Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

I asked Jensen about TPU competition, Nvidia’s lock on the ever more bottlenecked supply chain needed to make advanced chips, whether we should be selling AI chips to China, why Nvidia doesn’t just become a hyperscaler, how it makes its investments, and much more. Enjoy! Watch on YouTube https://youtu.be/Hrbq66XqtCo; read the transcript https://www.dwarkesh.com/p/jensen-huang. Sponsors * Crusoe’s https://crusoe.ai/dwarkesh cloud runs on state-of-the-art Blackwell GPUs, with Vera Rubin deployment scheduled for later this year. But hardware is only part of the story—for inference, Crusoe’s MemoryAlloy tech implements a cluster-wide KV cache, delivering up to 10x faster TTFT and 5x better throughput than vLLM. Learn more at crusoe.ai/dwarkesh https://crusoe.ai/dwarkesh * Cursor https://cursor.com/dwarkesh helped me build an AI co-researcher over the course of a weekend. Now I have an AI agent that I can collaborate with in Google Docs via inline comment threads! And while other agentic coding tools feel like a total black-box, Cursor let me stay on top of the full implementation. You can try my co-researcher out at github.com/dwarkeshsp/ai_coworker https://github.com/dwarkeshsp/ai_coworker, or get started on your own Cursor project today at cursor.com/dwarkesh https://cursor.com/dwarkesh * Jane Street https://janestreet.com/dwarkesh spent ~20,000 GPU hours training backdoors into 3 different language models, then challenged my audience to find the triggers. They received some clever solutions—like comparing the base and fine-tuned versions and extrapolating any differences to reveal the hidden backdoor—but no one was able to solve all 3. So if open problems like this excite you, Jane Street is hiring. Learn more at janestreet.com/dwarkesh https://janestreet.com/dwarkesh Timestamps (00:00:00) – Is Nvidia’s biggest moat its grip on scarce supply chains? (00:16:25) – Will TPUs break Nvidia’s hold on AI compute? (00:41:06) – Why doesn’t Nvidia become a hyperscaler? (00:57:36) – Should we be selling AI chips to China? (01:35:06) – Why doesn’t Nvidia make multiple different chip architectures? This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.dwarkesh.com https://www.dwarkesh.com?utm_medium=podcast&utm_campaign=CTA_1

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Dwarkesh Patel?

  We've seen the evaluations of a bunch of software companies crash because people are expecting AI to commoditize software. And there's a potentially naive way of thinking about things, which is like, look, Nvidia sends a GDS2 file to TSMC. TSMC builds the logic dies, it builds the switches, then it packages them with the HBM that SK Hynix and Micron and Samsung make, then it sends it to an ODM in Taiwan where they assemble the racks. And so Nvidia is fundamentally making software that other people are manufacturing. And if software gets commoditized, does Nvidia get commoditized?

• 0:31Speaker 01

  Well, in the end, something has to transform electrons to tokens. That transformation, there's no, the transformation of electrons to tokens and making those tokens more valuable over time, I don't, I think that that, that's hard to, hard to completely commoditize. The transformation from electrons to tokens is such an, such an incredible journey. And, and making that token, you know, it's like making one molecule more valuable than another molecule, making one token more valuable than another.

• 1:13Speaker 01

  The amount of artistry, engineering, science, invention that goes into making that token valuable, obviously, we're, we're watching it happening in real time. And so, so the, the, the transformation, the manufacturing, all of the science that goes in there is far from deeply understood and is far from, the journey is far from, far from over. And so, so I doubt that it will happen. We're going to make it more efficient, of course. I mean, the whole, the whole thing about Nvidia, in fact, the way that you framed the question is my mental model of our company. The input is electron, the output is tokens. That is in the middle, Nvidia.

• 1:57Speaker 01

  And our job is to, to do as much as necessary, as little as possible, to enable that transformation to be done at incredible capabilities. And, and what I mean by as little as possible, whatever I don't need to do.

• 2:13Speaker 01

  I partner with somebody and I make it part of my ecosystem to do. And if you look at Nvidia today, we probably have the largest ecosystem of partners, both in supply chain upstream, supply chain downstream. All of the computers, computer companies and all the application developers and all the model makers and all the, you know, AI is a five-layer cake, if you will. And we have ecosystems across the entire five layers. And so we try to do as little as possible. But the part that we have to do, as it turns out, is insanely hard. And I don't think that that gets commoditized. In fact, I also don't think that the enterprise software companies, the tools makers,

• 2:57Speaker 01

  you know, most of the software companies today are tools makers. Some of them are not, but some of them are workflow codification, you know, systems. But for a lot of companies, they're tool makers. For example, Excel is a tool, PowerPoint is a tool, Cadence makes tools, Synopsys makes tools. I actually see the opposite of what people see. I think the number of agents are going to grow exponentially. The number of tool users are going to grow exponentially. And it's very likely that the number of instances of all these tools are going to skyrocket.

• 3:39Speaker 01

  It is very likely the number of instances of Synopsys design compiler is going to skyrocket. And the number of agents that are going to be using the floor planners and all of our layout tools and our design role checkers, the number of agents that are, today we're limited by the number of engineers, tomorrow those engineers are going to be supported by a bunch of agents. They're going to be exploring out the design space like you've never seen explore before and want to use the tools that we use today. And so I think tool use is going to cause these software companies to skyrocket. The reason why it hasn't happened yet is because the agents aren't good enough at using their tools yet. And so either these companies are going to build the agents themselves.

• 4:23Speaker 01

  Or agents are going to get good enough to be able to use those tools. And I think it's going to be a combination of both.

• 4:29Dwarkesh Patel?

  I think in your latest filings, you had almost $100 billion in purchase commitments with people, foundries, memory, packaging, and then semi-analysis has reported that you will have $250 billion of these kinds of purchase commitments. And so one interpretation is Nvidia's mode is really that you've locked up many years of these scarce components that are, somebody else might have an accelerator, but can they actually get the memory to build it? Can they actually get the logic to build it? And this is really Nvidia's big mode for the next few years.

• 5:00Speaker 01

  Well, it's one of the things that we can do that is hard for someone else to do. The reason why we could, we've made enormous commitments upstream. Some of it is explicit, these commitments that you mentioned. Some of it is implicit. For example, a lot of the investments that are upstream are made by our supply chain because I said to the CEOs, let me tell you how big this industry is going to be and let me explain to you why, and let me reason through it with you, and let me show you what I see. And so as a result of that, that process of informing, inspiring, aligning with CEOs

• 5:43Speaker 01

  of all different industries upstream, they're willing to make the investments. Now, why are they willing to make the investments for me and not someone else? And the reason for that is because they know that I have the capacity to buy it, buy their supply, and sell it through my downstream. The fact that Nvidia's downstream supply chain and our downstream demand is so large, they're willing to make the investment upstream. And so if you look at GTC, and people are marveled by the scale of GTC and the people that go, it's a 360 degrees, the entire universe of AI, all in one place. And they're all in one place because they need to see each other.

• 6:27Speaker 01

  I bring them together so that the downstream could see the upstream, the upstream could see the downstream, and all of them could see all the advances in AI. And very importantly, they can all meet the AI natives and all the AI startups that are all being built and all the amazing things that are happening so that they could see firsthand all the things that I tell them. And so I spend a lot of my time informing directly or indirectly our supply chain and our partners and our ecosystem about the opportunity that's in front of us. You know, most of my keynotes, you know, some people always say, you know, Jensen, in most keynotes, it's like one announcement after another announcement, after another announcement, after another announcement.

• 7:12Speaker 01

  Our keynotes are, there's always a part of it that's a little torturous in the sense that it almost comes across like an education. And in fact, that's exactly on my mind. I need to make sure that the entire supply chain, upstream and downstream, the ecosystem, understands what is coming at us, why it's coming, when it's coming, how big is it going to be, and be able to reason about it systematically, just like I reason about it. And so I think the modus you describe it, we're able to, of course, build for a future,

• 7:52Speaker 01

  if our next several years is a trillion dollars in scale, we have the supply chain to do it. Without our reach, the velocity of our business, you know, just as there's cash flow, there's supply chain flow, there are turns. Nobody's going to build a supply chain for an architecture if the architecture, the business turns is low. And so our ability to sustain the scale is only because our downstream demand is so great and they see it and they all hear about it, they see it all coming. And so that's, it allows us to do the things that we're able to do at the scale we're able to do.

• 8:32Dwarkesh Patel?

  I do want to understand more concretely whether the upstream can keep up. For many years now, you guys have been 2Xing revenue year over year, you guys have been more than tripling the amount of flops you're providing to the world year over year.

• 8:44Speaker 01

  And 2Xing at the scale now is really incredible.

• 8:47Dwarkesh Patel?

