Nvidia's 14 Billion Dollar Hugging Face Acquisition

0:00Nvidia is reportedly closing in on a $14 billion acquisition of hugging face, which is made up of a $12.9 billion purchase price and a $1 billion dollar staff retention pool. Right. And if you look at hugging face, I mean, they generate roughly 150 million in annual revenue. Yeah. So when you run the math on that purchase price against that, you know, that revenue stream, you're looking at an 86 times revenue multiple. So the real question that we have to figure out is can an open source ecosystems stay genuinely hardware neutral when the world's most dominant AI chip company actually owns it? Yeah, because you look at that 86 times multiple and it becomes pretty obvious. This isn't a traditional software purchase. No, not at all. A company doesn't pay that kind of premium just to acquire a $150 million business. I mean, that revenue is essentially a rounding error for Nvidia right now. Right. They make that in like no time. Exactly. What they're actually buying is chip demand insurance because if you look at the strategy of the major tech players, Google, Amazon, Meta, they're all aggressively building their

1:04own custom silicon. Yeah, you've got TPUs from Google, Trainiam from Amazon. Right. They're designing these applications specific integrated circuits that are, you know, tailored specifically to handle their own internal AI workloads. If you're not subscribed yet, take a second and hit follow whenever podcast app you're using. Yeah. It helps us keep making this. We appreciate you being here. And the entire goal of those custom chip programs is just to stop buying so many general purpose GPUs from Nvidia. Because every time one of those closed labs successfully ships a custom chip and shifts a workload onto it, Nvidia loses a customer. Yeah. Or at least they lose a portion of that customer's future infrastructure spend, you know. Right. Like Google using a TPU to serve search results means they aren't buying an H100 to do it. But the open source ecosystem, it operates on a totally different set of incentives. Totally different because you have these thousands of developers, smaller startups, academic researchers. They don't have custom silicon fabs. No, there aren't a billions to spin up a chip

2:05factory. Right. So they all default to running on Nvidia hardware because it's just, you know, the path of least resistance. Yeah. And hugging face functions as the distribution layer for that entire ecosystem. Yeah. And it comes down to a hardware concept called the compilation target. Because when a researcher writes a model, they write it in a framework like a pie torch. Right. But underneath pie torch, that code has to be translated into instructions. A specific piece of silicon can actually understand. And for over a decade, Nvidia's proprietary software layer, CUDA has just been the default language for that. Exactly. Google's TPUs use a different compiler. Amazon's chips use a different compiler. But the open source world builds for CUDA, which means you're basically looking at the concept of funneling demand. Yeah. Because by owning the central hub, where developers gather to find and test and share models, Nvidia basically ensures that open source remains this reliable corporate channel that perpetually routes compute demand directly

3:06back to their own GPUs. Because if hugging face remains independent, there's a real risk for Nvidia there. Huge risk. Right. An independent platform might eventually optimize for alternative hardware. Like they might build translation layers that make it frictionless for a developer to just pull a model and instantly deploy it on say Amazon Trainiam or Google's custom chips without even having to rewrite their code. Right. So Nvidia is paying 14 billion to ensure that friction remains incredibly low for their own hardware. And you know, potentially high for everything else. You control the library, you control what books people read. They want to control the default behavior of a million developers. Like when someone clicks deploy on hugging face, Nvidia wants the underlying system to automatically provision a server running their chips. Yeah. They do not want to drop down menu where an Amazon chip is suddenly the cheaper, easier option. No. So the 14 billion is really a defensive maneuver to protect their core monopoly on the data center floor. But you have to question the sustainability

4:06that strategy, right? How so? Like a hardware giant trying to force software developers to stay loyal simply by owning the platform they use because developers are, um, they're highly sensitive to being boxed in. Oh, absolutely. If you're building a startup and you feel a platform is artificially constraining your choices or like inflating your compute costs by hiding cheaper non-invidea options, you're going to look for alternatives. Yeah. Software engineers view friction as a defect and they just route around it. Exactly. But you know, you see that developer flight risk in other platforms. But this whole strategy of buying distribution channels connects directly to moves happening all across the sector right now. Yeah, we're watching this massive wave of infrastructure consolidation. Right. We're the underlying plumbing of the internet is basically trying to buy the interface. I mean, Stripe recently executed the exact same playbook acquiring open router for over $7 billion. Which is wild because Stripe is fundamentally a payments company. Like they process credit cards,

