Most people who run a language model (LLM) on their own laptop today are running on the same software under the hood, which few of them realize. That software is called llama.cpp. It was written by Georgi Gerganov, and in March 2023 it was an evening hack meant to get the Llama model running in plain C and C++ with no dependencies, so even on a machine without a graphics card.1 That hack gave rise to the GGUF format, which became the de facto way of distributing quantized models, and to an inference engine that Ollama, LM Studio, KoboldCpp and a number of other desktop tools now stand on.
This year the software changed owners twice. In February 2026, Hugging Face announced that the team around ggml.ai, meaning Gerganov and his people, was moving over to it.2 The terms sounded good: the project stays under the MIT license, and the team gets full-time positions and technical autonomy. Then on September 3, 2026, Jensen Huang announced that NVIDIA is buying all of Hugging Face for $12,930,300,000.3 Of that, $11.9 billion goes to shareholders, and another billion is in stock earmarked for retaining employees. The deal is expected to close in the first half of 2027, so for now it is an agreement, not a done deal.4
Not everyone sees why NVIDIA is paying roughly $13 billion for a company “whose website you just download open models from.” That is a culpable underestimation of the position Hugging Face has built.
First of all, running open-weight models locally on personal computers today rests on llama.cpp and the GGUF format, and it flows through one main distribution channel, the Hugging Face Hub. Could you download from somewhere else? Maybe. The 300 most downloaded GGUF repositories on the Hub have about 83 million downloads over the last thirty days, roughly three times the MLX format for Apple Silicon. That is hard to compete with.5 Second: llama.cpp has moved into the business of its own users. One-command installation, a model catalog, a web interface: Ollama used to offer all of that, and now the engine can do it by itself. And third: all of this is now being bought by NVIDIA, a company that makes its money on the very cards that llama.cpp can work around. Cha-ching. A factual note: the deal has not yet been closed and finalized.
An engine that stopped being just an engine
Local AI is usually described as the floors of a house. On the ground floor is the engine, the program that actually computes the model. On the first floor is the application you work with, such as Ollama or LM Studio, which takes care of downloading models, running them, and the chat window you type into. On the top floor are the paid services, the cloud, and enterprise features. Each floor used to have its own owner, and each of them earned money in a different place. But the tenants started moving: the engine built its own application, and the applications are moving to the cloud. The market is reshuffling, and everyone is reacting to it.
llama.cpp has long since stopped being a bare inference core. Its server component, llama-server, has an OpenAI-compatible API and ships with its own built-in web interface.6 The model can also call tools through it. It can ask on its own for a web search or for a piece of code to be run, and it can use several of them at once.7 And in December 2025 came router mode, in which llama-server dynamically loads, unloads, and switches models without a restart.8 Those were the first big shifts.
By summer 2026 the project had gone even further. It has its own consumer website, llama.app, one-line installation in the terminal, unified llama serve and llama cli commands, and above all its own model catalog.9 On top of that comes integration of the local coding agent Pi, whose connecting plugin lives right in the Hugging Face organization on GitHub. One-command installation, a model catalog, an agent. That is exactly the offering Ollama came out with three years ago.
So the engine is moving closer to users, while the applications built on top of it are moving toward the cloud. You can see it clearly with Ollama. It started as a convenient wrapper around llama.cpp, and today it has its own model catalog, its own management of how models are run, and its own cloud. You get a model running on your laptop with one command, and when it stops being enough, you switch to a bigger model in Ollama’s cloud.10 An application that was supposed to be local thus becomes an entry gate to the cloud.
Why a chip maker is buying a model catalog
This is a good place to go through the numbers, because they usually explain the motive. By available estimates, Hugging Face has annual revenue of around $150 million. NVIDIA earns roughly $96 billion per quarter. From the buyer’s point of view, the acquired company is an accounting rounding error. If this were about revenue, a $12.9 billion purchase would make no sense. It makes sense as an investment in distribution.
