It’s a core part of the open model infrastructure. Their hardware business benefits greatly if this works really well and open models proliferate to every corner of the economy, with everyone buying GPUs with a lower capacity factor/duty cycle than the centralized ones. It also helps people across the world to collaborate on new uses for this, effectively running a massive search in the search space of what’s possible with these things, for every year (see the endless variety of quants and finetunes to fit in different cards, and bias their abilities on different tasks). More use cases pop up, some of them are killer apps, buying Nvidia’s hardware becomes more valuable at whatever price point, more people do it at higher prices, Nvidia stock chart goes up.

If Anthropic, OpenAI buy them, this org gets very different priorities, owning them ensures that doesn’t happen.

Hugging Face's internal storage system is more sophisticated now after they acquired XetHub. They have a clever way to chunk and content dedupe the data that makes it a lot more efficient to version control large files.

Xet's tech is very interesting and I greatly enjoyed their blog posts and other technical writeups[0]

But in the era of frontier models and hardware designed for fast model loading (eg: LLM on a chip) I can't imagine that this is worth so much. There's *a lot* of competition to improve Git LFS.

So while I agree with you that this plays a role I also have to believe that it's to understand model usage by competition and thus stifle it.

[0] - https://huggingface.co/docs/hub/en/xet/overview

git lfs just dumps your bin on s3, and is abandonoware since ms bought github, since github historically hosts on aws s3 and improving it would immediately benefit amazon. gh, the maintainers of it, is the only forge that does not allow clean-up unreachable artifact from the server.

Surely you could now vibe code such system for under $1,000.

I highly doubt it, there's about a million things that only show up at scale that aren't lessons the bots can easily digest from Medium posts and bypass

You are assuming the operator of vibe coding tools is not experienced in the domain.

There's more to than just the website. You have network bandwidth costs, security costs, infrastructure costs.

You can't just spin up a VPS and throw a few brain jar's in; the network bandwidth must be a fortune. Hosting over 700GB+ for many of the same models has storage cost too.

That is fair point, but still you can build such system for under a grand and then pimp yourself out to VCs.

> Their brand alone surely is not worth this much.

HF has become the front runner in open model hosting. They have a lot of trust from the wider community. That's gotta count for something.

Lawyers (and bean counters) call it 'goodwill'. One word.

Accounting goodwill might be one term, but it has multiple components (trust, reputation, market share, mindshare, track record, relationships, online and offline community presence, etc. etc.)

I only meant the artful term is 'goodwill' and not 'good will'.

Oh

It's a narrative to pump the stock. Nothing more. Nvidia could have build out this infrastructure themselves for less than 1% of the cost, but it wouldn't have a name like Huggingface does.

It's a customer funnel. They've been attempting a similar product for years and nobody cares.

They are buying the user base.

Sometimes I wonder if anybody on this website has ever actually done professional software engineering at scale. What a ridiculous take.

It isn't an awful 'surface level take'.

I'm not that familiar either, and so had same question.

I thought they just hosted other peoples models. So why the value of 13B?

Buying a company is very rarely about buying the tech, at least principally. It is true, as others have pointed out, that Nvidia could have "rolled their own" Hugging Face far more cheaply and probably in about the same amount of time as it will take to execute the deal end to end. That should give you a clue that the tech isn't really the target here

More than anything, they are buying a position within the ecosystem. Many of us could get a version of an app deployed in short order that implements most core HF functionality, but I am sorry to say that this is very far from the "hard part" of building a business, and this was the case even before coding agents were a thing. It's no small thing to become the first tool pretty much everyone reaches for in a particular niche, and therefore the incumbent in that niche

And there's lots of inertia keeping companies and developers on these incumbents like HF and GitHub -- see also Microsoft's play here dumping billions into a company that is basically peanuts and probably nowhere near as profitable as Microsoft normally would like. But they bought it at a time when they were trying to make Microsoft synonymous with open source, and it was a great play to that end. Likewise Nvidia wants to be synonymous with AI. Even if HF is a loss leader, the positioning is probably worth it

To a lesser degree, Nvidia probably also might like to reinvigorate its flagging cloud compute business and this could be a natural way to do that. At least TechCrunch thinks this, but I, an overconfident layperson, feel that the positioning is where most of the value is for Nvidia, and consolidating that will act as a force multiplier for that + anything else they may want to do with this space

13B is a lot for an online platform that hosts weights. What's the special secret sauce that makes this so valuable? The fact that it's the go-to place for public weights?

No secret sauce is needed to be valuable

Did you read the model hack that occured on their platform? I wouldn't exactly call them "advanced" if their security practices are anything to go by.

Code is hard but more likely it is a brand thing which is much harder to do. Nvidia for years now has tried to get a public ai model hub of sorts working. This is obviously them throwing in the towel.

It's such an obvious customer funnel they have been trying for years to make.

1. Customer googles "some model" 2. Hugging face is often top result 3. Thinks it's great 4. Buys

The customers are already on hugging face.

Acting like code is the blocker here would be very wrong. Nvidia 200% has the technical knowhow to make a hugging face platform and they'd likely make it better.

I think they should worth much much more than this

Call me a conspiracy theorist, but I think it might have something to do with OpenAI's model hacking HF thing, especially considering OpenAI is one of Nvidia's most important customers.

>So what are they paying for?

They are paying for an insurance policy. Elon Musk and SpaceXAI must be incredibly interested in buying the company who was the victim of a mass hacking campaign by OpenAI. Any lawsuit, criminal investigation, or slowdown of OpenAI's model training would put OpenAI's massive data center build out, which Nvidia just backstopped, at risk.

Any DA can investigate it, they don't need Musk or some other big tech dipshit to tell them otherwise. Real crimes were committed and no charges are being pressed.

DAU