  Exactly. So then you look at logic, say, you're the biggest customer on TSMCs. N3 node and you're one of the biggest on N2. AI as a whole this year is going to be 60% of N3. It's going to be 86% next year, according to some analysis. How do you 2X if you're the majority? And how do you do that year over year? So are we in a regime now where the growth rate in AI compute has to slow because of upstream? Do you see a way to get around these? How do we build 2X more fabs year over year ultimately?

• 9:21Speaker 01

  Yeah, at some level. The instantaneous demand is greater than the supply upstream and downstream in the world. And it could be at any instance, we could be limited by the number of plumbers, which actually happens.

• 9:46Dwarkesh Patel?

  The plumbers are admitted to next year's GTC.

• 9:49Speaker 01

  By the way, great idea. But that's a good condition. You want a market, you want an industry where the instantaneous demand is greater than the total supply of the industry. The opposite is obviously less good. If we're too far apart, if one particular item, one particular component is too far away, obviously the industry swarms it. So for example, notice people aren't talking very much about COAS anymore. And the reason for that is because for two years, we swarmed a living daylights out of it. And we double, double, double on several doubles. And now I think we're in a fairly good shape.

• 10:30Speaker 01

  And TSMC now knows that COAS supply has to keep up with the rest of the logic demand and the memory demand. And so they're scaling COAS and they're scaling future packaging technologies at the same level as they scale logic, which is terrific because for a long time, COAS was rather specialty and HBM memory was rather specialty. But they're not specialties anymore. People now realize they're mainstream computing technology. And then, and of course, we're now much more able.

• 11:05Speaker 01

  To influence a larger scope of our supply chain. In the past, in the past, you know, in the beginning of the AI revolution, all the things that I say now, I was saying five years ago. And some people believed in it and invested in it. For example, Sanjay and the Micron team, I still remember the meeting really well, where I was clear about exactly what's going to happen and why it's going to happen and the predictions that I've today. And they really doubled down on it and we partnered with them and across LPDDR, across HBO Memories, they really invested in it. And it obviously has been tremendous for the company.

• 11:49Speaker 01

  Some people came a little bit later, but now they're all here. And so I think each one of these generations, each one of these bottlenecks gets a great deal of attention. And now we're pre-fetching the bottlenecks years in advance. So for example, the investments that we've done with Lumentum and Coherent and all of the Silicon Photonics ecosystem. The last several years, we really reshaped the ecosystem and the supply chain Silicon Photonics. We built up an entire supply chain around TSMC. We partnered with them on Coop, invented a whole bunch of technology. We licensed those patents to the supply chain,

• 12:34Speaker 01

  keep it nice and open. And so we're preparing the supply chain through invention of new technologies, new workflows, new testing equipment, double-sided probing, investing in companies, helping them scale up their capacity. And so you could see that we're trying to shape the ecosystem so that it's ready, the supply chain, so that it's ready to support the scale.

• 12:57Dwarkesh Patel?

  It seems like some bottlenecks are easier than others. And so scaling up Coos versus scaling up...

• 13:02Speaker 01

  I went to the hardest one, by the way.

• 13:04Dwarkesh Patel?

  Which is?

• 13:05Speaker 01

  Plumbers.

• 13:07Dwarkesh Patel?

  Yeah, that's true.

• 13:08Speaker 01

  Yeah, I actually went to the hardest one. Yeah, plumbers and electricians. And the reason for that is because, and this is one of the concerns that I have about the doomers, describing the end of work and killing of jobs. And you know, one of the things that if we discourage people from being software engineers... We're going to, we're not a software engineers. And, and the same prediction 10 years ago, some of the, some of the doomers were, were saying that, we're telling people, whatever you do, don't be a radiologist. And you might hear some of those, some of those videos are still on the web. You know, radiology is, is going to be the first career to go. Nobody's, the world's not going to need any more radiologists. Guess what we're short of? Radiologists.

• 13:54Dwarkesh Patel?

  Oh, but, okay, so going back to this point about, well, some things you scale, other things like, how do you actually get, how do you actually manufacture 2X the amount of logic a year, ultimately that's bottlenecked by, memory and logic are bottlenecked by EUV. How do you get to 2X as many EUV machines a year?

• 14:09Speaker 01

  Yeah.

• 14:10Dwarkesh Patel?

  Year over year?

• 14:10Speaker 01

  None of, none of that's impossible to scale quickly. You just need to, you, you could do, all of that is easy to do within two or three years. You just need a demand signal. It's not, it, it, once you, once you can build one, you could build 10, and once you can build 10, you can build a million. And so, these things are not, not hard to replicate.

• 14:30Dwarkesh Patel?

  How far down the supply chain do you go? Where you, do you go to ASML and say, hey, if I look out three years from now, for me to, for Nvidia to be generating 2 trillion a year, in a year in revenue, we need way more AUV machines and?

• 14:41Speaker 01

  Some of them I have to directly, some of them indirectly, and some of them, if I can convince TSMC, ASML will be convinced. And so that's, that, you know, we have to think about the critical, critical pinch points and, but if TSMC is convinced, you'll have plenty of EUV machines in a few years. And so, none of that, my point is that, none of the bottlenecks last longer than a couple, two, three years. None of them. And meanwhile, meanwhile, we're improving computing efficiency by 10X, 20X, in the case of Hopper to Blackwell, some 30, 50X. We're, we're coming up with new algorithms because CUDA is so flexible.

• 15:24Speaker 01

  We're, we're developing all kinds of new techniques so that we drive efficiency in addition to increasing capacity.

• 15:30Dwarkesh Patel?

  Yeah.

• 15:31Speaker 01

  And so, so there, those, those are, those are things that, that none of that worry me. It's the stuff that's downstream from us. Energy policies that prevent energy from, you know, you can't grow, you can't create, you can't create an industry without energy. You can't create a whole new manufacturing industry without energy. We want to re-industrialize the United States. We want to bring back chip manufacturing and computer manufacturing and packaging. And we want to build new things like EVs and robots. And we want to build AI factories. And you can't build any of these things without energy. And those things take a long time. But more chip capacity, that's a two, three year problem. More COAS capacity, two, three year problem.

• 16:13Dwarkesh Patel?

  Interesting. I feel like I have guests tell me the exact opposite thing sometimes. And I don't, in this case, I just don't have the technical knowledge to adjudicate, but.

• 16:20Speaker 01

  Well, the beautiful thing is you're talking to the expert.

• 16:22Dwarkesh Patel?

  Yeah. True, true.

• 16:25Dwarkesh Patel?

  Okay, I want to ask about your competitors.

• 16:28Speaker 01

  Yeah.

• 16:28Dwarkesh Patel?

  So if you look at TPU, arguably, two out of the top three models in the world, Claude and Gemini, were trained on TPU.

• 16:39Dwarkesh Patel?

  What does that mean for Nvidia going forward?

• 16:41Speaker 01

  Well, we have, we have a very different, we build a very different thing.

• 16:46Speaker 01

  You know, what Nvidia built is accelerated computing, not a tensor processing unit. And accelerated computing is used for all kinds of things. You know, molecular dynamics and quantum chromodynamics. And it's used for data processing, data frames, structured data, unstructured data. It's used for fluid dynamics, particle physics, you know? And in addition, we use it for AI. And so accelerated computing is, is much more diverse. And although AI is the conversation today is obviously very important and impactful, computing is much broader than that.

• 17:30Speaker 01

  And what Nvidia has done is reinventing, reinvented the way computing is done from general purpose computing to accelerated computing. Our market reach is far greater than any, any TPU, any ASIC can possibly have. And so if you look at our position. We're the only company that accelerates applications of all kinds. We have a gigantic ecosystem. And so all kinds of frameworks and algorithms all run on Nvidia. And because our computers are designed to be operated by other people, anyone who's an operator could buy our systems.

• 18:12Speaker 01

  Most of these home-built systems, you have to be your own operator because we're never designed to be flexible enough for other people to operate. And so as a result of the fact that anybody can operate our systems, we're in every cloud, including Google and Amazon and Azure and OCI. And so whether you want to operate it to rent or operate it, if you want to operate it to rent, you better have a large ecosystem of customers in many industries that be the off-takers. If you're operating it, if you want to operate it for yourself, we obviously have the ability to help you operate yourself, like for example, for Elon with XAI.

• 18:54Speaker 01

  And because we could enable operators in any company, in any industry, you could use it to build a supercomputer for scientific research and drug discovery at Lilly. And so we can help them operate their own supercomputer and use it for the entire diversity of drug discovery and biological sciences that we accelerate. And so there are just a whole bunch of applications that we can address that you can't do so with TPUs. Because Nvidia's built CUDA as a fantastic tensor processing unit as well, but it does every lifecycle

• 19:36Speaker 01

  of data processing and computing and AI and so on and so forth. And so our market opportunity is just a lot larger. Our reach is a lot greater. And because we have such a large, we basically support every application in the world now. You could build Nvidia systems anywhere and know that there will be customers for it. And so it's a very different thing.