But it does make sense as an investment in distribution. Hugging Face is the place where more than three million model repositories “live,” used by eighteen million developers and two hundred thousand companies.3 Anyone who wants to download an open model today goes there. And NVIDIA has a reason to make sure that ecosystem stays alive: big closed labs such as OpenAI, Google, Amazon, and Anthropic are building their own chips and thereby reducing their dependence on its hardware. A flourishing world of open models gives customers an alternative to the closed labs, and that in turn opens the market to NVIDIA’s cards and chips.11
In the announcement, Huang promises that nothing will change. Hugging Face will remain an open platform, and developers will pick models, frameworks, clouds, and compute platforms as they see fit. He literally writes that working with Hugging Face will not require computing on NVIDIA hardware.3 That is a remarkable promise. Mainly because it occurred to the head of NVIDIA to say it out loud. Why guarantee things that nobody thought of questioning?
An engine for hardware the buyer doesn’t sell
And here I see a clash, because llama.cpp became famous for running models where there is no NVIDIA graphics card. It supports Apple Silicon via Metal, has a backend for AMD via HIP, for Intel via SYCL, a universal Vulkan backend, and it runs on a bare CPU or on a Raspberry Pi. That independence from the iron is precisely its main value.
Gerganov himself responded publicly and matter-of-factly. He writes that NVIDIA engineers have been actively contributing to the project for over a year, working with the community, and providing hardware for development and testing. He adds that hardware independence is one of the project’s founding principles, so development and support of all backends will continue as before, driven by the community.12 That is not an empty phrase, because the collaboration is documented. This June, Gerganov described joint work by llama.cpp maintainers and NVIDIA engineers on tensor parallelism, which resulted in speedups on RTX cards and a foundation for hardware-independent parallelism in ggml.
The February promise that the project would stay open source and free, however, was made by Hugging Face on its own behalf. The new owner has not repeated it so far, because formally there is no new owner yet. If NVIDIA explicitly confirms the governance of llama.cpp and the GGUF format after the deal closes, that will be good news for the local ecosystem. If that sentence is never uttered, that will also be a kind of answer…
Escape from the cloud to local
The promise of local AI sounds appealing: run the model at home and the data never leaves your machine. Ollama uses this argument as a key to regulated industries, where data must not be sent to someone else’s cloud.10 But it’s not that easy.
Data stays on the machine only under certain conditions. A local model must really be used, the application must not send telemetry, RAG data must not leak out, and external tools and search must be local. And above all: running locally doesn’t remove platform dependence, it just moves it elsewhere. The user escapes the API provider and stays dependent on the model catalog, the runtime, CUDA, a specific GGUF quantization, and the model’s license.
You can probably see where I’m going with this. Once the deal closes, the catalog, the engine, and the graphics card will all belong to one company. The escape from the cloud leads to the chip maker, who, after all, never let you out of its all-embracing AI arms.
What you’re actually allowed to run when you’re in Europe
For the Czech reader, an important point is that “open weight does not mean open source.” Model licenses differ, and a European user can legally run fewer of them than an American one. The Acceptable Use Policy of Llama 4 explicitly does not grant rights to the multimodal models to individuals domiciled in the EU or to companies headquartered in the EU.13 The same restriction applies to the Llama 3.2 vision models. Text models pass, multimodal ones don’t. Meta has not given an official reason; commentators link it to uncertainty around European regulation, mainly the AI Act and GDPR.
This gives the debate about sovereign AI a concrete outline. A European company that wants to run locally and multimodally runs into a licensing boundary drawn around the continent. The catalog of models available to it narrows before it even gets to technical limitations. And starting next year, that catalog will be operated by an American chip maker.
Open code, concentrated influence
Does this mean that one single company will control how we run open models? Not necessarily; there will still be alternative ways, with varying degrees of usability, quality, and general awareness.
It is not an entirely new situation, because before February 2026 the concentration of influence was even higher. llama.cpp had a broad base of contributors, but decision-making authority ultimately rested with Gerganov in the role of BDFL. The February move transferred the concentration of power from a person to an institution. Moreover, Hugging Face was no newcomer to the project; it already had key llama.cpp contributors on its team.2 The naive picture, sometimes presented, of “a corporation swallowing the commons” thus gets a crack.
