For a total of $12.9303 billion, Hugging Face is becoming part of Nvidia. According to Nvidia CEO Jensen Huang, the AI model platform will remain open to all kinds of free-to-download LLMs. Nevertheless, the acquisition raises questions about both that promise and Nvidia’s long-term status as the king of AI hardware.
Rumors had been circulating for weeks, with even the amount involved an open secret. Hugging Face (HF), the largest platform for downloading open AI models, had been an independent private company until today. As a linchpin of the LLM software chain, HF has become a household name; coincidentally, the same is true for Nvidia’s GPUs, which have been the primary platform for training and running LLMs on a daily basis since the rise of GenAI. It’s important for Nvidia to keep it that way. Let’s explore why.
Nvidia has always been a platform company
Some coverage of this acquisition will undoubtedly view HF as a somewhat surprising choice for an acquisition; after all, Nvidia is an AI chip company that not long ago acquired its potential competitor, Groq. However, the “moat,” the strength behind what amounts to a near-monopoly on data center GPUs, is often identified as the software layer surrounding the GPUs. Summarized as CUDA, which can be described as both software and a driver, the familiarity and maturity of Nvidia’s software are just as crucial to its success as the hardware. However, it goes beyond that: Nvidia’s platform has always been broader than just GPUs; consider the Mellanox acquisition for networking or Run:ai for orchestration.
Now, an additional dimension is being added. Nvidia is ensuring it keeps control of a far larger portion of the AI value chain than before, with the ability to keep a watchful eye on the development of open models to improve its own hardware, drivers, networking and overall ecosystem. Even just knowing firsthand which models are deployed by whom will tell the company where it should be pushing its products long-term. For now, nothing changes for HF users, but long-term this may make Nvidia’s reach more invasive.
With Hugging Face now part of Nvidia, the result is a somewhat incongruous combination of openness and a closed ecosystem. HF offers AI models of all shapes and sizes, with the generous licenses accompanying the available LLMs as their common thread. These have included Nvidia models for years, which naturally run optimally within Nvidia’s “walled garden.” There’s no reason to assume Nvidia will upset the applecart too much early on, but it will want to recoup its investment somehow.
Hugging Face’s adoption figures speak for themselves: by its own account, it has 18 million users, more than 3 million models, 500,000 datasets, and 1 million applications, with 200,000 companies using the platform for their AI needs.
The acquisition price is significantly higher than the last publicly reported valuation. During the Series D round in 2023, Hugging Face was valued at $4.5 billion. Given a reported annual revenue of approximately $150 million, the price amounts to about 86 times revenue.
No mandatory Nvidia hardware
Nvidia CEO Huang emphasizes that nothing will change regarding the platform’s neutrality. “Nvidia computing power will not be required to build or deploy via Hugging Face,” he writes. The platform will continue to support multi-cloud and multi-accelerator environments and will also retain its well-known 🤗 brand name.
That promise should help calm some concerns. Several observers had previously questioned whether a neutral hub could remain neutral under an owner that also provides the dominant GPU infrastructure on which many of those open models run.
In any case, the connection between the two parties has long existed. Nvidia has published more than 500 models and 250 open datasets on the platform, making it, according to Huang, the largest contributor of open models and data. In his announcement, Huang refers to an open letter about open weights that he recently co-signed.
Nvidia is feeling the pressure elsewhere
You wouldn’t think so given that it’s the world’s most valuable company, but Nvidia is also under pressure in the long term. For the first time, there are signs that the GPU manufacturer’s customers are truly looking elsewhere. It’s not AMD, Google Cloud, Amazon, or any of the AI model developers themselves that are driving this trend, though they all play a small part. LLM workloads vary enormously, from training to inference, and Nvidia has had little to fear in terms of performance so far. But change is on the horizon.
Specifically, this involves concepts such as Cerebras, Google’s Frozen v2, or OpenAI’s Jalapeño chip in collaboration with Broadcom: processors that, while impressive on paper, are nowhere near as generically powerful as a GPU from Nvidia or AMD, but do excel at highly specialized workloads. If AI models ever standardize or plateau in their performance improvements, it will become cost-effective to essentially etch the architecture of these models directly onto hardware. GPUs don’t do that: they’re filled with layers of abstraction between the model and the hardware and are widely applicable to virtually any LLM. A chip design that runs only a specific AI model optimally can operate many times faster and more efficiently for that use case than a GPU.
It is therefore important for Nvidia that AI models continue to innovate and evolve, including open-source options that drive down the cost of LLM usage and make it more accessible. As long as that process remains active and adoption grows, Nvidia can benefit from sustained demand and the corresponding prices. It won’t be able to just buy up its chip competition like it effectively did with the Groq acqui-hire, meaning a wider ecosystem shift keeps it relevant for all AI adopters.
Read also: Nvidia doubles revenue thanks to sustained AI boom