5:09they manage subscription billing. Yeah. Yet they just bought the marketplace that dictates which AI models developers utilize. But the logic is identical to Nvidia buying hugging face. You secure the routing mechanism. Right. Open router lets a developer write an application and then in the background, open router dynamically sends the prompt to whichever provider has the cheapest or fastest version of a model at that exact millisek. And Stripe wants to own the billing layer for all of those millions of microtransactions. Yeah. So they bought the router. And Nvidia wants to own the compute layer. So they're buying the repository. Exactly. You control the flow of traffic. You can extract the toll. And you saw AWS just purchased duck labs too. Right. And duck labs builds tools for developers to manage cloud environments. So the infrastructure giants are just swallowing the developer ecosystem hole. They're securing the on ramps. Because if you own the physical server infrastructure, whether that's Amazon data centers or Nvidia hardware, you want to own the digital interface where people make the

6:12decision about which infrastructure to use. The distance between the developer writing a line of code and the hardware actually executing it is shrinking. And the hardware companies are just trying to own that entire journey. And Nvidia's purchasing power right now is just staggering. I mean, this hugging phase deal is part of a larger $26 billion buying spree. Yeah. They've got agreements with startups like poolside and they just spent 20 billion to acquire rival Grookes assets. Grookes. Right. They were building those highly specialized chips language processing units. Right. LPUs designed explicitly to run models faster than Nvidia GPUs. They were a legitimate threat. So Nvidia just bought their assets. Yeah. They're using their unprecedented hardware margins to just buy up the software layer and neutralize any alternative hardware layers. And when you contrast this AI spending frenzy with the traditional economy, the divergence is just it's striking. Oh, it's completely disconnected. Because while AI companies are dropping tens of billions on digital infrastructure acquisitions,

7:15traditional tech and services are contracting. Like Uber is cutting 3,000 jobs in a restructuring effort right now. And young brands just finalize the sale of pizza hut for 1.5 billion. Right. We have an environment where a global pizza empire with physical locations, supply chains, ovens, thousands of employees sells for a fraction of what a digital routing platform or an open source model repository sells for. The capital flow is just entirely decoupled from traditional physical assets. Investors are placing this massive premium on the choke points of the future digital economy. And discussing that capital flow leads directly into how hugging face achieved this specific $14 billion valuation. Because the backstory here is really interesting. Right. The founders recently turned down a $500 million offer from Nvidia at a $7 billion valuation. Which requires an immense amount of nerve. Yeah. Turning down half a billion dollars in guaranteed liquidity for the founding team, especially when you only have 150 million revenue run rate.

8:15That is a severe test of conviction. I mean, most venture capitalists would push so hard for an exit in that scenario. But the founders understood their own leverage. They knew they held the keys to a very specific kingdom. Right. By holding out, the founders effectively set their own price. Because they realized their value wasn't derived from their revenue stream or their cash flow. No, their value was the threat of what happens to Nvidia if a competitor acquired hugging face instead. Just imagine if Amazon had purchased hugging face. Amazon could deeply integrate the platform with their training chips subsidized the compute costs for developers. And suddenly the default path for open source AI runs entirely on Amazon hardware just bypassing Nvidia completely. Exactly. That threat alone is worth $14 billion to Nvidia. It's essentially a ransom payment disguised as an acquisition. Capital is flowing into AI tooling in a way that's totally disconnected from traditional financial metrics. We see this across the board. Like Cognition

9:15recently closed a funding round, valuing it at $47 billion. Right. Investors and acquires aren't valuing these companies based on discounted cash flows or EBITDA or profit margins. No, they're valuing them based on their structural position in the AI supply chain. If you're the load bearing wall for how developers build artificial intelligence, hardware monopolies will basically pay whatever it takes to ensure you don't collapse or worse, get bought by their enemy. Exactly. And that structural position is built on what hugging face actually created. Like if we shift from the strategy of the acquisition to the technical reality of what they built to become this valuable, it's entirely about removing friction. Yeah, you really have to look at what AI development looked like before hugging face existed. It was a nightmare. Complete nightmare. When a research lab published a paper describing a breakthrough model, developers faced just so many hurdles. You'd read the PDF, realize the model could be useful and then try to actually find the code. Right. And you'd have to track down some incomplete implementation on a random forum.

10:17Yeah. And then you had to recreate the exact software environment the researcher used, which usually involve tracking down obsolete versions of Python or like specific deprecated graphics drivers. And then you had to figure out the tokenizers, which you know, a model doesn't understand text, it understands numbers. Right. And the tokenizer is the tool that chops up English words into numbers. If you use a slightly different tokenizer than the original researcher, the model would just output complete gibberish. Exactly. Then you had to locate the actual neural network weights, which were often hosted on some slow random university FTP server. You untangled all the dependencies, mapped out a deployment path, and tried to force it onto your hardware. It would regularly take weeks of engineering time just to make a model output a single word on your own machine weeks and hugging face solve this by standardizing the handoff between research and production. Yeah, a model transition from being a theoretical paper or just a loose collection of files into a discoverable versioned artifact. They bundled the model weights,