The project today has a publicly documented three-tier ladder. Contributors open pull requests, collaborators are invited based on work done, and maintainers can merge; there are roughly three dozen maintainers.1 Among them are people from Hugging Face, independent developers, and Bartowski, one of the best-known community quantizers. Developers on Hacker News describe a regime where a review sometimes arrives in five minutes and a pull request is merged within an hour, because Gerganov delegates and is not a “gatekeeper.”14 Decision-making power is spread across more people than before.
The code also stays under the MIT license, so anyone can fork it. But the technical possibility of a fork and its real feasibility are two different things. A fork would have to keep pace with new model architectures, with the GGUF format, with hardware backends, and with the whole sprawling ecosystem of servers, APIs, and web UI. It would need people who understand optimizations, access to pre-release models, and test hardware. That is not a small thing, and it’s definitely not cheap.
So in my view the contradiction lies elsewhere. The code is more forkable than ever, maintainers are multiplying, and yet decision-making about what the default choice is converges on one company. The question of power in open source is “who decides that a change enters upstream” and “whose models and quantizations will be the default.”
The concentration also doesn’t have just one center. Ollama runs its own model registry, a second distribution point with its own defaults and with a pinned, modified version of llama.cpp. Here too someone decides; the problem of gatekeepers is as old in IT as IT itself…
The starting grid gets shuffled right at the start…
We’ve described the market just as the operating systems themselves, and their makers, are entering it. At WWDC 2025, Apple introduced the Foundation Models framework, which gives developers direct access to an on-device model in three lines of Swift, with no API key, no network, and no per-token fee.15 In June 2026 it took that further: AFM 3 brings a twenty-billion-parameter on-device model and image input.16 Microsoft is going in the same direction via Windows ML and the runtime for Copilot+ machines. Of course, having an on-device LLM makes perfect sense.
The difference is in who supplies the model. Today the user downloads it as a GGUF and runs it through Ollama or llama.cpp. In the model that Apple and Microsoft are pushing, the operating system delivers it directly as its own API. If this channel wins out, everything will change, because “ordinary users” will simply use these LLMs the way they learned to use Internet Explorer and other bundled tools. The battle over the catalog of open models would then be a battle over a shrinking part of the market.
What follows from this
After September 3, the answer to the question in the title looks like this: from running models locally, the one who sold you the hardware profits. The software is free. The engine is free, the runtime is free, the weights are there to download, and yet someone has just sent thirteen billion dollars into that layer. Not to charge for it, but to keep it open and alive, because open models keep customers on its cards.
In a certain sense, that is good news. Nobody has a reason to put a paywall in front of llama.cpp, because the money is rather in the iron. At the same time it means that the independence of local AI now depends on the business interest of one company. As long as that interest coincides with the community’s, it works. Once they diverge, the MIT license will remain, but the toll booth, along with the gatehouse, will move elsewhere.
A side note, but an eloquent final thread: Nat Friedman and Daniel Gross, who funded the launch of ggml.ai, vanished in mid-2025 inside Meta, which bought out their fund and hired both.17 The financiers ended up in Big Tech, and the engine, after two moves, ended up at a maker of graphics cards. In three years, an evening hack on a MacBook became an asset in a billion-dollar transaction, and it is still the same code under the same license.
What to watch next is easy to say. Whether NVIDIA, after the deal closes, repeats the February promise of openness through Huang’s mouth. Whether commits to the AMD, Metal, and CPU backends continue at the same pace. And whether a regulator gets involved, because a formal antitrust review would give the ecosystem years, not months, to build alternatives.
Sources
A note on methodology: the terms of the acquisition are based on NVIDIA’s announcement and on the regulatory filing as described in the media. The deal has not yet been closed; completion is expected in the first half of 2027. Hugging Face’s revenue and NVIDIA’s quarterly revenue are estimates from trade media, not audited figures. Claims about the product surface of llama.cpp and about Llama’s licensing restriction in the EU are backed by primary sources. The predominance of the GGUF format is based on my own measurement via the public Hugging Face Hub API. The data reflect the state as of Sept 13, 2026.