• 20:00Dwarkesh Patel?

  This is going to be sort of a long question, but you know, you have spectacular revenue. And this revenue is mostly, you're not making 60 billion a quarter from pharma and quantum. You're making it because AI is unprecedented technology that is growing unprecedentedly fast. And so then the question is, what is best for AI specifically? And I'm not in the details, but I talked to my AI researcher friends and they say, look, when I use a TPU, it's this big systolic array that's perfect for doing matrix multiplies, whereas a GPU is very flexible. It's great when you have lots of branching, when you have irregular memory access. But with these, you know, what is AI? Just like these very predictable matrix multiplies again and again and again.

• 20:40Dwarkesh Patel?

  And you don't have to give up any diarrhea for warp schedulers, for, you know, switches between threads and memory banks. And so the TPU is really optimized for the majority, the bulk of this growth in revenue and use case for a compute that is coming online right now. Yeah, I wonder how you react to that.

• 21:01Speaker 01

  Matrix multiplies is an important part of AI, but it's not the only part of AI. And if you want to come up with a new attention mechanism, or if you want to disaggregate in a different way, if you want to come up with a whole new type of architecture altogether, for example, you know, a hybrid SSM. If you want to use a, if you want to create a model that, that, that fuses diffusion and autoregressive somehow, you want an architecture that's just generally programmable. And, and we run everything you can imagine. And so that's the advantage.

• 21:42Speaker 01

  It allows for invention of new algorithms a lot more, a lot, a lot more easily. And so, because it's a programmable system. And, and the ability to invent new algorithms is really what makes AI advance so quickly.

• 22:00Speaker 01

  TPU is like anything else is impacted by Moore's law. And we know that Moore's law is increasing about 25% per year. And so the only way to really get 10x leaps, 100x leaps, is to fundamentally change the algorithm and how it's computed every single year. And that's Nvidia's fundamental advantage. The only reason why we were able to make Blackwell to Hopper 50 times, you know, I said it was 35 times. And, and, and when I first announced it was going to, Blackwell is going to be 35 times more energy efficient than Hopper, nobody believed it. And, and, and then Dylan wrote an article.

• 22:43Speaker 01

  He said, he said, in fact, in fact, I sandbagged, it's actually 50 times. And you can't reasonably do that with just Moore's law. And so the, the way that we solve that problem is new out, new models, MOEs, paralyzed and disaggregated and, and distributed across a computing system. And without the ability to really get down and come up with new kernels with CUDA, it's really hard to do. And, and so the combination of the programmability of our, of our architecture, the, the fact that Nvidia is an extreme co-design company,

• 23:25Speaker 01

  where we could even offload some of the computation into the fabric itself, NVLink, for example, into the network, Spectrum X. And that we could affect change across the processors, the system, the fabric, the libraries, the algorithm. All of that was done simultaneously. Without CUDA to do that, I wouldn't even know where to start.

• 23:53Dwarkesh Patel?

  My sponsor, Crusoe, was among the first clouds to offer Nvidia's Blackwell and Blackwell Ultra platforms. And they just announced their Nvidia Vera Rubin deployment, scheduled for later this year. But access to state-of-the-art hardware is only part of the story. For example, most inference engines already do KV caching for a single user's forward passes. But Crusoe does it across users and GPUs. So if 1,000 agents are running on the same system prompt, Crusoe only has to compute the KV cache once for it to become available to every single GPU in the cluster. This is especially important as systems get more agentic and require much longer prefixes in order to use tools and access files. In a recent benchmark, Crusoe was able to deliver up to 10 times faster time to first token and up to five times better throughput than VLLM.

• 24:35Dwarkesh Patel?

  This is just one among many reasons that you should run your inference workload with Crusoe. And if you need GPUs for training, you don't need to switch clouds. Crusoe has got you covered there, too. Go to crusoe.ai/thoracash to learn more. So this gets at an interesting question about Nvidia's clientele, where if 60% of your revenue is coming from these big five hyperscalers. You know, in a different era with different customers, let's say it's professors who are running experiments, and they are helped a bunch by, they need CUDA. They can't use another accelerator. They need to just run PyTorch with CUDA and have everything optimized. But if you've got these hyperscalers, they have the resources to write their own kernels.

• 25:17Dwarkesh Patel?

  In fact, they have to to get that extra last 5% that they need for their specific architecture. Anthropic, Google are mostly running their own accelerators or running TPUs. And Tranium, but even OpenAI, using GPUs has Triton, which they're like, we need our own kernels. So they've down to CUDA C++. They've, instead of using Kublas and Nickel and everything, they've got their own stack, which compiles to other accelerators as well. And so if most of your customers can and do make replacements for CUDA, to what extent is CUDA really the thing that is going to make Frontier AI happen on Nvidia?

• 25:58Speaker 01

  CUDA is a rich ecosystem. And so if you want to build on any computer first, building on CUDA first is incredibly smart.

• 26:11Speaker 01

  And because the ecosystem is so rich, we support every framework. If you want to create custom kernels, if you need, for example, we contribute enormously to Triton. And so the backend of Triton, huge amounts of Nvidia technology. We're delighted to help every framework become as great as it can be. And there's lots and lots of frameworks. There's Triton, there's VLM, there's SG Lang, and then there's more, right? And now there's a whole bunch of new reinforcement learning frameworks coming out. You know, you've got Verol, you've got Nemo RL, you've got a whole bunch of new. And then now with post-training and reinforcement learning, that entire area is just exploding.

• 26:52Speaker 01

  Right? And so if you want to build on an architecture, building on a CUDA makes the most sense. Because you know that the ecosystem is great. You know that if something happens, it's more likely in your code and not in the mountain of code underneath. You know, don't forget the amount of code that you're dealing with when you're building these systems. When something doesn't work, was it you or was it the computer? You would like it always to be you and to be able to trust the computer. And obviously, we still have lots and lots of bugs ourselves. But our system is so well wrung out that you could at least build on top of the foundation. So that's number one, is that the richness of the ecosystem, the programmability of it, the capability of it.

• 27:36Speaker 01

  The second thing is, if you were a developer and you were building anything at all, the single most important thing you want more than anything is install base. You want the software that you run to run on a whole bunch of other computers. You don't want to build a software, you're not building software just for yourself. You're building software for your fleet or for everybody else's fleet because you're a framework builder. And Nvidia's CUDA ecosystem is ultimately its great treasure. We are now, I don't know how many, several hundred million GPUs. Every cloud has it, goes back to A10, A100, H100, H200, you know?

• 28:15Speaker 01

  The L series, the P series, I mean, there's a whole bunch of them. And they're in all kinds of sizes and shapes. And if you're a robotics company, you want that CUDA stack to actually run in the CUDA in the robot itself. We're literally everywhere. And so the install base says that once you develop the software, once you develop the model, it's going to be useful everywhere. And so the install base is just too incredibly valuable. And then lastly, the fact that we're in every single cloud makes us genuinely unique. Because you're an AI company and you're an AI developer, you're not exactly sure which CSP you're going to partner with and where you would like to run it. And we run it everywhere, including on-prem for you, if you like.

• 28:58Speaker 01

  And so I think that the richness of the ecosystem, the expansiveness of the install base, and the versatility of where we are, that combination makes CUDA invaluable.

• 29:15Dwarkesh Patel?

  That makes a lot of sense. I guess the thing I'm curious about is whether those advantages matter a lot to your main customers. Like, there's many people who they might matter for, the kind of person who can actually build their own software stack, who will make up most of your revenue. Especially if you go to a world where AI is getting especially good at the things which have tight verification loops, where you can RL on them. And then this question of how do you write a kernel that does attention or MLP the most efficiently across a scale up? It's a very verifiable sort of feedback loop. And so, oh, can everybody, can all the hyperscalers write these custom kernels for themselves?

• 29:58Dwarkesh Patel?

  And they might still, Nvidia still has great Python performance, or they might still prefer to use Nvidia. But then the question is, does it just become a question of who is offering the best specs, the best flops and memory and memory bandwidth for a given dollar? Where historically Nvidia has just had, and still has, you know, the best margins in all of AI across hardware and software, 70% plus because of this CUDA mode? And the question is, oh, can you sustain those margins if for most of your customers, they can actually afford to build instead of the CUDA mode?

• 30:33Speaker 01

  The number of engineers we have assigned to these AI labs is insane. Working with them, optimizing their stack. And the reason for that is because, because nobody knows our architecture better than we do. And these architectures are not, not as general purpose as a CPU. The reason, the reason why a CPU is so, you know, CPU is kind of like, like a Cadillac, you know? It's, it just always, you know, it's a nice cruiser. It never goes too fast. Everybody drives it pretty well, you know? It's got cruise control, you know, and everything is easy. But in a lot of ways, Nvidia's GPUs are accelerators are kind of like F1 racers.

