Transcript 0:00 In 2012, Linus Torvalds called NVIDIA the single worst company the Linux community had dealt with. Then he looked into the camera, raised his middle finger, and addressed NVIDIA directly. 0:12 And NVIDIA has been the single worst company we've ever dealt with. So NVIDIA, [censored] you. 0:19 14 years later, in 2026, NVIDIA maintains open Linux kernel modules, publishes major AI libraries, and releases model weights, data sets, and training recipes. 0:31 On its own open source page, it counts more than 1,400 projects released under OSI-approved license on GitHub, over 650 open models, and more than 250 open data sets. And recently, this story got a new dimension. 0:49 NVIDIA has agreed to buy Hugging Face for almost $13 billion. If the deal closes, NVIDIA would own the main place where much of the open model world distributes its work. 1:02 I'm a little bit surprised to hear this news because Hugging Face always tremendously valued their independence, but I guess the burden to make money and fight for compute got too overwhelming. 1:13 One disclosure before I go further. NVIDIA has sponsored Turing Post before, during GTC in March. This episode is not sponsored. NVIDIA did not review it. 1:23 But I really wanted to do this episode and this little investigation because I got tired of two types of comments that I, for some reason, keep receiving. 1:33 One type says, "Jensen Huang made every correct decision for three decades." The other says, "NVIDIA contributed almost nothing to open source." [upbeat electronic music] Span. Attention span. 1:46 Welcome to Attention Span by Turing Post. I'm Ksenia. You asked me about all of that, and here you go. It's time to see what is true and what is false about NVIDIA and open source. 1:58 To understand why NVIDIA opens anything at all, you have to go back to the moment it almost died for closing the wrong thing. NVIDIA's first graphics chip, NV1, used curves and quadrilaterals. 2:12 Microsoft and game developers were converging on triangles. 2:16 NVIDIA had chosen an architecture that did not match the emerging software standard and was building a second chip for Sega around the same time, and it was failing magnificently, to the point that Jensen Huang had to go to Sega executive Shoichiro Irimajiri and admit the project was wrong. 2:36 He asked Sega to release NVIDIA from the contract while continuing its support. What Irimajiri arranged was a $5 million equity investment converted from money Sega already owed under the contract. 2:50 Jensen has since said it was all the money NVIDIA had and that it bought the company six months. NVIDIA used that time to abandon its architecture and build Riva 128 around the standard developers had chosen. 3:05 Irimajiri later said that the confession reduced his confidence in the chip but increased his confidence in Jensen. 3:12 NVIDIA survived through an honest admission, Sega's generosity, and a replacement chip that worked when the company could afford just one attempt. The lesson was painful. 3:24 Developers decide which platform becomes a standard, and this is the first half of the pattern. But there is the second half in which NVIDIA is right, but nobody cares. 3:35 A decade after that near-death experience, NVIDIA made a strong comeback and encountered a different problem. 3:42 At Stanford, graduate student Ian Buck helped build Brook GPU, an open source project for using graphics processors in ordinary calculations. Before Brook, a GPU only accepted graphics commands. 3:55 Scientists who wanted to use one GPU had to hide their math inside a picture, store the numbers as pixel colors, write the calculation as a shader, and read the results back out of the rendered image. 4:09 Nothing about the work was graphical. That was the only language the hardware spoke. Buck later said that moving an application to a GPU effectively required a doctorate in computer graphics. 4:20 NVIDIA hired him in 2004, and Buck began developing CUDA. It gave developers a direct way to use a GPU beyond graphics. Jensen then made the expensive decision to put that capability into GeForce gaming cards. 4:35 Here is the trap that it created. 4:37 Developers could find CUDA hardware almost everywhere, but every chip now carried features gaming customers did not value, while the applications that could justify the investment barely existed. 4:49 Jensen has a name for markets like that. 4:51 He calls them zero billion dollar markets, and he means it as a compliment because, yes, there are no customers, but there's another side of it, because there's also no competitors, and CUDA was one of those markets. 5:04 GeForce sales was strong in fiscal 2008, so CUDA alone did not cause NVIDIA's later downturn. 5:10 It was a commercial failure because scattered scientific users did not form a market large enough to cover what NVIDIA was spending, so NVIDIA went looking for one. 5:21 It searched across seismic processing, medical imaging, and even frozen pizza simulation. In Taiwan, Jensen visited a homemade GPU supercomputer cooled with ordinary fans. 5:33 So he knew and the company knew the technology worked, but the large business was missing. CUDA downloads peaked in 2009 and then declined for three years, according to people interviewed by The New Yorker. 