Don't expect things to go differently this time around. Nvidia wants control over the software stack. Acquiring HF fits in perfectly. The play is long term.
They even share many of their pre-training and even post-training datasets for Nemotron on HuggingFace; for example: https://huggingface.co/datasets/nvidia/Nemotron-Post-Trainin...
Which other lab shares this?
Yes, there is no question NVIDIA wants to lock you into CUDA and their hardware. But also, they’ve consistently demonstrated the most openness when it comes to model training, datasets, and research; even before the LLM era (e.g. StyleGAN).
There’s also modelscope.cn (china’s huggingface) which is worth checking out. I would not be surprised if one day, we have to use China VPNs to download open weight models.
Of course they are. They're commoditizing their complement.
I want to own the hardware, not play around in an nvidia fiefdom full of nvidia rules.
Nvidia probably likes that ASML is a monopolist (it's called a monopsony). The price is high and Nvidia can't scale as hard as they want to (more general: capital intensive market). This monopsony makes sure that other chip companies can't rapidly scale up and try to beat Nvidia. That there only ever was one other GPU firm (I'm prehistoric; once there were more) and they bungled it on software, is pretty sweet for Nvidia.
On the side of their customers. It would be best for Nvidia if there is free competition for the outputs generated from there GPUs. This maximizes consumer surplus and thus demand. Maximum demand for tokens, is maximal demand feeded in their monopoly. If there is a monopoly right from you, demand is curtailed, and your value is limited.
And that exactly is why you see interest from token generators for chips. Bridge that moat and gain a larger value surplus. Both NVidia and, say, China actively undercutting the token-supplier value chain is quite interesting to watch. It's like the Opium wars with us as somewhat happy customers.
In this same vein, why isn't ASML raising thousands of billions for building their own (subsidiary) chip foundries, while raising prices and starving the market (a little) for their machines.
You're looking this purely through an economic lens, while in reality geopolitical factors play a huge role in what ASML can and cannot do. The US government would likely take an extremely dim view of any new external competitor popping up for their chip foundry industry (especially with all the new US plants being built or planned) and would lean heavily on their vassal/ally the Netherlands to prevent this. Unlike with China, the US has more leverage over the Netherlands[1]
The US security state and US tech giants are joined at the hip, as they have been since the beginning of Silicon Valley[1], right through the Snowden revelations through to the present day[2].
[1] https://nltimes.nl/2026/08/20/us-preparing-force-netherlands...
[2] https://www.brennancenter.org/our-work/research-reports/sect...
> Of course they are. They're commoditizing their complement
Then why don't they sell consumer GPUs with tons of memory. They clearly segment the market into consumer versus server/business.
There are a few problems though, primarily, a GPU with lots of VRAM and very high bandwidth is inherently very expensive (on top of which there is also the CUDA premium); AI use cases are better served by SoCs with lower (but still high) bandwidth and more RAM.
If you don't want to use CUDA, they expose the PTX bindings to write your own CUDA alternative too: https://docs.nvidia.com/cuda/parallel-thread-execution/index...
Also, when running OpenCL, NVIDIA hardware disables multiple DMA engines, and allows only one memory transfer at a time to prevent OpenCL running as fast as CUDA.
Did NVIDIA finally allow open source drivers to access all parts and features of the card to allow feature parity? Last time I checked they were considering a plan for planning a solution to that.
I don't think bits like the GSP firmware will ever be open-sourced, but the opportunity to write better OpenCL drivers has always existed. Some of Nvidia's other proprietary driver backends (eg. GBM, Vulkan) are also decently neglected, but mostly out of disuse rather than malice. I don't think any of these things mean you don't own the hardware.
Having said that, I'll try compiling a OpenCL 3.0 program in the cluster, so I can report whether NVIDIA runs this software, and if yes, how well.
Likewise SYSCL although built on top of OpenCL 3.0 primitives, is mostly Intel, which also owns CodePlay, the company that delivered the first working SYSCL compute experience, again neither AMD nor Intel (until it bought CodePlay).
Also if we were just discussing labs, Ai2 opens ~everything with dramatically less resources than Nvidia.
Nvidia’s history with linux shows the opposite. And as a user running models on a linux/AMD stack, this information does not fill me with hope.
They might be right w.r.t. openness about LLM at the moment, but w.r.t. general software openness they are definitely the opposite of open.