• 31:16Speaker 01

  And yeah, I could imagine everybody's able to drive it at 100 miles an hour. But it takes quite a bit of expertise to be able to push it to the limit. And we use, we use a ton of AI to create the kernels that we have. And I'm pretty sure we're going to still be needed for quite some time. And so our expertise helps our, our, our AI labs partners get another 2X out of their stack easily, oftentimes. It's not unusual that we, you know, by the time that we're done optimizing their stack or optimizing a particular kernel, their model sped up by 3X, 2X, 50%.

• 32:01Speaker 01

  That's a huge number, especially when you're talking about the install base of the fleet that they have, of all the hoppers and black walls that they have. When you increase it by a factor of two, that doubles their revenues. That directly translates to revenues. Nvidia's computing stack is the best performance particio in the world, bar none.

• 32:24Speaker 01

  Nobody can demonstrate to me that any single platform in the world today has better performance TCO ratio, not one company. And in fact, in fact, the benchmarks are out there. Dylan's inference max is sitting out there for everybody to use. And not one, TPU won't come, Trainium won't come. I encourage them to use inference max and demonstrate their incredible inference cost. It's really, really hard. Nobody wants to show up. ML perf, I would, I would welcome Trainium to demonstrate their 40% that they claim all the time.

• 33:07Speaker 01

  I would, I would love to hear them demonstrate the, the cost advantage of TPUs. It makes no sense in my mind. It makes absolutely zero sense. On first principles, it makes no sense. And so I, I think the, I think the, the reason why we're so successful is simply because our TCO is so great. There's a second, you say 60% of our customers are the top five, but most of that business is external. For example, most of AWS is, most of Nvidia in AWS is for external customers, not internal use. Most of our customers at Azure, obviously, all of our customers are external, all of our customers at OCI are external, not internal use.

• 33:49Speaker 01

  The reason why they, they favor us is because our reach is so great. We can bring them all of the great customers in the world. They're all built on Nvidia. And the reason why all these companies are built on Nvidia is because our reach and our versatility is so great. And so, so I think, I think the flywheel is really install base, the programmability of our architecture, the richness of our ecosystem. And the fact that there's so many AI companies in the world, there's tens of thousands of them now. And if you were one of those AI startups, what architecture would you, would you choose? You would choose an architecture that's most abundant, where the most abundant in the

• 34:30Speaker 01

  world, the one has the largest install base, where the most largest install base, and one that has a rich ecosystem. And so that's the flywheel, that's the reason why between the combination of one, our perf per dollar is so great. That they have the lowest cost tokens. Second, our perf per watt is the highest in the world. And so if one of these companies, if our partners built a one gigawatt data center, that one gigawatt data center better deliver the maximum amount of revenues that and number of tokens, which directly translates to revenues. You wanted to generate as many tokens as possible, maximize the revenues for that data center.

• 35:13Speaker 01

  We are the highest tokens per watt architecture in the world. And then lastly, if your goal is to rent the infrastructure, we have the most customers in the world. And so that's the reason why the flywheel works.

• 35:25Dwarkesh Patel?

  Interesting. I guess the question comes down to what is the actual market structure here? Because even if there's other companies, there could have been a world where there's tens of thousands of AI companies that have roughly equal share of compute. But if even through these five hyperscalers, really, the people on Amazon using the computer, Anthropic, OpenAI, and these big, big foundation labs who can themselves afford and have the ability to make different accelerators work.

• 35:53Speaker 01

  No, I think your assumption is premise is wrong.

• 35:58Dwarkesh Patel?

  Maybe. Let me ask you a slightly different question.

• 36:01Speaker 01

  Come back and make me correct your premise.

• 36:04Dwarkesh Patel?

  Okay. Let me just ask you a different question, which is, okay.

• 36:07Speaker 01

  But still, make sure that make me come back and fix, because it's just too important to AI. It's too important to the future of science, it's too important to the future of the industry. That premise, the premise, look.

• 36:21Dwarkesh Patel?

  Let me just finish the question and then address it together. So what do you think if, if all these things are true about price performance and performance per watt, et cetera, are true? Why do you think it is the case that, say, Anthropic, for example, just announced a couple of days ago, they have a multi-gigawatt deal with Broadcom and Google for TPUs and majority of their compute, obviously for Google, it's TPUs are majority of compute. So if I look at these big AI companies, it seems like a lot of their, there was some point where it was all Nvidia, and now it's not. And so I'm curious how to square, if these things are true on paper, why are they going with other accelerators?

• 37:03Speaker 01

  Yeah. Anthropic is a unique instance and not a trend. Without Anthropic, why would there be any TPU growth at all?

• 37:14Speaker 01

  It's 100% anthropic. Without anthropic, why would there be any uranium growth at all? It's 100% anthropic. I think that's fairly well-known and well-understood. It's not that there's an abundance of ASIC opportunities. There's only one anthropic.

• 37:31Dwarkesh Patel?

  But OpenAI deals with AMD. They're building their own Titan accelerator.

• 37:35Speaker 01

  Yeah, but they're mostly, and we could all acknowledge they're vastly Nvidia. And we're going to still do a lot of work together. Yeah. And I'm not offended by other people using something else and trying things. If they don't try these other things, how would they know how good ours is? And sometimes you've got to be reminded of it. And we have to continuously earn the position that we're in. There are always big claims. And look at the number of ASICs that have been canceled. Just because you're going to build an ASIC, you still have to build something better than Nvidia. And it's not that easy, building something better than Nvidia. It's not sensible, actually.

• 38:19Speaker 01

  Nvidia's got to be missing something seriously. Because our scale, our velocity, we're the only company in the world that's cranking it out every single year. Big leaps every single year.

• 38:31Dwarkesh Patel?

  I guess their logic is that, hey, it doesn't need to be better. It just needs to be not more than 70% worse because they're paying you 70% margins.

• 38:39Speaker 01

  No, no, no, don't forget. Even in ASICs, margins are really quite high. Nvidia's margin is 70%, let's say, but in ASIC, margin is 65. What are you really saving?

• 38:51Dwarkesh Patel?

  Oh, you mean from Broadcom or something like that?

• 38:52Speaker 01

  Yeah, sure. You've got to pay somebody. And so I think the ASIC margins are incredibly good, from what I can tell. And they believe it so too. And so they're quite proud of their incredible ASIC margins. And so you ask the question, why? A long time ago, we just didn't have the ability to do it.

• 39:17Speaker 01

  And this is, and at the time, I didn't deeply internalize how difficult it would be to build a foundation AI lab like OpenAI and Anthropic.

• 39:34Speaker 01

  And the fact that they needed huge investments from the supplier themselves. We just weren't in a position to make the multi-billion dollar investment into Anthropic so that they could use our compute. But Google and AWS were, and they put in huge investments in the beginning so that Anthropic in return used their compute. We just weren't in a position to do so at the time, nor did I, I would say my mistake is I didn't deeply internalize that they really had no other options, that a VC would never

• 40:13Speaker 01

  put in five, $10 billion of investment into an AI lab with the hopes of it turning out to be Anthropic. And so that was my miss, but even if I understood it, I don't think we would have been in a position to do that at the time. But I'm not going to make that same mistake again. And I'm delighted to invest in OpenAI and I'm delighted to help them scale and I believe it's essential to do so. And then when I was able to, when Anthropic came to us, I'm delighted to be an investor, delighted to help them scale.

• 40:54Speaker 01

  But we just weren't at the time able to do so. If I could rewind everything, Nvidia could have been as big back then as we are now, I would have been more than happy to do it.

• 41:06Dwarkesh Patel?

  This is actually quite interesting, which is for many years Nvidia has been this

• 41:13Dwarkesh Patel?

  D company in AI, making money, making lots of money. And now you're investing it, it's been reported that you've done up to $30 billion in OpenAI and $10 billion in Anthropic. But now their valuations have increased and I'm sure they'll continue to increase. And so over all these many years, you were giving them the compute, you saw where AI was headed, and then they were worth like one tenth what they are now a couple of years ago, or even a year ago in some cases.

• 41:42Dwarkesh Patel?

  And you had all this cash.

• 41:45Dwarkesh Patel?

  There's a world where either Nvidia themselves becomes a foundation lab, does a huge investment to make that possible, or has made the deals you've made now at current valuations much earlier on, and you had the cash to do it. So I am curious, actually, why not have done it earlier?

• 42:02Speaker 01

  We did it as soon as we could have. We did it as soon as we could have. And if I could have, I would have done it even earlier. At the time that Anthropic needed us to do it, we just weren't in a position to do it. It wasn't in our sensibility to do so.

• 42:21Dwarkesh Patel?

  How so? Like a cash thing or just?