5:47 Jensen kept expecting the market the following year, then the following one, then the year after that, while wondering whether NVIDIA had arrived too early. Investors were furious. 5:57 NVIDIA even opened university research centers and held its first GPU technology conference in 2009. Yep, that's right. 6:04 The famous GTC that now is treated as the Super Bowl of AI was created to promote a product that was failing. The market NVIDIA could not find was already forming in labs on hardware bought at retail. 6:18 In 2009, Stanford researchers showed that a gaming GPU could cut a neural network training job that had taken weeks down to about a day. 6:28 Geoffrey Hinton later recalled recommending NVIDIA cards to around 1,000 researchers, then asking NVIDIA for one free GPU. Jensen said no. 6:37 NVIDIA had built the technology deep learning needed, and still the company itself did not recognize the market forming around. 6:46 In 2012, Alex Krizhevsky and Ilya Sutskever trained AlexNet on two GeForce cards, both from Amazon. They ran in the bedroom at Krizhevsky parents' house, and his parents paid the electricity bill. 6:59 AlexNet finished far ahead of the second place ImageNet system. It made history. And with that, GPUs made history as well. 7:09 NVIDIA believed GPUs could become general purpose computers, but it did not know that deep learning would justify that investment. 7:17 Researchers found the application with hardware built for gamers, and the interesting part is what NVIDIA did once it was told about that. 7:25 Brian Cadanzaro explained the field of deep learning and the AlexNet result and the phenomenal result of that to Jensen in 2013. Was that an aha moment for Jensen? 7:37 Very likely so, because an executive remembers Jensen sending a Friday email that made deep learning a company priority by Monday. 7:45 That decision did not make NVIDIA an AI company over one weekend, but that weekend made it possible for NVIDIA to become an AI company when no one was even thinking about it. 7:57 Jensen told Cadanzaro, allegedly, to imagine all 8,000 NVIDIA employees standing in the parking lot and choose anyone he needed. 8:05 There was no literal gathering on this parking lot, but Jensen was giving Brian a permission to recruit across organizational boundaries because he realized how important that breakthrough in deep learning with GPUs became. 8:19 The products followed over several years, cuDNN in 2014, DGX-1 in 2016, and Tensor Cores in 2017. 8:29 Yes, Jensen protected CUDA during years of doubt, listened when outsiders produced convincing evidence, and moved resources super quickly once he understood what they had found. 8:41 NVIDIA learned that the decisive application for its platform might come from people who did not work there. That is the habit to keep in mind because everything that follows is NVIDIA trying to build it into its system. 8:52 CUDA remained proprietary and tied to NVIDIA hardware, while NVIDIA gradually opened many of the tools developers needed around it. 9:01 Let's see what they opened, because a lot of comments I receive is about NVIDIA not opening anything, which is not true. And Linux shows why they decided to open a lot of things. 9:14 In 2012, NVIDIA's official driver was closed, and the community-built Nouveau driver reconstructed much of the hardware behavior through reverse engineering. 9:25 Over the next decade, Linux became the operating system beneath AI clusters and clouds and supercomputers. 9:32 NVIDIA's largest customers were buying thousands of GPUs and running them on Linux, so poor cooperation had become a problem inside its core business. 9:43 In 2022, NVIDIA opened the kernel modules that connect Linux to the GPU. More GPU management had moved into NVIDIA-controlled firmware, which allowed the kernel-facing code to open while the rest stayed controlled. 9:57 In 2024, the open modules became the recommended option for newer GPUs. User space components, firmware, CUDA, and cuDNN did not open. The same shape shows up in smaller places. 10:10 NCCL is NVIDIA's collective communications library. Once training expanded across main GPUs, they had to exchange tensors constantly. 10:20 NVIDIA solved that difficult communication problem once, published the source, and made large NVIDIA systems easier to use. NCCL became part of the major deep learning frameworks. 10:30 Stack enough of this together and you get the current project list which follows the workflow for calculation and data processing to training, serving, simulation, and robotics. 10:40 There is still a lot of conversations around what is open and what is closed, and if it creates a closed garden. When I spoke with Spencer Juan, I asked, "But doesn't it create sort of like a closed walled garden? 