11:19the tokenizer configuration files, and the exact software dependencies into a single digital package. And they built the necessary tooling around it so a developer could download it, test it, and deploy it with like two or three lines of code using their Transformers library. Before hugging face, trying to run an open model was like buying a box of complex machinery, where the instructions are in a dead language and half the screws are missing. And they basically gave you the fully assembled machine and plugged it into the wall for you. Right. We spent a lot of time analyzing the algorithms and the math behind artificial intelligence. But this plumbing is really the primary reason open model spreads so rapidly. It democratized AI development. It moved the capability beyond just a few well-funded labs with dedicated hardware engineers. It allowed a single developer in a basement to pull down a state of the art model and start building a product around it by the end of the day. Because without that standardization, open source AI remains an academic exercise. With it, it becomes an industry. And standardizing that handoff

12:21leads right into the economic cycle of model deployment. We're seeing a very specific flywheel in action right now. Yeah. A researcher or a large lab releases an open model. We're seeing this constantly with models like Lama, Mistral, Deepseek, and you know, Meta's new Mew Spark 1.3. The moment that file hits the hugging face servers, a massive ecosystem activates. Developers immediately download it and test it locally. Then startups take that base model and fine-tune it for specific use cases. Right. They feed it legal documents to make an AI pair legal. Or medical texts for clinical coding or support transcripts for customer service. And then enterprises take those fine-tune models and evaluate them against their own proprietary data in secure environments. But there's a critical economic flip in that process because the initial phases, the testing and the fine-tuning are relatively cheap and open. Yeah. You can run those processes on a high-end workstation or just a small cloud instance. Right. Technologies like Lore-Lower Rank Adaptation allow you to fine-tune a massive model by only tweaking a tiny fraction of its parameters.

13:25It requires very little compute. But the final step training, the final iteration, serving it to millions of users and scaling it at real volume, is intensely compute heavy and extremely expensive. And that final step is exactly where Nvidia captures extraordinary value. Because the open-source model itself stays free. You don't pay a licensing fee to meta to use llama. And you don't pay hugging phase to download it. But Nvidia owns the pipe it flows through when it scales. Right. By owning hugging phase, they ensure that the entire cheap open phase of the flywheel is optimized to seamlessly transition into the expensive, compute heavy phase, running on their GPUs at a massive data center. They're giving away the razor to sell the blades essentially. But in this case, they're buying the store that gives away the razors to make sure the blades only fit their specific handles. And the economics of scaling at volume points directly to a technical shift occurring in how AI operates. Yeah. The industry is actively shifting from a training-centric model toward an infant-centric model.

14:27Right. For the last few years, the massive capital expenditure has been focused on training. Companies buying 10,000 GPUs and running them at maximum capacity for six months straight just to teach a frontier model how to understand language, which requires specialized networking and centralized hardware. But once a model is trained, it needs to be run. That process of generating an answer, which is called inference, is where the bulk of compute will eventually happen over the next decade. And this brings in the threat of local inference. Oh, big time. Technologies like GGML and Lama.cpp are changing the physical location of where inference actually happens. Right. Georgia Gurgenoff and others built software that allows for highly efficient local and edge inference. They use a process called quantization to compress the model weights, turning high precision numbers into lower precision numbers, which just drastically reduces the amount of memory required. Which means capable models can run on laptops, phones, and edge servers without relying on massive centralized GPU farms. You look at Apple Silicon. The M-series chips in

15:30a standard laptop have unified memory architectures, meaning the C2U and the GPU share the same pool of RAM. Yeah, you can fit a heavily quantized, highly capable model directly into the memory of a consumer device. And this pushes back on the narrative that Nvidia is completely untouchable. Because if inference moves to the edge, to devices powered by Apple or Qualcomm chips, Nvidia's data center monopoly becomes less relevant for the daily operation of AI. So Nvidia needs to influence the developer stack to maintain control against that shift to the edge. Yeah, combining hugging faces ecosystem with Nvidia's hardware and their proprietary CUD software architecture allows them to maintain dominance across the entire life cycle. Models to distribution to inference to hardware. Right. If a developer builds a tool on hugging face, Nvidia wants the default inference engine to rely on tensor cores found in their server GPUs, not some system optimized for an edge device processor. And you see how thoroughly Nvidia dominates the current hardware cycle when you look at their competitors. Broadcom's recent

16:33financial outlook failed to rival Nvidia's booming forecasts. Yeah, Broadcom builds networking equipment in custom ASICs, but the market clearly sees Nvidia's merchant silicon as the singular beneficiary of the current data center build out. Every company is buying off the shelf Nvidia GPUs. But the hugging face acquisition is about securing the next phase. It's about making sure that when the build out ends and the optimization phase begins, the software defaults still point back to their hardware. And Nvidia's attempt to control the whole stack naturally brings up developer anxiety. I mean whenever a dominant hardware player buys a neutral software hub, the community reacts. Yeah, developers are experiencing a mix of fear and hope right now and you see widespread comparisons to Microsoft absorbing GitHub. The Microsoft GitHub acquisition is really the closest historical parallel we have for this. It is when Microsoft announced they were buying GitHub, which was, you know, the central repository for the world's open source code, developers panic. Many assumed Microsoft