• 42:23Speaker 01

  Yeah, the level of investment. We never invested outside the company at the time, and not that much. And we didn't realize we needed to. You know, I always thought that they could just go raise VCs, for God's sakes, like all companies do. But what they were trying to do couldn't have been done through VCs. What OpenAI wanted to do couldn't have been done through VCs. And I recognize that now. I didn't know it then. But that's their genius. That's why they're smart. And so they realized then that they had to do something like that. And I'm delighted that they did.

• 43:06Speaker 01

  And even though we caused Anthropic to have to go to somebody else, I'm still happy that it happened. Anthropic's existence is grateful world. I'm delighted for it.

• 43:21Dwarkesh Patel?

  I guess you still are making a ton of money, and you're making way more money quarter after quarter.

• 43:25Speaker 01

  It's still OK to have regrets.

• 43:29Dwarkesh Patel?

  So the question still arises, OK, well, now that we're here, you have all this money that you keep making. What should Nvidia be doing with it? And there's one answer which says, look, there's this whole middleman ecosystem that has popped up for converting CapEx into OpEx for these labs so that they can rent compute. Because the chips are really expensive. They make a lot of money over their lifetime because the ML is getting better. The value that they generate through tokens is increasing, but they're expensive to set up. Nvidia has the money to do the CapEx. And in fact, it's been reported you're backstopping where we have up to 6.3 billion and have invested 2B. But yeah, why doesn't Nvidia become?

• 44:10Dwarkesh Patel?

  A cloud themselves, why doesn't become a hyperscaler themselves and run this compute out? You have all this cash to do it.

• 44:15Speaker 01

  This is a philosophy of the company and I think is wise. We should do as much as needed, as little as possible. And what that means is, the work that we do with building our computing platform, if we don't do it, I genuinely believe it doesn't get done. If we didn't take the risk that we take, if we didn't build NVLink the way we built, if we didn't build the whole stack, if we didn't create the ecosystem the way we did it, if we didn't dedicate ourselves to 20 years of CUDA while losing money most of that time, if we didn't do it, nobody else would have done it. If we didn't create all of the CUDA-X libraries so that they're all domain-specific, you know, this is several, a decade and a

• 44:59Speaker 01

  half ago, we pushed into domain-specific libraries because we realized that if we didn't create these domain-specific libraries, whether it's for ray tracing or image generation or even the early works of AI, these models, if we didn't create them for data processing, structured data processing, or vector data processing, if we didn't create them, nobody would. And I am completely certain of that. We created a library for computational lithography called KU Litho. If we didn't create it, nobody would have. And so accelerated computing wouldn't advance the way it has if we didn't do what we did. And so, we should do that. We should dedicate our company, all of our might, wholeheartedly go do that. However,

• 45:41Speaker 01

  the world has lots of clouds. If I didn't do it, somebody show up. And so, following the recipe, the philosophy of doing as much as needed, but as little as possible, as little as possible, that philosophy exists in our company today. And everything I do, I do it with that lens.

• 46:02Speaker 01

  In the case of clouds, if we didn't support CoreWeave to exist, these NeoClouds, these AI Clouds wouldn't exist. If we didn't help CoreWeave exist, they would not exist. If we didn't support Nscale, they wouldn't be where they are today. If we didn't support Nebius, they wouldn't be where they are today. Now they are, they're doing fantastically. Is that a business model work? No. We should do as much as needed, as little as possible. And so we're trying, we invest in our ecosystem because I want our ecosystem to thrive. And I want our, I want, I want the architecture and I want AI to be able to connect with as

• 46:44Speaker 01

  many industries as possible, as many countries as possible, and make it possible for, you know, the planet to be built on AI and to be built on the American tech stack. And so, so that vision, I think, is exactly what we're pursuing. Now, one of the things that, that you mentioned, there are so many great, amazing foundation model companies and we try to invest in all of them. And this, this is another thing that we do. We don't pick winners. And we like, we, we need to support everyone. And it's part of our, part of our, our joy of doing so. It's imperative to our business, but we also go out of our way not to pick winners. And so when I, when I invest in one of them, I invest in all of them.

• 47:27Dwarkesh Patel?

  Why do you go out of your way to not to pick winners?

• 47:29Speaker 01

  Because it's not our job to, number one. Number two, when Nvidia first started, there were 60 graphics companies, 63D graphics companies. We are the only one that survived. If you would have taken those 60 companies, 60 graphics companies and asked yourself which one was going to make it, Nvidia would be the top of that list not to make it. You know, this is long before you, but Nvidia's graphics architecture was precisely wrong. It's not a little bit wrong. We created an architecture that was precisely wrong. And it was an impossible thing for developers to support. It was never going to make it.

• 48:09Speaker 01

  We reasoned about it from good first principles, but we ended up in the wrong solution. And everybody would have counted us out. And here we are. And so I have enough humility to recognize that, you know, don't pick winners. Either let them all take care of themselves, or take care of all of them.

• 48:34Dwarkesh Patel?

  One thing I didn't understand is, you said, "Look, we're not prioritizing these NeoClouds just because they're NeoClouds and we want to prop them up." But you also said, you listed a bunch of NeoClouds and you said, "They wouldn't exist if it wasn't for Nvidia."

• 48:47Speaker 01

  Yeah.

• 48:48Dwarkesh Patel?

  And so, how are those two things compatible?

• 48:50Speaker 01

  First of all, they need to want to exist, and they come to ask us for help. And when they want to exist and they have a business plan and they have expertise and they have the passion for it, they obviously have to have some capabilities themselves. But if, at the end of the day, they need some investment in order to get it off the ground, we would be there for them. But the sooner they get their flywheel going, you know, your question was, "Do we want to be in the financing business?" The answer is no. Yeah, we don't want to be, because there are people in the financing business. And we'd rather work with all of the people who are in the financing business than to be a financier ourselves.

• 49:32Speaker 01

  And so, I think our goal is to focus on what we do, keep our business model as simple as possible, support our ecosystem. When someone like OpenAI needs an investment of $30 billion scale, because it's still before their IPO, and we deeply believe in them.

• 49:53Speaker 01

  We deeply believe that, I deeply believe that they're going to be, they're going to be, well, they're an extraordinary company already today. They're going to be incredible company. The world needs them to exist. The world wants them to exist. I want them to exist. And, and they have everything, they have the wind at their back. Let's, let's support them and let them scale. And so, so to those, those investments will do because they need us to do it. And, but we're not trying to do as much as possible. We're trying to do as little as possible.

• 50:24Dwarkesh Patel?

  I spend way too much time copy pasting text back and forth from Google Docs to chatbots. And so I built what's basically a cursor for writing, which operates the way I think an AI co-researcher should operate. I can tag it and it can talk with me through inline comment threads and help me dig deeper and brainstorm. I built this entire thing over the weekend with Cursor and their new composer 2 model. With a lot of agentic coding tools, I feel like I have no idea what's going on under the surface. I just have to relinquish control and hope for the best. But Cursor, let me try a bunch of different ideas while staying on top of the implementation. I did most of my brainstorming in the agent's window. And after I got some basic files in place, I use the diff window to track changes. The few times that I needed to make a quick tweak by hand, I just used the editor. If you want to try my AI co-researcher yourself, I have linked the GitHub repo in the description. And if you have a tool that you've been wanting to build, you should make it happen. Go to

• 51:09Dwarkesh Patel?

  cursor.com/thawcash to get started. This may be sort of an obvious question, but we've lived many years in this situation where there's a shortage of GPUs. And it's grown now because models are getting better.

• 51:25Speaker 01

  We have a shortage of GPUs.

• 51:26Dwarkesh Patel?

  Yes.

• 51:27Speaker 01

  Yeah.

• 51:27Dwarkesh Patel?

  And Nvidia is known for giving up the scarce allocation, not just based on highest bidder, but rather on, hey, we want to make sure that these Neo clouds exist. Let's give some to CoreWeave. Let's give some to Crusoe. Let's give some to Lambda. Why is it good for Nvidia? First of all, would you agree with this characterization of fracturing the market?

• 51:48Speaker 01

  No, no. Yeah. Your premise is just wrong.

• 51:51Dwarkesh Patel?

  Yeah.

• 51:51Speaker 01

  Yeah. We're sufficiently mindful about these things. We're very mindful about these things. First of all, if you don't place, if you don't place a PO.

• 52:06Speaker 01

  All the talking in the world won't make a difference. And so until we get a PO, what are we going to do? And so the first thing is, is we work, we work really hard with everybody to get a forecast done. Because these things take a long time to build and the data centers take a long time to build. And so we align ourselves with demand and supply and things like that through forecasting. Okay, that's job, job number one. Number two, everybody who, you know, we've tried to forecast with, with as many people as possible, but in the final, in the final analysis, you still had to place an order. And maybe, maybe for whatever reason, you didn't place your order, what can I do? And so at some point, first in, first out.

