles at his father's garden-equipment shop in northern France.

#11Clément Delangue grew up in La Bassée, a French town of roughly 6,000 people.

#12At 17, he started importing ATVs and dirt bikes from China and selling them on eBay.

#13There was a practical problem. Where do you put the inventory?

#14Delangue used his father's garden-equipment shop.

#15That early online business eventually helped him land an internship at eBay. But another encounter pulled him in a different direction.

#16At an e-commerce event, he met a cofounder of Moodstocks, a small French startup working on image-recognition technology—AI that could identify what a camera was looking at.

#17What struck Delangue was not simply the technology.

#18It was the size of the company building it.

#19A tiny AI startup could attempt something that looked like the territory of giants.

#20Delangue chose not to extend his time at eBay and spent time with Moodstocks instead.

#21Moodstocks would eventually be acquired by Google.

#22Delangue didn't follow it into Google. He started a company of his own.

#23It didn't become a major success.

#24Delangue's earlier startup was called UniShared.

#25Its idea was collaborative note-taking: students could create lecture notes together and make that knowledge available to others.

#26Someone shouldn't need to attend Harvard or Stanford to benefit from what was being taught there.

#27The premise already contained two ideas that would later sound familiar.

#28Build together. Share what gets built. But UniShared struggled.

#29Delangue later reflected on a problem he had encountered there: people who want to collaborate are not necessarily the same people who want everything to be open.

#30He also learned that pushing a set of values too far, too quickly, could make it harder to reach a broad audience.

#31The startup didn't work out. The values didn't disappear with it.

#32Years later, Delangue would explicitly connect the openness and collaboration he had pursued at UniShared with what excited him about Hugging Face.

#33Only the people sharing things would be different.

#34They wouldn't be students sharing notes.

#35They would be people building AI.

#36Delangue knew Julien Chaumond through the French startup world.

#37Chaumond, in turn, had an old connection to Thomas Wolf.

#38Chaumond and Wolf had attended engineering school together.

#39For a while, they even played in a rock band that covered Alanis Morissette.

#40Wolf's path after that hardly looked like preparation for founding one of the world's best-known AI platforms.

#41He earned a PhD in physics. Then he studied law.

#42Then he spent roughly five years working as a patent attorney.

#43Around 2015, deep learning pulled him back toward technical work. Some of the mathematics looked familiar from statistical physics.

#44He began teaching himself machine learning.

#45Wolf also became part of a study group connected to an online Stanford engineering course that Delangue and Chaumond were taking.

#46Eventually, the three decided to work together.

#47Their ambition was not to build infrastructure for the world's AI developers.

#48They wanted to make an AI friend for teenagers.

#49In the summer of 2016, Delangue and Chaumond went to New York.

#50Betaworks put roughly $200,000 behind them through its chatbot accelerator, a program that funded and supported young companies building conversational software.

#51The idea was difficult to finance in France.

#52It also had no obvious business model. This wasn't supposed to be an assistant that booked appointments or completed tasks.

#53The product was meant to talk with you, play with you, and develop something resembling an ongoing relationship.

#54They needed a company name before the trip.

#55They chose the name of an emoji. 🤗. Hugging Face.

#56In New York's Meatpacking District, Delangue and Chaumond spent three weeks working together almost day and night.

#57They had known each other before, but this was the first time they had spent more than a day working side by side in person.

#58Years later, Chaumond would remember those weeks as the beginning of their “bromance.”

#59The company kept building the AI companion for nearly three years.

#60From a distance, what happened next looks like a familiar startup story.

#61Consumer chatbot doesn't work. Company pivots.

#62Developer platform succeeds. Except the first part isn't right.

#63People were actually talking to Hugging Face's AI.

#64A lot. The founders later said users exchanged more than a billion messages with it.

#65Investors were interested too. In 2017, the company raised $1.2 million. The backers included Richard Socher, then a prominent figure in Salesforce's AI work, and NBA star Kevin Durant.

#66This was not a team desperately abandoning a product nobody touched.

#67But building the chatbot required Hugging Face to develop increasingly sophisticated language technology.

#68The team began releasing some of that work publicly.

#69Something unexpected happened. A second audience appeared.

#70Not teenagers looking for an AI friend. Developers and researchers trying to build AI themselves.

#71Then Google released BERT.

#72BERT was a new kind of language model from Google that became a major advance in natural-language processing—the technology used to help computers work with human language.

#73Google's implementation used TensorFlow, one of the main software frameworks researchers used to build machine-learning systems.

#74Thomas Wolf noticed something else: many researchers were working in another framework, PyTorch.