It’s a little like the trend of calling developers “engineers”. There’s no actual engineering in the traditional sense but I’m sure developers think it sounds cool to call themselves that.
My point was just that I just find it amusing that people call themselves engineers when the code they produce is so far removed from the level of rigour one would expect in literally any other engineering industry.
I say this as someone who also has family and friends who are actual engineers, if they built bridges and buildings to the same standards that many developers write code, then people would die.
This isn’t meant as a criticism of developers, by the way. Just an observation at the vast differences in the domains and thus the tolerance for errors in the process.
Likewise for labs. I’ve worked in AI startups and the science departments are not something one would think of when you say “laboratory”. I get why the term is used, but it’s still amusing.
I know language isn’t static. It’s something that evolves, like how a “computer” used to refer to a person rather than a thing, but that doesn’t stop me from being amused. But maybe the real issue here is I take myself less seriously than others so I have that capacity to be amused by the titles I’ve held?
Just like everything AI is a "model". It's actually not, but it sounds cool/sciency.
It sounds cool because people in our age are almost obsessed with scientistic performance.
At least AI labs are actually doing experiments.
No shit Sherlock. Name one company that shared their code to make you NOT to consume their stuff?
A few years ago everyone said _Open_AI is the the most open labs. How did that turn out? Lots of coy, deceiving actions till the whole company was turned into whatever rent seeking amoral borg adjacent shell of it's former past it is now.
Dude, where's the src for GPU drivers and the firmware blobs?
Sorry, you can't say it is one of the most open labs without qualifying a proper response to the question above.
Nvidia is also one of the most closed hardware developers around. Two things can be true.
But it is a good example of people talking past one another and not communicating well, I would think.
Sure, the models are open weight, but porting the code needed to run them on non-Nvidia hardware is not trivial.
They supported OpenCL when Khronos floated the idea of a GPGPU standard to manufacturers, but OEMs didn't want to design scalable hardware or sponsor the software.
Evidently we should, because Linus has been more positive about Nvidia in the last 2 years [0]. I've been using the open driver for years now, for both gaming and CUDA.
[0] https://binarymusings.org/posts/talks/linus-on-ai-linux-in-k...
> This is actually one of the benefits brought by AI; it has made Nvidia a good participant in the Linux kernel space. [...] Now, when Linux is so important for AI clouds, Nvidia suddenly cares very much about Linux.
Now Nvidia is also premium sponsor of the Linux Foundation.
Nvidia is a big company. They are good about some things and bad about others.
I think they really do like open weights because they make some of the best hardware for training, and the more open weights models there are, the more people are training and fine-tuning them, mostly on Nvidia hardware.
I feel like Nvidia is one of the better choices for buying Huggingface. Not perfect, but definitely far from the worst.
The worry is that Nvidia is trying to control the way you run those models. Trying to bake CUDA assumptions into model design, and pushing the software ecosystem to be as Nvidia first as they can.
1. Oracle
2. A16Z
3. GameStop
2. Microsoft
3. Google
Are probably the worst of the realistic options for acquiring HuggingFace
My impression I've developed in the years of working there is that Nvidia's relationship with opensource is... not intentional. They kinda suck at it because they genuinely don't know how to do it more than they want to make money out of it.
Here's an anecdotal "success story" which is also an illustration to how things might not work out well otherwise.
So, I was on the team that deals with server infrastructure. One day we get a new "feature" which was supposed to allow Slurm (the workload manager, a kind of software used to run "jobs", including eg. model training) to be deployed with distributed MySQL as a backend. The feature is all obviously written by a single developer with an enormous amount of "help" from AI. I was tasked with testing it.
Trying to figure out what it does... I realized that the "distributed" part of the feature was to be achieved by integrating with Oracle's MySQL by means of using MySQLShell (another proprietary Oracle's product). Until that point, by default, we integrated with MariaDB. Not only was it using Oracle's proprietary tool, the tool, actually, didn't support the "distributed" part of the "solution". It was pitched as the "first step on the way there".
So, I was able to push back on it, mentioning Galera, arguing that the "solution" doesn't solve the problem and will require from customers to change databases (even if they are mostly compatible... they never quite 100% compatible). And the misfeature was rolled back.
I made an effort to investigate how did we even get there, and turned out that whoever authored the "solution" had an experience of working with Oracle products, but never really tried the open-source ones. So, he didn't do a research. He just used what he knew.