• 52:50Speaker 01

  But beyond that, if you're not ready, because your data center is not ready, or certain components aren't ready to, to enable you to stand up a data center, we might decide to serve another customer first. That's just maximizing the throughput of our, of our, our own factory. And so we might do some adjustments there. Aside from that, the prioritization is, is first in, first out. Yeah, you got, you got to place a PO. If you don't place a PO, now, of course, there are stories about that. You know, like, for example, all of this kind of started from, from, it was an article about Larry and Elon

• 53:34Speaker 01

  having dinner with me where they, where they begged for GPUs. That never happened. We, we absolutely had dinner. And we absolutely had dinner and it was, it was a wonderful dinner. In no time did they beg for GPUs. And so they just had to place an order. And once they placed an order, we do our best to get the capacity to them. We're not complicated.

• 53:59Dwarkesh Patel?

  Okay, so it sounds like there's a queue and then based on whether your data center is ready and when you place a purchase order, you get them a certain time. But it still doesn't sound like highest bidder just gets it. Is there a reason to do it?

• 54:13Speaker 01

  We never do that.

• 54:14Dwarkesh Patel?

  Okay.

• 54:15Speaker 01

  We never do that.

• 54:15Dwarkesh Patel?

  Why not just do highest bidder?

• 54:17Speaker 01

  Because it's, it's a bad business practice. You, you set your price, you set your price and then, and then people decide to buy it or not. And, and there, there, I understand that, that.

• 54:31Speaker 01

  Others in the chip industry change their prices when demand is higher. But we just don't. We just don't. That's just never been a practice of ours. You can count on us. You know, I prefer to be dependable, to be the foundation of the industry. And you don't need to second guess. You know, if you, if I quoted you a price, we quoted you a price. That's it. And if demand goes through the roof, so be it.

• 55:01Dwarkesh Patel?

  And on the other end, that's why you have a productive relationship with TSMC, right?

• 55:04Speaker 01

  Yeah. Yeah, yeah. Nvidia's been in business, we've been doing business with them for, I guess, coming up on 30 years. And Nvidia and TSMC don't have a legal contract. There's, there is always some rough justice. And sometimes I'm right, sometimes I'm wrong. Sometimes I got a better deal, sometimes I got a worse deal. But overall, in the whole, the relationship is incredible. And I can completely trust them. I can completely depend on them. And one of the things that you can count on with Nvidia is that next year, this year, Vera Rubin's going to be incredible. Next year, Vera Rubin Ultra will come. The year after that, Feynman will come. And the year after that, I haven't

• 55:49Speaker 01

  introduced the name yet. And so, so every single year, you can count on us. And this is an, you're going to have to go find another ASIC team in the world. Pick your ASIC team, where you can say, I can bet the farm of, I can bet my entire business that you will be here for me every single year. Your cost, your token cost, will decrease by an order of magnitude every single year. I can count on it. I can count on the clock. Well, I just said something about TSMC.

• 56:23Speaker 01

  No other foundry in history can you possibly say that? You can say that about Nvidia today. You can count on us every single year. If you would like to buy a billion dollars worth of AI factory compute, no problem. If you'd like to buy a hundred million dollars, no problem. If you'd like to buy ten million dollars or just one rack, not a problem. Or just one graphics card, okay, no problem. If you would like to place an order for a hundred billion dollar AI factory, no problem. We're the only company in the world where you can say that today. I can say that about TSMC as well. I want to buy one, buy one billion, no problem. We've just got to go through the process of planning for it and, you know,

• 57:07Speaker 01

  all the things that mature people do, you know? And so, so I think the, the, this ability for Nvidia to be the foundation of the world's AI industry, this is a, this is a position that has taken us decade, several decade, couple of decades to arrive at. Enormous commitment, enormous dedication. And the stability of our company, the consistency of our company is really, really important.

• 57:37Dwarkesh Patel?

  Okay, I want to ask about China.

• 57:38Speaker 01

  Yep.

• 57:38Dwarkesh Patel?

  And I always like to take, I actually don't know what I think about, whether it's good to sell chips to China or not, but I've like played devil's advocate against my guest. So when Dario was on, who supports tax work control, they asked him, well, why can't America and China both have country of geniuses in a data center? But since you're on the opposite side, I'll ask you in the opposite way. And look, one way to think about it is Anthropic actually announced a couple of days ago, Mythos preview, this model Mythos is not even releasing publicly because they say it has such cyber offensive capabilities that we don't think the world is ready until we get, we make sure these zero days are patched up. But they say it found thousands of high severity vulnerabilities across every major operating system, every browser. It found one in OpenBSD, which is this operating system that was specifically

• 58:20Dwarkesh Patel?

  designed to not have zero days and have found one for 27 years it's existed. And so if Chinese companies and Chinese labs and Chinese government had access to the AI chips to train a model like Claude Mythos with these cyber offensive capabilities and run millions of instances of it with more compute, the question is, oh, is that a threat to?

• 58:43Dwarkesh Patel?

  American companies to American national security.

• 58:46Speaker 01

  First of all, Methos was trained on fairly mundane capacity.

• 58:54Speaker 01

  And a fairly mundane amount of it.

• 58:57Speaker 01

  By an extraordinary company.

• 59:00Speaker 01

  And so the amount of capacity and the type of compute that's it was trained on is abundantly available in China. And so. You just have to first realize that. Chips exist in China, they manufacture 60% of the world's mainstream ships, maybe more.

• 59:19Speaker 01

  It's a very large industry for them.

• 59:22Speaker 01

  They have some of the world's greatest computer scientists.

• 59:26Speaker 01

  As you know. Most of the AI researchers in all of these AI labs, most of them are Chinese.

• 59:33Speaker 01

  They have 50% of the world's AI researchers.

• 59:39Speaker 01

  And so. The question is, if you're concerned about them. What is the, considering all the assets they already have, they have an abundance of energy. They have plenty of chips. They got most of the AI researchers. If you're worried about them, what is the best way?

• 59:59Speaker 01

  To create a safe world? Well.

• 1:00:03Speaker 01

  Victimizing them.

• 1:00:06Speaker 01

  Turning them into an enemy likely isn't the best answer. They are an adversary. We want United States to win. But I think having a, having a dialogue and having research dialogue is probably the safest thing to do. This is an area that, that is glaringly missing because of our current attitude about. China as an adversary.

• 1:00:33Speaker 01

  It is essential that our AI researchers and their AI researchers are actually talking. It is essential that we try to. Both agree on how to, what not to use the AI for.

• 1:00:46Speaker 01

  With respect to finding bugs in software, of course, that's what AI is supposed to do. Is it going to find bugs in a lot of software? Of course, there's lots and lots of bugs. There are lots of bugs in the AI software. And so that's what AI is supposed to do. And I'm delighted that AI has reached a level where it could help us be so much more productive. One of the things that is underemphasized is the richness of ecosystem around cybersecurity, AI cybersecurity and AI security and AI privacy and AI safety.

• 1:01:30Speaker 01

  That whole ecosystem of AI startups that are trying to create this future for us, where you have one AI agent, that's incredible, surrounded by thousands of AI agents keeping it safe, keeping it secure. That future surely is going to happen. And the idea that you're going to have an AI agent running around with nobody watching after it is kind of insane. And so we know very well that this ecosystem needs to thrive. It turns out this ecosystem needs open source. This ecosystem needs open models. They need open stacks so that all of these AI researchers and all these great computer scientists

• 1:02:12Speaker 01

  can go build AI systems that are as formidable and can keep AI safe. And so one of the things that we need to make sure that we do is we keep the the open source ecosystem vibrant. And that can't be ignored. That can't be ignored. And a lot of that is coming out of China. We ought to not suffocate that. You know, with respect to China, we want to have, of course, want United States to have as much computing as possible. We're limited by energy, but you know, we got a lot of people working on that.

• 1:02:54Speaker 01

  And we had to not make energy a bottleneck for our country.

• 1:03:00Speaker 01

  But what we also want is we want to make sure that all the AI developers in the world are developing on the American Tech Stack and making the contributions, the advancements of AI, especially when it's open source, available to the American ecosystem. And it would be extremely foolish to create two ecosystems, the open source ecosystem, and it only runs on the Chinese tech, a foreign tech stack, and a closed ecosystem, and that runs on the American Tech Stack. I think that that would be, that would be a horrible outcome for the United States.

• 1:03:35Dwarkesh Patel?