#75He suggested that Hugging Face port BERT to PyTorch and release it.

#76It wasn't a declaration that the company should abandon its business.

#77It wasn't a master plan for a new platform.

#78The idea was much simpler. People would probably find this useful.

#79The team produced the implementation in about a week and open-sourced it, meaning anyone could inspect, use, and modify the code.

#80Developers came. They downloaded it. They modified it.

#81They asked for support for other models.

#82They used the code in their own projects.

#83Something Hugging Face had released while building its own AI was suddenly spreading faster among developers than the AI product itself.

#84Now the founders faced an unusual choice.

#85On one side was a consumer product with real users and an enormous volume of messages.

#86On the other was code they were giving away.

#87They chose the free code.

#88That was the consequential turn. Hugging Face stopped centering the company on building a better AI companion.

#89Instead, it would help other people build AI.

#90For Delangue, the direction had an echo.

#91At UniShared, openness and collaboration had failed to produce the company he wanted.

#92Now those same values were finding traction among AI developers.

#93And BERT created more requests. So Hugging Face supported more models.

#94The library expanded into Transformers, a toolkit that let developers use a growing range of pretrained AI models without rebuilding everything from scratch.

#95More models created another problem: people needed somewhere to store and share them.

#96Hugging Face built the Hub. Then those models changed over time, which meant versions had to be managed.

#97Git—the version-control system widely used to track changes in software—became part of the infrastructure.

#98Companies wanted similar workflows without exposing proprietary work.

#99Hugging Face added private repositories and enterprise features.

#100There had been no grand blueprint on day one that said: build the GitHub of AI.

#101The company kept solving the next problem created by the people already using what it had built.

#102Eventually, Hugging Face contained far more AI made by everyone else than AI made by Hugging Face itself.

#103The company pushed the collaborative idea further.

#104In 2021, the BigScience project brought together more than 1,000 researchers from over 70 countries and more than 250 institutions.

#105They were attempting something normally associated with the research labs of very large technology companies:

#106build a huge language model. But do it through an open international collaboration.

#107In 2022, that effort produced BLOOM, a 176-billion-parameter language model. Parameters are the adjustable values a model learns during training; at the time, a model of that scale was the kind of project associated with enormous computing resources.

#108Only a few years earlier, Hugging Face had been helping developers use a model created by Google.

#109Now it was organizing a global community to build a giant model together.

#110The company's role was becoming harder to describe as a single AI product.

#111It was becoming a place where AI got made.

#112Then came an investment round that made the change impossible to miss.

#113In 2023, Hugging Face raised $235 million at a $4.5 billion valuation.

#114Look at the investors: Google. Amazon. Nvidia.

#115AMD. Intel. Qualcomm. IBM. Salesforce. Some of these companies competed directly with one another.

#116Nvidia and AMD fought in AI chips. Google and Amazon fought in cloud computing.

#117Many of them had their own models, developer tools, or AI strategies.

#118Yet they could all participate in Hugging Face.

#119That was possible partly because Hugging Face wasn't asking the ecosystem to bet on one proprietary model as the winner.

#120Different models could live there. Different hardware companies could connect to it.

#121Delangue described the position Hugging Face wanted to occupy with a geographic metaphor:

#122Switzerland. A neutral place. Its neutrality wasn't simply an absence of power.

#123It was part of what allowed competing companies to share the same space.

#124Then one of those companies tried to get much closer.

#125Nvidia later proposed investing another $500 million in Hugging Face, according to reporting, at a valuation of roughly $7 billion.

#126Hugging Face declined. By then, the company was profitable and still had a substantial portion of the money it had previously raised.

#127It didn't urgently need the cash. More striking was the concern reported by the Financial Times: Hugging Face wanted to remain independent and avoid giving one investor too much influence.

#128That mattered for a company whose usefulness depended in part on being a place where rivals could coexist.

#129If one competitor became too powerful inside that place, would everyone else still see it the same way?

#130Then, months later, the situation changed dramatically.

#131In August 2026, a report said Nvidia had agreed to acquire Hugging Face for about $12.9 billion.

#132Not another investment. The whole company.

#133The companies had not publicly announced the transaction as completed.

#134But the report opened a much larger question.

#135What would Nvidia be buying?

#136By 2026, Hugging Face hosted millions of models, alongside huge numbers of datasets and AI applications.

#137Most models attracted very little attention.

#138A tiny fraction accounted for the overwhelming majority of downloads.

#139That wasn't necessarily a contradiction.

#140It was what a platform could look like. Hugging Face didn't have to know in advance which model would matter.

#141Someone uploads one. People try it. A model catches on.