Unfortunately, this is a rare win, where the evidence of disadvantages of using proprietary solution was huge and enough to turn the tide. But often it doesn't face any resistance because nobody is even aware of the problem.
I do think that they still might be a bit reticent to release fully open-source drivers, though, only because of the extent to which hardware design could be inferred from the driver source. OTOH, AI itself is making disassembling and reverse engineering binary code easier and easier, so there might not be much of a point to withholding the source in the near future.
Modular on the other hand creates the Mojo compiler gets criticised for not open sourcing it immediately and now once they do, no-one cares anymore.
Huggingface was not just a target for open source, but as a force to have open weight models run better on Nvidia against the rest.
this is the crux - if nvidia makes it so that open weights end up running better on nvidia hardware than competitor's, then it's going to prevent hardware innovation and competitiveness in the entire sector.
It's like as tho General Motors buys out oil refinery to make gas for all, but the gas somehow runs smoother in GM cars.
NVidia is delivering commodity hardware to the hyperscalers, and would eventually get commodity margins (when hyperscalers make their models work on their own hardware).
Amazing article on relevant economics of squeezing vendors - actually about antitrust ad-models but:
The answer is that [advertising based] companies are an ideal test case for an increasingly common business meta-model, where companies try to create a consumer surplus at one end in order to maximize their negotiating leverage for capturing the producer surplus everywhere else in their supply chain.
Perhaps not relevant to huggingface, sorry.
Couldnt they just take those models anyway?
Whereas mojo is a general purpose language, and we're absolutely spoiled for choice on modern languages with open source compilers.
I'm not saying it's fair or right, I still think mojo is neat, but isn't exactly comparing apples to apples.
Nvidia on the other hand has not and the best they have done is a bunch of closed-source blobs which they do more closed source releases than the rest.
Mojo is open source and targets all GPU architectures for their compiler regardless of the vendor and nvcc targets their own (and both that and CUDA are closed source).
So this is directly an apples to apples comparison.
AMD is another story. Their binary closed source blobs were even worse than nvidias for many years. So much so people used terrible performance open source tries just to avoid the headache. That they finally slopped together an open source version after they'd lost the GPU race isn't exactly noble. But as an open source supporter in general, I commend the effort still.
possibly because it took qualcomm buying them to make that happen.
Mojo was partially open source before Qualcomm bought them, and they were going to do open source it anyway.
Was NVCC or CUDA ever open source since the lifetime of its development?
uh huh.
> Was NVCC or CUDA ever open source since the lifetime of its development?
you ever ask Nvidia why? i did.
Exactly. You're one of those that don't care.
Mojo's standard library was open sourced a year before the acquisition.
> you ever ask Nvidia why? i did.
The correct answer to my question is "No". CUDA or NVCC was never open sourced during its lifetime.
Since you "claim" to have asked them and know the true reason why is not open source, just say what Nvidia told you verbatim right here.
Intel and AMD have no problem open sourcing their GPU drivers for Linux, but Nvidia has a problem with that.
We can only speculate that "open sourcing" it reveals their GPU trade secrets and their intellectual property for others to copy.
If it was going to "eventually" be made open source, why not do it from the start? Why not do it a year in?
From what I was told, your speculation is incorrect. I can't say anything more than that, sorry.
I don't know why you're so aggressive in tone towards me, chill out.
> If it was going to "eventually" be made open source, why not do it from the start? Why not do it a year in?
That isn't how open sourcing software works. It gets released when it is ready or gradually.
But of course as I predicted that you would ask that question. Why didn't Linus release Linux immediately? Why did he wait 29 days to do it instead of the day when he announced it? Why didn't Rust release 4 years early as soon as is was eventually open sourced?
> From what I was told, your speculation is incorrect. I can't say anything more than that, sorry.
Who told you? Nvidia directly?
The correct answer is that you do NOT know the reason yourself because there was NO discussion at all.
Just say you don't know.
I'm going to skip past all of your asinine aggressive response and just answer one thing again since you missed it the first time...
> you ever ask Nvidia why? i did.
So that was not even open sourced “from the start” either.
You understand that you just undermined your own point? It was released when it was ready, rather than from the start.