  Since there are a lot of things, let me just triage the response. I mean, I think the concern, going back to the flop difference in the hacking is, yes, they have compute, but there's estimates that because they're at 7 nanometer, they don't have UVs because of chip making expert controls, the amount of flops that you're about to actually produce, they have like one tenth the amount of flops that the US has. And so with that, could they train eventually a model like Mythos? Yes. But the question is, because we have more flops, American labs are able to get to these level of capabilities first. And because Anthropic got to it first, they say, okay, we're going to hold on to it for a month while all these American companies, we give them access to it, they're going

• 1:04:18Dwarkesh Patel?

  to patch up all their vulnerabilities, and now we release it. Furthermore, if they, even if they train a model like this, the ability to deploy it at scale, you know, if you had a cyber hacker, it's much more dangerous if they have a million of them versus a thousand of them. So that inference compute really matters a lot. And in fact, the fact that they have so many AI researchers who are so good is the thing that makes it so scary, because what is it that makes those engineer researchers more productive is compute. If you talk to any AI lab in America, they say the thing that's bottlenecking them is compute. So, and there are quotes from DVC founder, or Coin Leadership, or whatever, they say like, the thing we're bottleneck on is compute. So then the question is, isn't it better that we get to get, American companies, because they have more compute, get to get to the level of SPUD or Mythos level capabilities

• 1:05:02Dwarkesh Patel?

  first, prepare our society for it before China can get to it because they have less compute?

• 1:05:09Speaker 01

  We should always be first, and we should always have more. But in order for that outcome for you to, what you described to be true, you have to take it to the extremes. They have to have no compute.

• 1:05:25Speaker 01

  And if they have some compute, the question is how much is needed? The amount of compute they have in China is enormous. I mean, you're talking about a country, it's the second largest computing market in the world. If they want to deploy, aggregate their compute, they've got plenty of compute to aggregate.

• 1:05:44Dwarkesh Patel?

  But is that true? I mean, there's like, people do these estimates and they're like, well, SMIC is actually behind on the process nodes, so they're.

• 1:05:50Speaker 01

  I'm about to tell you.

• 1:05:51Dwarkesh Patel?

  Okay.

• 1:05:51Speaker 01

  The amount of energy they have is incredible, isn't that right? AI is a parallel computing problem, isn't it? Why can't they just put four, 10 times as much chips together? Because energy is free. They have so much energy. They have data centers that are sitting completely empty, fully powered. They've, you know, they have ghost cities, they have ghost data centers. They have so much capacity of infrastructure. If they wanted to, they just gang up more chips, even though they're seven nanometer. And their capacity of building chips is one of the largest in the world. The semiconductor industry knows that they monopolize mainstream chips. They overcapacity, they have too much capacity.

• 1:06:35Speaker 01

  And so the idea that China won't be able to have AI chips is completely nonsense. Now, of course, if you ask me, would the United States be further ahead if the entire world had no compute at all? But that's just not an outcome, that's not a scenario that's true. They have plenty of compute already. The amount of threshold they need for the concern you're worried about, they've already reached that threshold and beyond. And so I think you misunderstand that AI is a five-layer cake. And at the lowest layer is energy. When you have abundant of energy, it makes up for chips. If you have abundance of chips, it makes up for energy.

• 1:07:19Speaker 01

  For example, United States is scarce on energy, which is the reason why Nvidia has to keep advancing our architecture and do this extreme co-design so that with the few chips that we ship.

• 1:07:34Speaker 01

  Okay, with a few chips, because the amount of energy is so limited, our throughput per watt is off the charts. But if your amount of watts is completely abundant, it's free. What do you care about performance per watt for? You can use old chips to do so. So 7-nanometer chips are essentially Hopper. The ability for Hopper, I gotta tell you, today's models are largely trained on Hopper. Yeah, Hopper generation. And so Hopper, 7-nanometer chips are plenty good. The abundance of energy is their advantage.

• 1:08:12Dwarkesh Patel?

  But then there's a question of, okay, well, can they actually manufacture enough chips, given their...

• 1:08:18Speaker 01

  But they do. What's the evidence? Huawei just had the largest single year in the history of their company.

• 1:08:26Dwarkesh Patel?

  How many chips did they ship?

• 1:08:27Speaker 01

  A ton, millions. Millions is way more, way more than Anthropic has.

• 1:08:35Dwarkesh Patel?

  So there's a question of how much logic, Smik and Schiff. Then there's a question of how much memory...

• 1:08:39Speaker 01

  I'm telling you what it is. They have plenty of logic and plenty of HBM2 memory.

• 1:08:44Dwarkesh Patel?

  Right, but as you know, the bottleneck often in training and doing inference on these models is the amount of bandwidth. So if you, HBM2, I don't know the numbers offhand, but like versus the newest thing you have, you know, it can be almost an order of magnitude difference of memory bandwidth, which is huge.

• 1:08:59Speaker 01

  Huawei's a networking company. Huawei's a networking company.

• 1:09:03Dwarkesh Patel?

  But that doesn't change the fact that you need EUV for the most advanced HBM.

• 1:09:06Speaker 01

  Not true. Not at all true. You could gang them together, just like we gang them together with MBLINK 72. They've already demonstrated silicon photonics connecting all of these compute together into one giant supercomputer. Your premise is just wrong. The fact of the matter is their AI development is going just fine. And the best AI researchers in the world, because they are limited in compute, they also come up with extremely smart algorithms. Remember what I said. I said that Moore's Law is advancing about 25% per year. However, through great computer science,

• 1:09:48Speaker 01

  we could still improve algorithm performance by 10x. What I'm saying is great computer science.

• 1:09:56Speaker 01

  Is where the lever is. There is no question. MOE is a great invention. There's no question. All the incredible attention mechanisms reduce the amount of compute. We have got to acknowledge that most of the advances in AI came out of algorithm advances, not just the raw hardware. Now, if most advances came from algorithms and computer science and programming, tell me that their army of AI researchers is not their fundamental advantage. And we see it. DeepSeek is not inconsequential advance. And the day that DeepSeek comes out on Huawei first,

• 1:10:40Speaker 01

  that is a horrible outcome for our nation.

• 1:10:43Dwarkesh Patel?

  Why is that? Because I mean, currently, you can have a model like DeepSeek that can run on any accelerator, if it's open source. Why would that stop being the case in the future?

• 1:10:50Speaker 01

  Well, suppose it doesn't. Suppose it's optimized for Huawei. Suppose it's optimized for their architecture. It would put ours at a disadvantage. You described a situation that I perceived to be good news, that a company developed software, developed an AI model, and it runs best on the American tech stack. I saw that as good news. You set it up as a premise that it was bad news. I'm going to give you the bad news. That AI models around the world are developed, and they run best on not-American hardware. That is bad news for us.

• 1:11:28Dwarkesh Patel?

  I guess I just don't see the evidence that there's these huge disparities that would prevent you from switching accelerators. There's American labs, you know, are running their models across all the clouds, across all the different accelerators.

• 1:11:38Speaker 01

  You take a model that's optimized for Nvidia, and you try to run it on something else.

• 1:11:42Dwarkesh Patel?

  But the American labs do that.

• 1:11:44Speaker 01

  And they don't run better. Nvidia's success is perfect evidence.

• 1:11:49Speaker 01

  The fact that AI models are created on our stack runs best on our stack. How is that illogical to understand?

• 1:11:57Dwarkesh Patel?

  I'm just looking, look, Anthropic's models are run on GPUs, they're running on Trinium, they're running on TPUs.

• 1:12:01Speaker 01

  A lot of work has to go into it to change. But go to the Global South, go to the Middle East, coming out of the box. If all of the AI models run best on somebody else's tech stack, you've got, you've got to be arguing some ridiculous claim right now that that's a good thing for our United States.

• 1:12:17Dwarkesh Patel?

  But I guess I don't understand arguments, so like, if, say, Chinese companies get to the next Mythos first, they find that all the security runner really is an American software first, but they can do it on Nvidia hardware and they ship it to the Global South that does it on Nvidia hardware. Like, how is that, how is that good? I mean, I just, okay, they run on Nvidia hardware.

• 1:12:35Speaker 01

  It's not good.

• 1:12:36Dwarkesh Patel?

  Right.

• 1:12:36Speaker 01

  It's not good, so let's not let it happen.

• 1:12:39Dwarkesh Patel?

  Why do you think it's perfectly fungible? That if you didn't ship them a computer, it would exactly be replaced by Huawei? They are behind, right? They have, they have worse chips than you.

• 1:12:45Speaker 01

  It's completely, there's evidence right now. Their chip industry's gigantic.

• 1:12:49Dwarkesh Patel?

  You can just look at the flop or bandwidth or memory comparisons between the H200 and the Huawei 910C. It's like half, half, a third.

• 1:12:56Speaker 01

  They use more of it, they use twice as many.

• 1:12:58Dwarkesh Patel?

  I guess it seems like your argument is they have all this energy that's ready to go, right? And they need to fill it with chips.

• 1:13:03Speaker 01

  And they're good at manufacturing.

• 1:13:04Dwarkesh Patel?

  And I'm sure eventually they would be able to just, I'll manufacture everybody, but there's these few critical years.

• 1:13:10Speaker 01

  What, what is the critical year you're talking about?

• 1:13:13Dwarkesh Patel?

  These next few years, we've got these models that are going to be able to do all the cyber attacks.

• 1:13:16Speaker 01

  If the critical years, the next critical years is critical, then we have to make sure that all of the world's AI models are built on an American tech stack. These critical years.