#142Other developers build with it. More tools and services appear around it.

#143And as more people run AI, they need compute—the processing power required to train or operate those models.

#144Nvidia dominates that part of the market.

#145That makes Hugging Face potentially more valuable to Nvidia than its software revenue alone would suggest.

#146It sits along the route through which developers discover models, obtain them, and put them to work.

#147But the reported acquisition also contains a paradox.

#148If part of Hugging Face's value comes from being a common ground for many competing companies, Nvidia ownership could make some of those competitors question how neutral that ground really is.

#149In other words, owning Hugging Face could put pressure on one of the qualities that made Hugging Face worth owning.

#150That tension remains unresolved. Meanwhile, Hugging Face is already extending the same pattern somewhere new.

#151On August 27, 2026, around the same time as the Nvidia acquisition report, Hugging Face's Pollen Robotics introduced a small robot.

#152Microduck. Roughly 25 centimeters tall. $399.

#153A tiny two-legged machine with a name to match.

#154But the interesting part wasn't just the robot.

#155Users could train new behaviors in a simulation—a virtual environment where the robot could learn without first performing every attempt on the physical machine.

#156Then they could run those behaviors on the real robot and share what they had trained.

#157The familiar Hugging Face loop was moving into robotics.

#158Someone builds something. Someone shares it.

#159Someone else takes it and changes it. A decade earlier, Hugging Face had wanted to make one AI product people would love.

#160Now it was trying to make it easier for other people to create and share not only AI models, but behaviors that AI could perform in the physical world.

#161Microduck reportedly took more than $2.6 million in orders in its first 24 hours.

#162No one knows how far that experiment will go.

#163And Hugging Face didn't map out this destination when it began.

#164Again and again, some of its most important turns came from noticing what other people were doing with things the company had put into their hands.

#165For Clément Delangue, there is a smaller scene that captures just how far that journey traveled.

#166When Delangue was 17, the machines he imported from China had to go somewhere.

#167They ended up at his father's garden-equipment shop.

#168Years later, Delangue was showing his father generative AI—software that could create new content from a simple request.

#169He told him to ask for anything. The system would make an image.

#170His father chose something familiar. A John Deere tractor.

#171Then he changed one detail. Put it on the moon.

#172The AI generated the picture. The teenager who had once filled his father's shop with real machines bought online was now sitting beside him, using AI to produce an image of a machine that had never been there at all.

#173His company had changed just as radically.

#174Hugging Face did not become the company with the one AI everyone had to use.

#175It became the place where millions of people could bring their own, share them, take someone else's, and build again.

#176Ten years after three founders hurriedly named their chatbot startup after a 🤗 emoji, that place had become difficult for even the world's largest AI companies to ignore.

#177At 17, Clément Delangue was importing ATVs from China and selling them on eBay. Later, he passed up a path that could have taken him into Google and kept starting companies instead. Why did he keep choosing the uncertain route?

#178Before Hugging Face, Moodstocks showed Delangue that a tiny AI startup could take on problems that looked reserved for giants. He walked away from an extension at eBay to spend time there—and Moodstocks was later acquired by Google. What did he see in that small team?

#179UniShared was already built around people creating something together and sharing it. It failed. Years later, strikingly similar values helped Hugging Face become a global platform. What changed?

#180Thomas Wolf left physics, studied law, and spent roughly five years working in patents. Then deep learning pulled him back—and one small suggestion he made on a Friday became a turning point for Hugging Face. Why did he return?

#181In 2016, Betaworks put roughly $200,000 behind an AI-friend startup with no obvious business model. Even the founders had only just settled on the name Hugging Face. What did Betaworks think it was betting on?

#182One of Hugging Face's early investors wasn't a technologist at all. It was NBA star Kevin Durant. Why did a basketball player put money into a tiny AI chatbot company before its famous pivot?

#183BERT belonged to Google, not Hugging Face. Yet making Google's model easier for another group of developers to use helped change Hugging Face's entire direction. What was it about BERT that made that possible?

#184More than 1,000 researchers across over 70 countries joined BigScience to attempt something usually done inside giant corporate labs. How do you get that many people to build one AI together?

#185BLOOM had 176 billion parameters and emerged from an enormous open collaboration. Why would researchers rebuild, in public, something giant technology companies were already racing to create behind their own walls?

#186Nvidia reportedly offered to invest another $500 million in Hugging Face and was turned down. Months later came a report that it had agreed to buy the entire company for $12.9 billion. What changed?

#187Pollen Robotics had been building robots long before Hugging Face acquired it. Why would a company known for sharing AI models decide it needed a robotics company of its own?