The entire point is one company gave a gradual open source release over several months to a year, verses another company that has open sourced close to nothing related around CUDA or nvcc for decades.
nvcc remains completely closed source despite it even using Lattner’s LLVM (which he open sourced) as a dependency! (shocker)
> I'm going to skip past…
Exactly. Skip it because you seem to be struggling to answer those basic questions.
Nothing was missed as I already responded to your last sentence by asking you to post exactly what Nvidia just told you verbatim since you claimed to know the true reason. You failed to answer.
If that is too difficult then a paraphrasing is far better than: “I can’t for just reasons.”
Or just be honest and say: “I do not know the reason”.
Nope. It was released when they realized we were building something that was going to be competitive with theirs. It was viewed as better to join forces than do it separately because we were already helping them design the servlet spec.
> you seem to be struggling to answer those basic questions
I have yet to see a question, just a bunch of ad hominem. You really have some weird personal beef with me. Seek help and support for your anger issues.
The rest of the details to this are irrelevant. We already know that it was not open-sourced 'from the start' and as you just admitted, it took a week to get it ready and transition it from closed to open source.
So even with your example, it fails your own qualification for a project to "eventually" be made open source.
> I have yet to see a question, just a bunch of ad hominem. You really have some weird personal beef with me. Seek help and support for your anger issues.
I don't think you even know what an 'ad hominem" is or even realize that you just did one in your own comment as a distraction for not answering a basic question. But I'll remind you about what that question was since you missed it:
"Who told you? Nvidia directly?"
These are very basic question(s) that you are really struggling to answer. In this entire thread, you could not even give the real reason why Nvidia never would open source their CUDA or nvcc software when you "claimed" to know.
So again, Repeat exactly what Nvidia told you right here, given you claim to know the "true" reason or just say "I don't know".
The only struggle is your understanding. Maybe english isn't your first language, but that is me saying I asked Nvidia.
No you did not. You are referring to a question that I already answered for you [0]. Your answer was incorrect.
>> The correct answer to my question is "No". CUDA or NVCC was never open sourced during its lifetime.
> The only struggle is your understanding. Maybe english isn't your first language, but that is me saying I asked Nvidia.
Says the person that does not know what an "ad hominem" was nor did you even realize you just did one in a previous reply. The question I am basically asking you is to elaborate on what the true reason was since you claimed to have asked them (Nvidia).
If Nvidia gave you their response, It should be very easy to just say it right here word-for-word or paraphrase what they told you then. Many replies later still, you are struggling to even do that.
Such as basic question to answer and yet you continue pretending to know the reason when you clearly don't know.
The only thing I can see them being able to get away with is increasingly bending the hugging face python API and any other features of that sort they develop to NVIDIA only. I personally don't use that and don't see a reason to and I am not sure how many people do use the hugging face python library.
They need to take a machete to all the cross coupling they’ve metastasized.
Their strategy for AI is pretty clear and they bank on on-premises OSS models for the busines, with their really cool open-source software: https://www.youtube.com/watch?v=tmcn1-jFLWY
This is a really good watch and is a glimpse into what the actual future shapes up to be considering the current situation in where the OSS Chinese models successfully compete with proprietary US ones.
Nvidia wants people and companies to go choose a free open model, run that model on Nvidia hardware. And because Nvidia can't fully control what hardware an AI model can run, Apple Silicon and AMD hardware users will benefit as well.
Open weights correspond to "binary available" for software. Nobody would call that "open source".
(Even ignoring the licensing which "open source" normally entails.)
Though we agree that the term isn't typically used that way in the LLM context, right?
They're trying to mix up the competitive landscape(that doesn't impact their bottom line, and I don't think opensource is eating their lunch), so I don't think this is fake, at least that's my initial take.
100%
> They want
[citation needed]
Or at least, $13b to stay at the head of the race (or keep the race running) must be worth it to someone's desk.
There's only $50b in datacenter buildout nationally (Source: Gemini, 2026).
So it is a bit of a puzzling choice for what amounts to a pile of software, in my opinion. but I don't know shit.
But more seriously, this is my first time seeing that as well, and I'm not sure I like it. Citing an LLM is a little like citing Wikipedia to me, you cite the primary source the LLM is quoting directly, not the secondary source.
It may as well say "(Socrates, probably)"
In fact, has HF ever wished to make money? They probably pay AWS more infra cost than they earn. The exit was planned all along.
Your reference lacks authority, veracity, and reproducibility.