10:53 You know?" No, no, we, we open source everything. The goal is that it should be open source, and it should be open source from the get-go, right? And so Isaac Lab, Isaac Lab Arena, Newton. 11:02 Newton's part of the, the Linux Foundation. It's not even NVIDIA-owned. Uh, we contribute Warp, right? And that's also open source now. 11:08 Apache V2 PhysX, uh, we've also open sourced, and so we've made great, great strides in the last couple of years to open source as much as we can, and the reason is exactly like you said. 11:17 You don't wanna have a closed garden, and the reason you don't wanna have a closed garden is if everyone has different tool chains, how do you enter or exit? 11:23 And so what we really want is to have highways between all these tools, and so create levels of interop, create levels of compatibility, and then also make sure that because they're open source, people can contribute what they need because we can't solve for everything, but we really wanna solve for the things that we can uniquely add value to. 11:38 That's our, our strategy. Spencer's everything refers specifically to the robotics project he was discussing, but NVIDIA still opens a lot, and his answer gives us the practical reason for the push. 11:51 NVIDIA cannot build every simulator, dataset, model, and application. Open tools allow others to fill the gaps while NVIDIA supplies the computing platform underneath them. 12:03 It is, if you think about it, a win-win for everyone. Fast-forward to today's world of generative AI. Open sourcing tools were the first layer. Models are the current one. 12:14 NVIDIA now applies this approach to Nemotron, Cosmos, Groot, and others. 12:18 Releases can include weights, datasets, evaluations, and training recipes, and NVIDIA's own open source page claims all four for its flagship families. Rare company open sourcing that much in AI. 12:32 And again, it's a win-win for everyone. Researchers and developers receive what they need for their experiments, and NVIDIA not shying away from its business logic. In its annual report, NVIDIA puts it this way. 12:44 The rise of high-quality open source foundation model is making advanced AI broadly accessible, and open source AI depends on developer adoption. 12:54 So if it gets deployed on competitor's platform, it could reduce demand for NVIDIA products. So yes, in public, open models are a movement. 13:02 In the risk factors, they are a bet that only pays at the workloads set on NVIDIA hardware. This made me think about the first OpenAI, the company with its initial idea. 13:14 In 2015, it began as a non-profit concerned that powerful AI could become concentrated inside a few companies. Its founders promised to publish research and code, share patents, and collaborate freely. 13:27 As frontier models became expensive and commercially valuable, OpenAI created a for-profit structure and placed its leading systems behind controlled access. NVIDIA has different economics and can allow to do more. 13:42 CUDA remains controlled, yes, while open models create workloads for the hardware, networking, and software NVIDIA sells. 13:50 This is a commercial form of openness, and the incentive may be much more durable than the prompts because NVIDIA earns when many successful models need accelerated computing. So again, it's a win-win. 14:03 Which brings us to the part of the stack NVIDIA does not yet own. What Hugging Face does, it brings NVIDIA closer to the builders. 14:13 At almost $13 billion, against annualized revenue of roughly one hundred and fifty million, the reported price for Hugging Face is about 86 times revenue. 14:23 Hugging Face's own mid-year report puts the public hub at close to three million models, repositories with data sets crossing one million for the first time this year, and it's allegedly around 16, 17 million developers on the platform. 14:38 Hugging Face connects versioning, downloads, libraries, evaluations, demos, enterprise controls, and deployment. It also shows what researchers are using before those workflows become large markets. 14:51 NVIDIA once spent years searching for the application that would justify CUDA. Hugging Face is where many new models and architectures first become visible. 15:00 NVIDIA has been a Hugging Face shareholder since the 2023 round that valued the company at four point five billion. 15:08 Late last year, according to the Financial Times, Hugging Face turned down a $500 million investment from NVIDIA at a $7 billion valuation, reportedly because it did not want one dominant investor influencing its decisions. 15:22 As I said, they always wanted to be very independent. Reporting now suggests the sale talks opened after a different suitor approached the company. 15:31 Nine months after saying no to a minority stake, they reported agreement with Jen-Hsun Huang the whole thing. I spoke with Clem in May, and most likely, the conversations with NVIDIA were already underway. 15:45 He described the ambition to reach tens and hundreds of millions of builders. I think, uh, the numbers of people who are going to be able to become AI builders is going to explode, right? 