17:34would ruin it, force integration with windows or just weaponize it against competitors. Some migrated to alternatives like GitLab. But Microsoft largely left GitHub alone to operate independently. Right. And eventually use their vast resources to build tools like co-pilot, which developers now rely on completely. So the hope among developers is that Nvidia follows the Microsoft playbook. Because Nvidia has incredibly deep pockets, they could provide hugging face with the resources to really enrich the platform. They could stimulate research and development in open-weight models. They could subsidize compute costs offering free H100 hours to academic researchers directly through the hugging face interface. Make the platform more robust and solve a lot of the scaling issues that grow in company faces. But the fear is equally strong and it's rooted in Nvidia's history of fiercely protecting its moat. Yeah, a massive corporate entity could easily ruin a neutral platform by subtly optimizing it exclusively for their own hardware. It doesn't even have to be an overt band.

18:38No, if hugging face simply stops updating the software libraries that make it easy to deploy models on Amazon Trainium. Or if they just bury models in the search results that are highly optimized for edge devices instead of server GPUs. The platform loses its utility. Developers fear a gradual closing of the ecosystem, you know. Death by a thousand small software updates that slowly force everyone onto Nvidia Silicon. And developer anxiety over monopolies, mirrors governmental anxiety over monopolies, which brings us to the regulatory reality of a deal like this. Yeah, a transaction of this magnitude featuring the dominant compute layer buying the default neutral layer for open models. That is exactly the type of combination antitrust regulators scrutinize heavily. We are not in an era where regulators just wave through tech acquisitions without looking at the secondary effect. No, we're operating in a highly active regulatory climate right now. Regulators are looking at concentration of power across all sectors from the food supply to digital advertising. Right, we see the US expanding beef price probes

19:40into major grocery chains, examining how a few dominant players control the meat packing industry and dictate prices to farmers and consumers. And we see Google actively defending its ad tech business in court where the argument is that Google owns the tools buyers use, the tools sellers use and the exchange where the transactions actually happen. Regulators are deeply skeptical of companies controlling multiple layers of a supply chain. So the debate really centers on how regulators will classify this specific acquisition. Will they view it as a straightforward software acquisition? Right. Nvidia makes hardware, hugging face makes software. Traditionally, regulators look for horizontal monopolies like a hardware company buying another hardware company. But this is a vertical integration play. Will regulators see it as a move explicitly designed to stifle custom silicon competitors? If Nvidia owns the distribution hub, can they disadvantage hardware startups trying to compete with their GPUs? Regulators look at the concept of foreclosure. Can the acquiring company use the acquired asset

20:42to foreclose competitors from accessing a market? Because if an AI startup builds a revolutionary new chip that's twice as fast as an Nvidia GPU, they need developers to write software for it. And if developers use hugging face and hugging faces code libraries don't support that new chip. That startup is foreclosed from the market. Exactly. Regulators might block the deal entirely based on that theory of harm. But even if regulators don't step in, developer trust might break the deal's value anyway. The fragility of trust is really the central issue here. Yeah, hugging face only became the default platform because it was a trusted place to work across all models, all cloud providers and all hardware platforms. It function as neutral territory. You could pull a model from meta, modify it on a Google cloud instance and deploy it on an Amazon server all using hugging face tools. And that is the ultimate risk for Nvidia. Developers are profoundly pragmatic. They don't have inherent loyalty to a specific code repository. No, if hugging face begins to feel like a closed

21:43Nvidia distribution channel, if that neutrality is compromised, developers and competing platforms will just adapt. They'll migrate quickly to an alternative repository or spin up a decentralized system. The friction of moving code is much lower than the friction of moving physical supply chains. Right, you can clone and get repository in seconds. You cannot build a new semiconductor fabrication plant in a weekend. You can buy the platform for $14 billion, but you cannot buy the trust and openness that made it the platform in the first place. The moment you buy it to exert control, you risk destroying the exact neutrality that made it valuable. Nvidia just dropped $14 billion, not just to acquire code or a brand, but to dictate the default behavior of the entire open source AI ecosystem. What happens to the fundamental concept of open source AI when the centralized hubs required to host it are all owned by the infrastructure monopolies. If you're not subscribed yet, take a second and hit follow on whatever app you're using. It helps us keep making this

22:43we appreciate you being here. Also check out our YouTube channel for more business and tech updates. There's a link in the description.