Enzo Ruffin

PortFolio 2026

Nvidia is buying Hugging Face: AI becomes a software battle

Nvidia has agreed to buy Hugging Face for $12.93 billion. What the deal says about software lock-in and the architecture choices behind an AI product.

5 min read
Enzo Ruffin
AI

On September 3, 2026, Nvidia announced an agreement to buy Hugging Face, valuing the company at $12.93 billion. The filing with the SEC breaks the amount down: about $11.9 billion for stockholders and up to $1 billion in retention for employees. The deal has not closed. Closing is expected in the first half of 2027, subject to regulatory approvals. For a developer shipping products with AI in them, it confirms one thing: a model has to stay a replaceable dependency, because lock-in now happens in software, open models and the tools teams adopt day to day.

Nvidia moves up into software

Nvidia has dominated the market for accelerators used to train and run models for years. Its GPUs have become baseline infrastructure for labs, large platforms and cloud providers. That position also creates a dependency: a significant share of its growth rests on the spending of a handful of major customers, several of which now design their own accelerators to cut costs. Buying Hugging Face is an answer to that shift, moving one layer up, towards models, developer tooling, deployment and the services that turn expensive infrastructure into something usable. After selling what runs AI, Nvidia wants to sell what people build with.

Hugging Face, the GitHub of AI

Founded in 2016 by Clément Delangue, Julien Chaumond and Thomas Wolf, Hugging Face became the crossroads of the open source AI ecosystem. It works much like GitHub: you discover, publish and pull models, datasets and demos, then wire them into your own applications. The platform hosts language, vision and audio models alongside the libraries used to run them, transformers first among them, present in a large share of Python AI projects. It also sells services to companies deploying their own models. Its value lies in its position: it is the place where developers pick the technology they are going to use.

What Nvidia is really buying: an audience

According to the figures released with the announcement, more than 18 million developers, researchers and creators use Hugging Face, and more than 200,000 companies rely on the platform. That audience matters more than the platform's revenue, because architecture decisions are made in a notebook, a README or a command line, the moment a developer tries a model and finds it good enough for the job. Whoever owns discovery owns the defaults, and defaults are what end up in production.

Open models and defending the CUDA ecosystem

Open source and open-weight models let a company download, adapt and deploy a model in house rather than depending entirely on a proprietary API. The approach gives more control over data, costs and hosting, which is exactly what companies embedding AI into their own systems are after. The need for compute, meanwhile, grows: more models adapted and deployed means more training, more fine-tuning and more inference. Nvidia therefore has every reason to keep that movement going, as long as it keeps running through CUDA, its main advantage for close to twenty years now.

The value chain gets longer

Until now the market could be summed up in three links: chips, models, applications. This deal sketches a longer one: chips, infrastructure, models, tooling, applications, end customers. Above all it points at where the economic value sits. Having a strong model is not enough: you have to run it at an acceptable cost, connect it to your data, and handle users, permissions, uptime and regressions. That work is ordinary software engineering, and it is precisely the link Hugging Face has settled on.

Lock-in now runs through software

A hardware vendor can be replaced: it is a purchasing decision, painful but finite. A software ecosystem is a different story. Once models, libraries, formats and workflows are embedded in a team's processes, leaving becomes a rewrite. That is the whole point of the deal for Nvidia, which brings developers even closer to its tools and infrastructure. It is also its limit: if the community starts feeling that tools presented as open are quietly tilting towards a single vendor, part of it will look elsewhere. Model weights are still files, and a repository can be mirrored.

The openness commitment

This is the point that will decide the real value of the acquisition. In its SEC filing, Nvidia commits to keeping the platform open, in line with Hugging Face's existing practices: models can still be uploaded and downloaded, and other silicon vendors remain supported. Jensen Huang added that Nvidia hardware would not be required to build or deploy on the platform. That promise determines what the asset is worth: Hugging Face built its position on the trust of a community that publishes its work there for free, far more than on technology that is hard to copy. If the neutrality holds, Nvidia inherits a wider ecosystem without having to close it to competitors. If it does not, the ecosystem will be rebuilt somewhere else, at a cost of $12.93 billion.

What it changes when you actually ship product

On the projects I ship, the consequence is immediate: a model has to be treated as a replaceable dependency, never as a foundation. In practice that means a single interface on the back end, a service exposing generate, summarize, classify, with adapters behind it, one per provider. Prompts, output schemas and evaluation sets are versioned with the code, exactly like tests. Inference cost is measured per feature rather than globally, otherwise nobody can tell which idea blew up the monthly bill. At that price, moving from a proprietary API to a self-hosted open model, or back, comes down to an implementation change.

For a startup, the value is in the integration

For a team building a SaaS product, an ERP or an internal tool, value comes from the way the model enters the product. An interface designed for one specific trade, integration with the tools already in place, clean and structured data, automated workflows, infrastructure you control: none of that can be downloaded. It is what I see on the products I work on, mobile apps and ERPs included: AI is a component there, almost never the product. You do not need to train your own model to build a serious technology company.

What is left to watch

Closing is expected in the first half of 2027, once regulatory approvals are in. Until then the openness commitment is a promise, to be judged on the two points named in the SEC filing: the freedom to upload and download models, and support for competing chips. For teams building with these tools, the question to settle now is which infrastructure they are willing to rely on for the next five years.

5 min read
Enzo Ruffin