15:57 It's gonna go from, uh, maybe a few hundred thousands or, or low millions of people who have the skills to do this kind of work to maybe, uh, tens of millions, fifties of millions, maybe a hundred million at some point. 16:12 And for that, you need compute, and for that, you need strong partner, which NVIDIA might become. 16:18 Because I don't think Hugging Face were able to figure out how to make the economics work, and the investors were not very happy about it. There was also an unrealized ambition that now is possible to become true. 16:32 Hugging Face has long wanted to become for machine learning what GitHub became for software. Yes, Hugging Face wanted to become GitHub. It lost its momentum. It didn't become GitHub in time. 16:44 It built the main distribution layer for open models, while the wider workflow remains fragmented across clouds and evaluation systems and inference providers and local runtimes. 16:55 And what we see currently in this somewhat acquisition spread is that it's not the only company being bought for its position in that middle layer. 17:04 It is Stripe that agreed to acquire OpenRadar for more than seven billion dollars, and SpaceX that closed its sixty billion dollar purchase of Cursor a few weeks ago. 17:15 And Cursor reportedly wants to build something like GitHub. And even OpenAI is trying to do something like that. 17:22 To control the platform where developers, AI engineers, machine learning engineers do all their work is a huge power. NVIDIA's position, if buying Hugging Face happened, would be broader than either. 17:36 It already supplies the dominant hardware and programming platform. It supplies training systems and inference software and models. Hugging Face would add the place where developers discover everybody else's models. 17:50 And NVIDIA already leads in inference. With Hugging Face, it would hold the distribution layer of OpenAI, not the company, but the whole industry as well, with influence ranging from cheap to the model page. 18:04 That is the tension. Hugging Face became valuable by supporting NVIDIA's partners and competitors, including AMD, Intel, Google, Amazon, and independent inference providers. 18:15 NVIDIA can fund its expansion and connect the scattered development process. It can also influence rankings, defaults, optimization priorities, and which architectures receive support first. 18:26 People looking at NVIDIA today tend to see one of two companies. Some see a company that was always on top. As of CUDA, AI and the current valuation followed one perfectly executed plan. 18:38 Others remember that NVIDIA, that Linus Torvalds attacked with proprietary drivers and little interest in cooperating with open source developers. 18:47 Both are snapshots of NVIDIA at different moments, and none of them are accurate. NVIDIA did not become open because Jensen suddenly adopted the philosophy of the Linux community. 18:57 The company, the rare company that we have, found a business model in which open source can generate enormous revenue, and it does opens hundreds of libraries, tools, models, and data sets, and encourage people to build on that. 19:13 Those projects create more training, more inference, and simulation workloads. 19:17 NVIDIA then makes money from the controlled computing platform underneath them, and this may be the most effective commercial model for open source in AI. NVIDIA does not have to charge for every model or library. 19:28 It can give away substantial technology because every successful project may create demand for GPUs, networking, CUDA, and enterprise software. 19:37 Hugging Face could finally gain the resources to pursue its original ambition. It wanted to become for machine learning what GitHub became for software. 19:45 But building the complete development platforms requires compute, enterprise distribution, deployment infrastructure, and a great deal of money. It also requires better technical vision. NVIDIA can provide all of that. 19:57 However, Hugging Face is not being transferred to the Linux Foundation, will become the property of the company that already dominates AI hardware. 20:05 Its neutrality would depend on NVIDIA's decisions rather than independent governance. That is the real meaning of this acquisition. 20:13 NVIDIA may have found a way to become one of the largest forces in open source AI while making its proprietary platform even stronger. The contributions of NVIDIA are very real and valuable. 20:25 The commercial interest is equally real, and these two facts support each other. Jensen has already become the inference king of AI. With Hugging Face under his umbrella, he may become the king of open AI too. 20:38 It will be a distinctly NVIDIA version of open source, open enough for millions of people to build and profitable enough for NVIDIA to benefit whenever they succeed. Thank you for watching this little overview. 20:49 What's your take on NVIDIA push into open source and their acquisition of Hugging Face? I'm very curious. [outro music]