There Were Never Just "Talks." There Was a Pattern.
There Were Never Just “Talks.” There Was a Pattern.
An open letter to Clément Delangue and the Hugging Face community
August 27, 2026
Clem, you have said the quiet part out loud for years.
Concentration of power is the biggest risk in AI.
You said Hugging Face was “a neutral platform, or the Switzerland” of the field.
You structured the 2023 round so Salesforce, Google, Amazon, Nvidia, Intel, AMD, Qualcomm and IBM each held a slice and none of them held the knife.
Late last year you turned down a 7 billion, because a single dominant investor could sway decisions. The Financial Times reported it. You did not need the money. You said so.
That was not branding. That was the product. Developers put models and datasets on the Hub because they believed the catalog was not an Nvidia storefront, not a Google storefront, not an OpenAI storefront.
This week the same company you refused a minority stake in is reported to be buying the whole building for about 13 billion have not produced a signed agreement and could still fall apart. CNBC says acquisition has been “part of ongoing and recent talks.” Neither of you has confirmed. That distinction matters. Calling a leaked process a closed acquisition is sloppy. Pretending the process is a rumor you “heard about” is sloppier.
You have spent 2026 talking about being careful, about long-term responsibility to people who parked their work on the Hub, about not optimizing for an exit. A sale to the firm that already sells the GPUs under most of those models is not a continuation of that sentence. It is the end of it.
And it did not come out of nowhere.
What actually happened to open-source AI
The history is not mysterious. It is a short list of outcomes.
1. The labs that tried to stay fully open at frontier scale ran out of money or people.
Stability AI raised 1 billion valuation after Stable Diffusion. By late 2023 it was burning far faster than it earned. Reports put cash near $4 million. AWS bills slipped. Emad Mostaque left in March 2024. Layoffs. A recap. A new CEO. Alumni founded Black Forest Labs. The models survived. The original institution did not.
Allen Institute for AI built the rare thing that was actually open — weights, data, code. In 2026 it pulled back from competing on general open models. Ali Farhadi, Hanna Hajishirzi, Ranjay Krishna, Kyle Lo, Luca Soldaini and others left for Microsoft’s Superintelligence team. The Paul Allen estate’s funding process changed. Ai2 still exists. The “American open lab that trains general models in the open” version of Ai2 does not.
LAION assembled the datasets the image generation boom ran on. After Stanford found suspected CSAM links in LAION-5B, the sets came down. A cleaned release followed. The organization is still a thin German nonprofit. OpenAssistant, the volunteer ChatGPT replica LAION helped run, concluded. Weights remain. The project does not.
EleutherAI still exists. It stopped being the place that trains the next GPT-J and became a smaller interpretability lab living on grants. That is survival. It is not scale.
2. The companies that built the tools everyone actually uses got bought.
MosaicML, which released MPT and sold efficient training, went to Databricks for $1.3 billion in 2023. The team became Mosaic AI. DBRX came later, under a data company, not a commons.
Then the shopping list got boring and complete:
- Neural Magic (vLLM, inference) → Red Hat / IBM
- Run:ai (GPU orchestration) → Nvidia, then promised open-source
- OpenPipe (open agent RL toolkit) → CoreWeave
- Flowise (open visual agent builder) → Workday
- Continue (open coding assistant) → Cursor
- Astral (uv, Ruff — the Python toolchain the field runs on) → OpenAI
- Promptfoo (open LLM eval/security) → OpenAI
- OpenClaw (open agent) → OpenAI acquihire, project waved toward a foundation
- Weights & Biases → CoreWeave
Every buyer said the repo would stay open. Sometimes the license stays. The people who understood the internals start reporting to a product org whose job is not the commons. Simon Willison’s point on Astral was the honest one: the risk is not a license revoke. It is neglect.