Footnotes
-
llama.cpp repository (first release March 2023, MIT license, documented ladder contributor - collaborator - maintainer, CODEOWNERS file, list of maintainers). Status as of Sept 13, 2026: over 123k stars, 21.6k forks, over 10,300 commits: https://github.com/ggml-org/llama.cpp ↩ ↩2
-
Hugging Face announcement on incorporating the ggml and llama.cpp team (Feb 20, 2026; full-time positions, technical autonomy, key contributors already earlier on the HF team): https://huggingface.co/blog/ggml-joins-hf ↩ ↩2
-
Jensen Huang, “NVIDIA to Acquire Hugging Face” (NVIDIA blog, Sept 3, 2026) - price $12,930,300,000, 18 million developers, 3 million models, 200k companies, commitment to openness, and the sentence that computing on NVIDIA hardware will not be required to work with Hugging Face: https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/ ↩ ↩2 ↩3
-
CNN Business (deal structure: $11.9 billion to shareholders and $1 billion in stock for employee retention, closing expected in the first half of 2027 according to the regulatory filing): https://www.cnn.com/2026/09/03/tech/nvidia-hugging-face-ai-acquisition ↩
-
My own measurement via the public Hugging Face Hub API, Sept 13, 2026. Sum of thirty-day downloads of the 300 most downloaded repositories by format: GGUF approx. 83 million, MLX approx. 27 million. Quantization accounts over 30 days: unsloth approx. 35 million, bartowski approx. 5.7 million. This covers the top of the distribution, not the sum of all repositories. MLX targets only Apple Silicon, and the numbers may include automated builds. ↩
-
Open WebUI documentation on llama.cpp (llama-server has a built-in web interface and an OpenAI-compatible API): https://docs.openwebui.com/alternatives/llama-cpp/ ↩
-
llama.cpp documentation, function calling: https://github.com/ggml-org/llama.cpp/blob/master/docs/function-calling.md ↩
-
llama.cpp documentation, server README (router mode): https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md ↩
-
Official project website llama.app (one-command installation, model catalog, Pi agent integration). Status as of Sept 13, 2026: https://llama.app ↩
-
Ollama Series B announcement ($88 million total, 8.9 million developers, local running with scaling to the cloud). Business Wire via Las Vegas Sun: https://lasvegassun.com/news/2026/jul/09/ollama-raises-65m-series-b-funding-to-grow-its-ope/ ↩ ↩2
-
TechCrunch (NVIDIA’s motivation: protecting its chip position, return to the cloud, context of closed labs building their own silicon): https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/ ↩
-
Georgi Gerganov on X (reaction to the acquisition: NVIDIA has contributed to the project for over a year, provided hardware for testing, hardware independence remains a founding principle, support of all backends continues, driven by the community): https://x.com/ggerganov/status/2095897173376618881 ↩
-
Llama 4 Acceptable Use Policy (multimodal models are not licensed to entities headquartered in the EU): https://www.llama.com/llama4/use-policy/ ↩
-
Hacker News discussion on the future of llama.cpp after the acquisition (description of review speed and delegation of authority): https://news.ycombinator.com/item?id=49567357 ↩
-
Apple, Foundation Models framework (WWDC 2025, on-device model via Swift API). Apple Machine Learning Research: https://machinelearning.apple.com/research/introducing-apple-foundation-models ↩
-
Apple AFM 3 (WWDC June 8, 2026, 20B on-device model with image input). 9to5Mac: https://9to5mac.com/2026/06/11/apples-new-foundation-models-explained-on-device-ai-cloud-ai-and-everything-in-between/ ↩
-
Meta hires Nat Friedman and Daniel Gross and buys out their NFDG fund (mid-2025). PYMNTS: https://www.pymnts.com/artificial-intelligence-2/2025/report-meta-aims-to-hire-ai-investors-nat-friedman-and-daniel-gross/ ↩