#e4mXplains:  e4mXplains: Nvidia just bought the AI sandbox: Why its $13 bn Hugging Face deal matters

I have a confession to make. Until today, I did not know what CUDA stood for.

I cover technology, artificial intelligence, and the companies making increasingly ludicrous amounts of money from both, but ask me to explain the intricacies of GPU architecture and things deteriorate fairly quickly.

I do, however, understand gold rushes.

When prospectors poured into California after the discovery of gold in 1848, fortunes weren't made only by the people who found it. There was excellent money to be made selling picks, shovels, clothing, transport and everything else required by thousands of people convinced there was gold in them thar hills.

Nvidia has spent the AI boom in an enviable modern version of that position.

It makes the specialised processors, or GPUs, that have become essential to training and running artificial intelligence systems. Everyone wants to find AI gold. Nvidia sells them the shovels.

On September 3, the world's dominant AI chipmaker announced that it would acquire Hugging Face for $12.93 billion. The deal is expected to close in the first half of 2027, subject to regulatory approvals, a fairly consequential caveat given the size of the transaction and the position Nvidia already occupies in the AI market.

And Nvidia isn't buying another sand shovel manufacturer. It is buying part of the sandbox.

First, what on earth is Hugging Face?

For people outside the technology industry, Hugging Face is easy to underestimate because describing it as an "AI developer platform" explains approximately nothing.

Think of it instead as part repository, part workshop and part marketplace for artificial intelligence.

Developers can go there to find AI models other people have created, access datasets, modify and test models, share their own work and deploy applications without having to build a foundational model from scratch.

It is enormous.

According to Nvidia, more than 18 million developers, researchers and creators use Hugging Face, which hosts more than three million models, 500,000 datasets and one million applications. More than 200,000 companies use the platform to discover, evaluate, customise and deploy AI.

Nvidia will pay around $11.9 billion to Hugging Face shareholders and provide an equity-based retention programme worth up to another $1 billion for employees joining the company.

The scale of the acquisition matters because Nvidia already occupies an extraordinarily powerful position somewhere else in the AI economy.

Which brings us, unfortunately, back to CUDA.

The shovel isn't just the chip

CUDA stands for Compute Unified Device Architecture.

There. Now we all know.

The name is considerably more complicated than the economic idea behind it.

A GPU provides computing power. CUDA is Nvidia's software platform that allows developers to put that computing power to work on tasks including artificial intelligence.

If the GPU is the factory, CUDA is closer to the assembly-line system: the tools, libraries and processes through which engineers make the machinery useful. (Yes, this is a comparison to the Ford factories and modern manufacturing, circa early 1900s. Who said this tech writer's references were dated?)

Nvidia has been developing CUDA for nearly two decades. Developers have learned to use it, software has been built around it, and companies have designed AI infrastructure around Nvidia hardware.

That matters because competition stops being purely about who can manufacture the cheapest or fastest chip.

Moving to another manufacturer's processor can also mean changing software, retraining engineers, testing workloads again and rebuilding parts of the infrastructure surrounding it.

Or, translated from tech into economics: switching costs money.

That is one of Nvidia's great advantages. It doesn't merely sell a remarkably successful product. An enormous ecosystem has grown around using that product.

There is one problem with selling shovels to trillion-dollar companies

Eventually, they start making their own.

Google already has its Tensor Processing Units, or TPUs. Meta, Microsoft and OpenAI are among the companies developing custom AI chips partly to reduce their dependence on Nvidia's processors. Reuters identified precisely this tension in reporting the acquisition: Hugging Face could help Nvidia strengthen its position as some of its largest customers develop alternatives to its hardware.

Nvidia cannot realistically stop Google or Meta from spending billions of dollars trying to build their way out of that dependence.

Hugging Face potentially gives it influence at a very different point in the life of an AI company: the beginning. And as could also be expressed in other situations, you never forget your first time.

A five-person startup experimenting with its first AI product does not have Google's engineering resources. It has neither the money nor, presumably, the masochistic inclination to optimise everything for several competing hardware ecosystems.

It has an incentive to choose whatever works.

If one combination of models, software and computing infrastructure is slightly easier to access, better documented, better optimised and faster to deploy, choosing it is perfectly rational.

Do that repeatedly and the five-person startup may eventually become a 5,000-person company whose engineers know the same tools, whose software has been designed around them and whose infrastructure has accumulated around the original choice.

Nobody had to lock the door. It just became progressively more inconvenient to leave.

That is what makes Hugging Face interesting.

Nvidia doesn't necessarily need its corner of the sandbox to be the only one available. It merely needs playing in its corner to be a little smoother.

Which is why Jensen Huang's promise matters

Nvidia appears acutely aware of the concern.

"Hugging Face will remain an open platform for the entire AI ecosystem," CEO Jensen Huang said in announcing the acquisition via blog post.

Developers will remain free to choose their models, frameworks, clouds, inference providers and computing platforms, he said, adding explicitly that Nvidia compute will not be required to build on or deploy through Hugging Face.

The company's regulatory filing goes further, committing Nvidia to continue supporting other silicon vendors. Hugging Face's existing backers include Nvidia competitors AMD and Intel.

And there is a compelling commercial reason for Nvidia to keep that promise.

Hugging Face derives much of its value from being broadly useful. Turn it into an Nvidia showroom and you risk driving away precisely the developers, models and competing ecosystems that made it worth $12.93 billion in the first place.

But open does not necessarily mean equal.

AMD hardware could remain perfectly available while Nvidia offers marginally easier deployment, tighter integrations, better optimisation or simply the familiarity of an ecosystem developers already know.

Indeed, Nvidia is already the largest contributor of open models and data to Hugging Face, according to Huang, with more than 500 models and 250 datasets on the platform. Its Hugging Face presence includes optimised versions of prominent open model architectures including Meta's Llama and work developed with Mistral AI.

The interesting question raised by the acquisition, then, isn't whether Nvidia will prevent developers from choosing someone else.

It is whether owning an important place where developers experiment can help ensure Nvidia remains the path of least resistance when those experiments become businesses.

And yes, this matters to advertising

Because advertising is rapidly becoming one of those businesses.

Agencies, martech companies and brands are building generative creative systems, AI agents, personalisation engines, automated media workflows and proprietary marketing applications.

Open-weight models make it possible for businesses to take an existing model and customise or deploy it for their own purposes rather than sending every task through a closed third-party AI service. Hugging Face has become one of the central places where developers find and work with those models.

For an agency building a campaign tool or a marketer deploying an AI agent, the processor eventually running underneath it may feel several floors below the actual work.

Economically, it isn't.

Every additional layer of AI inserted into advertising creates demand underneath it: for models, inference, cloud capacity and ultimately compute.

The advertising industry is currently racing to automate itself. The companies controlling the infrastructure underneath that automation consequently matter, even when marketers never see their logos.

Nvidia has spent the AI boom in perhaps the best business imaginable: selling the machinery everybody else needs to participate.

Its richest customers are now wealthy enough to start making some of that machinery themselves.

Buying Hugging Face won't stop them.

But it puts Nvidia inside one of the places where the next generation of AI developers will start experimenting long before they become rich enough to contemplate making their own chips.

During a gold rush, selling shovels is an excellent business.

Owning part of the sandbox where the next generation learns how to dig may be better.