3. Hugging Face was never a side project in this story. It was the square.
You hosted BLOOM. You hosted the models Stability, EleutherAI, Ai2, Mistral, DeepSeek, Qwen and Meta dropped into the world. You bought Pollen Robotics and talked about open robots. You told Axios the 2023 cap table was neutrality.
Nvidia was already on that cap table. So were Google and Amazon. That was the point: many landlords, no owner.
A 7 billion would have broken the design without buying the building. You said no.
A $12.9 billion purchase buys the building.
Why Nvidia wants the Hub, in one sentence
Nvidia does not need Hugging Face’s $150 million of revenue. At the reported price that is roughly 86 times sales. Nvidia needs the layer where developers choose which model to download, which checkpoint to fine-tune, which inference stack to trust, and which hardware that stack is tuned for.
That is control of distribution, not control of every weight file on earth. It is enough.
Open weights already run overwhelmingly on Nvidia silicon. Owning the catalog lets the company decide what is featured, what is default, what gets a one-click deploy onto NIM or DGX, what gets a CUDA kernel first, and which competing accelerator story has to fight the homepage. Jensen can sign letters defending open weights — he did, and Hugging Face signed too — and still want the checkout counter. Those two things are not in conflict. Open models that are easy to run on Nvidia boards are a demand engine for GPUs. A Hub Nvidia does not own can feature a stack that makes AMD, Google TPUs, or a Chinese cluster look like the rational choice. A Hub Nvidia owns will not treat that as a feature.
You already described the logic in March, on CNBC: Nvidia supports open source because it needs everyone to build, not a few players. That is true. It is also why the same firm wants the directory those builders walk through. “Everyone can build” on our chips, in our cloud leftovers, through our listing. That is vertical integration. It is not a secret plan. It is the 10-K.
The other bidders, if the reporting is right, understood the same asset. Salesforce was in the 2023 round and was said to have kicked the tires. Microsoft met and, per Business Insider, is not in active talks. Stripe just paid a reported $7.5 billion for OpenRouter — another distribution choke point. The market is pricing the catalog, not the chatbot you started in 2016.
“Talks” is the wrong word
You have used the language of caution: we hear things, we think long-term, we have a responsibility to the community. That language fit 2023. It fit the rejected 13 billion.
In 2024 you told Bloomberg you were getting about ten inbound acquisition pitches a week from other AI founders. You were the buyer. Argilla was $10 million. Pollen was a statement about open robotics. That was Hugging Face as consolidator of the open stack.
The arrow flipped. When the reports started this week, the question was not whether someone wanted the Hub. It was which incumbent was allowed to own the Switzerland.
Here is the part that is not fair to pin only on Nvidia:
Databricks bought MosaicML. OpenAI is buying the Python toolchain and the eval tools. CoreWeave bought the experiment tracker and an agent trainer. Workday bought the visual agent builder. Philanthropy blinked at Ai2. Stability nearly died on cloud invoices. Chinese labs now account for a large share of Hub downloads — you have said you were uneasy about that yourself.
Nvidia is not the only force that enclosed the last four years of open work. It is the force that can enclose the index.
That is why your old sentence still applies, and why it now points at you.
Concentration of power is the biggest risk in AI.
The Hub is one of the last pieces that was not already inside a model lab or a chip vendor.
If the reported deal closes, that sentence is no longer a warning you issue from the outside. It is a description of the company you sold.
Sources
- The Information — Nvidia agrees to buy Hugging Face for $12.9 billion
- CNBC — Nvidia reportedly agrees to buy Hugging Face
- Financial Times — Hugging Face’s rejected $500 million Nvidia check
- Business Insider — talks above $13 billion, no signed agreement
- Forbes — Stability AI’s cash crisis and recap
- GeekWire — Ai2’s pullback from general open models
- TechCrunch — Databricks acquires MosaicML
Neither Nvidia nor Hugging Face has confirmed a signed deal as of this writing. This letter treats the reported process as a process — not as a closed acquisition, and not as a rumor.