ATAILA Newsroom · Budapest · 2026-08-28
Nvidia is buying the place the world downloads open models from. We are not happy about it.
The Information reported on 27 August that Nvidia has agreed to buy Hugging Face for $12.9 billion; hwsw.hu carried it the same day. Nothing is signed as far as anyone outside the two companies knows, and neither has commented. ATAILA is built on open source AI, so we say it plainly: we are not happy about this.
What happened
Hugging Face is where open AI models live. Meta, Mistral, Alibaba, DeepSeek, Google — and Nvidia itself — publish their open weights there, and it maintains the toolchain the whole field imports: transformers, safetensors, and the GGUF builds that llama.cpp and Ollama consume. The reported numbers:
Two caveats. Business Insider, which broke the story, reports no signed agreement yet — the talks could still collapse. And this is a second approach: Nvidia reportedly offered $500 million at a $7 billion valuation last year and was turned down.
What we think
One company owning both the shelf and the toolchain is already concentration. That company also being the one that sells the GPUs is a different order of question. It makes nobody a villain — it means “open” stops being a property of the ecosystem and becomes a policy decision, made by an owner, reviewable every quarter.
To be fair: Nvidia publishes open weights itself, and hosting the world’s model files on roughly $150 million of revenue is a hard business — a funded owner is one honest answer to that. We hope for that version. We are not going to assume it.
Where we stand
ATAILA advocates for open source AI. Concretely:
- Downloading models is free.
- Loading, using and training them is free — on your own hardware, offline if you want.
- Independent repositories must seed open models so everyone can reach them — not only the people who bought the right accelerator.
This is our business, not a slogan. ATAILA’s primary purpose is to teach customers how to use open source AI — which model fits which job, on hardware they control. And the ATAILA Cloud Platform is itself built on open source from the hypervisor up: operating systems, databases, storage, identity, CI/CD, monitoring, serving. That is why we can say what we say about lock-in without crossing our fingers.
Three things we hope for
Each is checkable. In a year we will say here whether it held.
Hope 1 · Plurality
That other repositories like Hugging Face exist — and stay free.
Model weights are files. Nothing forces the world to have exactly one shelf, and this week showed how fragile that arrangement is. Independent repositories — including publicly funded European ones — should seed and mirror open weights the way research institutions mirror scientific data, so everyone can reach them.
Hope 2 · Neutrality
That Nvidia does not tweak the models to its own flavour — and keeps them free.
Open weights must stay accelerator-neutral: tested on AMD ROCm, Intel, Apple Silicon and plain CPUs, not merely permitted there. The danger is not a ban on anything. It is a slow drift where the CUDA path is the tested one and every other path is where the numbers come out wrong.
Hope 3 · Open ecosystem
That Nvidia does not kill the open source AI software ecosystem — no push into its own stack beyond drivers and CUDA.
We love vLLM and Ollama. We run both in production, alongside llama.cpp-family tooling. Nvidia has earned the driver and CUDA layer; above that line the ecosystem must stay open, and projects like these must keep optimising for whatever hardware they choose, at their own pace.
What this changes for our customers: nothing
The models we serve are pulled once onto our own storage, inside our own datacentre. Inference never calls a hub. And we deliberately run more than one serving engine — vLLM for throughput, Ollama where it fits better — so no single runtime, and no single repository, can strand us.
We are not happy about this deal, and we say so plainly. But nothing we run for a customer depends on it going well — the weights are on our hardware, served by engines we could swap. Open source is not a licence file. It is the ability to keep working when someone else changes their mind.
One question worth checking this week
If your model source disappeared tomorrow, would your system still start? If you have to check — that is the gap worth closing, whoever ends up owning the shelf.
Source: Az Nvidia kezébe kerülhet a Hugging Face — hwsw.hu, 2026-08-27
← Back to the Newsroom Press inquiries: contact us