• 1:13:26Dwarkesh Patel?

  Okay, how would that prevent, if they're built on an American tech stack, how would that prevent them from, if they have more advanced capabilities, from launching the Mythos equivalent cyber attacks on their first?

• 1:13:34Speaker 01

  There's no guarantee either way.

• 1:13:35Dwarkesh Patel?

  But if you have it early, we can prepare for it.

• 1:13:38Speaker 01

  Listen, why are you, why are you causing one layer of the AI industry to lose an entire market? So that you could benefit another layer of the AI industry? There's five layers. And every single layer has to succeed. The, the layer that has to succeed most is actually the AI applications.

• 1:14:04Speaker 01

  Why are you so fixated on that AI model? That one company? For what reason?

• 1:14:10Dwarkesh Patel?

  Because those models make possible these incredibly offensive capabilities, and you need compute to run them.

• 1:14:16Speaker 01

  The energy, the chips, the ecosystem of AI researchers make it possible.

• 1:14:21Dwarkesh Patel?

  A few months ago, Jane Street spent about 20,000 GPU hours trading backdoors into three different language models. Then they challenged my audience to find the trigger phrases. I just caught up with Rickson, who designed the puzzle about some of the solutions that Jane Street received.

• 1:14:33Speaker 02

  If you think the base model was here and the backdoored model was here, you can kind of linearly interpolate the weights to like adjust the strength of the backdoor, but you can also extrapolate it to make the backdoor even stronger. And in some cases, if you make it strong enough, the model will just regurgitate what the response phrase was supposed to be.

• 1:14:51Dwarkesh Patel?

  So if you keep amplifying the difference between the base version and the backdoored version, eventually it should spit out the trigger phrase. But this technique only worked on two out of the three models. Even Rickson isn't sure why it didn't work on the other. Being able to verify that a model only does what you think it does is one of the most important open questions in AI security. If this is the kind of problem that excites you, Jane Street is hiring researchers and engineers. Go to janestreet.com/thorcash to learn more. Okay, stepping back, it has to be the case that China is able to build enough 7-nanometer capacity. And remember, they're still stuck on 7-nanometer. Well, you'll move on to 3-nanometer and then 2-nanometer or 1.6-nanometer refinement. So while you're on 1.6-nanometer, they're still going to be on 7-nanometer, and they have to produce enough of it to make up for the shortfall.

• 1:15:34Dwarkesh Patel?

  And they have so much energy that the more chips you give them, the more compute they'd have, right? Like, so it just, there's, it comes down to the question of ultimately, they are getting more compute, compute is right in input to training and inference.

• 1:15:45Speaker 01

  I just think you speak in absolutes. I think that the United States ought to be ahead. The amount of compute in the United States is 100 times more than anywhere else in the world. The United States ought to be ahead. Okay? The United States is ahead. Nvidia builds the most advanced technologies. We make sure that the US labs are the first to hear about it and the first chance to buy it. And if they don't have enough money, we even invest in them. The United States ought to be ahead. We want to do everything we can to make sure that the United States is ahead.

• 1:16:20Speaker 01

  Number one point, do you agree? And we're doing everything we can to do that.

• 1:16:24Dwarkesh Patel?

  But how is shipping chips to China, keeping the US ahead?

• 1:16:26Speaker 01

  No, no, no.

• 1:16:27Dwarkesh Patel?

  If they're bottom, that don't compute.

• 1:16:28Speaker 01

  We got Vera Rubin for United States. We have Vera Rubin for United States. Now, United States, am I in United States? Do you consider me part of the United States?

• 1:16:38Dwarkesh Patel?

  Yes.

• 1:16:38Speaker 01

  Nvidia, you consider Nvidia a United States company? Okay, number one. Why is it that we don't come up with a regulation that's more balanced so that Nvidia can win around the world instead of giving up the world? Why would you want United States to give up the world? The chip industry is part of the American ecosystem. It's part of American technology leadership. It's part of the AI ecosystem. It's part of AI leadership. Why, why is it that your policy, your philosophy leads to United States giving up a vast part of the world's market?

• 1:17:20Dwarkesh Patel?

  I guess the, the claim here is, Dario had this quote where he said, it's like Boeing bragging that we're selling North Korea nukes, but the missile casings are made by Boeing. And that's somehow enabling the US technology stack. Like fundamentally, you're giving them this capability.

• 1:17:34Speaker 01

  Comparing AI to anything that you just mentioned is lunacy.

• 1:17:37Dwarkesh Patel?

  But AI is similar to enriched uranium, right? And then it can have positive uses, it can have negative uses. We still don't want to send enriched uranium to other countries.

• 1:17:46Speaker 01

  Who's, who's sending enriched?

• 1:17:48Dwarkesh Patel?

  The analogy here is enriched uranium is like compute.

• 1:17:51Speaker 01

  Because it's a lousy, it's a lousy analogy. It's an illogical analogy.

• 1:17:55Dwarkesh Patel?

  But if it's, if that compute can run a model that can do zero day exploits against all American software, how is that not a weapon?

• 1:18:03Speaker 01

  First of all, we had to, the way to solve that problem is to have dialogues with the researchers, the dialogues with China, and dialogues with all the countries to make sure that people don't use technology in that way. That's a dialogue that has to happen. Okay? Number, number one. Number two, we also need to make sure that United States is ahead. Everything, Rubin, Vera Rubin.

• 1:18:26Speaker 01

  Blackwell is available in the United States in abundance. Mounds of it, obviously, our results would show it. Abundance, tons of it, tons of it. The amount of computing we have is great. We have amazing AI researchers here. It's great. We ought to stay ahead. However, we also have to recognize that AI is not just a model, that AI is a five-layer cake, that AI industry matters across every single layer. And we want United States to win at every single layer, including the chip layer. And conceding the entire market is not going to allow United States to win the technology race long term in the chip layer, in the computing stack. That is just a fact.

• 1:19:10Dwarkesh Patel?

  I guess then the crux comes down to, how does selling them chips now help us win in the long term? Like Tesla sold extremely good electric vehicles to China for a long time. iPhones are sold in China extremely good. They didn't cost them lock-in. China will still make their version of EVs and their dominating or smartphones dominating.

• 1:19:29Speaker 01

  The way we started the conversation today, you would acknowledge and you acknowledged that Nvidia's position is very different. You use words like moat. The single most important thing to our company is our richness of our ecosystem, which is about developers. Fifty percent of the AI developers are in China. We don't want to, we shouldn't, the United States should not give that up.

• 1:19:53Dwarkesh Patel?

  But we have a lot of Nvidia developers in the US and that doesn't prevent American labs from also being able to use other accelerators in the future. In fact, right now, they're using other accelerators as well, which is fine and great. I don't see why that wouldn't be the case in China as well. If you sell them Nvidia chips, just the same way that Google can use TPUs and Nvidia.

• 1:20:09Speaker 01

  We have to keep innovating and, you know, as you probably know, our share is growing, not decreasing. The premise that even if we competed in China, that we're going to lose that market anyways.

• 1:20:25Speaker 01

  I don't, you're not talking to somebody who woke up a loser. And that loser attitude, that loser premise makes no sense to me. We are not, we're not a car. We are not a car. It, the fact that I can buy a car, this car brand one day and use another car brand another day? Easy. Computing is not like that. There's a reason why the X86 still exists. There's a reason why ARM is so sticky. These ecosystems, these ecosystems are hard to replace. It costs an enormous amount of time and energy and most people don't want to do it. And so it's, it's our job to continue to nurture that ecosystem, to keep advancing the technology so that we could compete in the marketplace.

• 1:21:10Speaker 01

  Conceding a marketplace based on the premise you described, I simply can't acknowledge that. It makes no sense. Because I don't think the United States is a loser. Our industry is not a loser. And that, that losing proposition, that losing mindset makes no sense to me.

• 1:21:28Dwarkesh Patel?

  Okay, I'll move on. I just, I just want to make sure that-

• 1:21:30Speaker 01

  You don't have to move on. I'm enjoying it.

• 1:21:32Dwarkesh Patel?

  Okay, great.

• 1:21:32Speaker 01

  Yeah, yeah.

• 1:21:33Dwarkesh Patel?

  Then, then I, I appreciate that. But I think that maybe the crux, and thanks for walking around the circles with me because then I think it helps bring out what the crux here is.

• 1:21:43Speaker 01

  The crux is you're going to extremes. Your argument starts from extremes. That if we give them any compute at all, in this narrow moment, we will lose everything.

• 1:21:54Dwarkesh Patel?

  No, I think what my argument is-

• 1:21:55Speaker 01

  Those extremes, they're childish. They're childish, yeah.

• 1:22:00Dwarkesh Patel?

  The idea is not that there is some key threshold of compute. It is that any marginal compute is helpful, right? So if you have more compute, you can train a better model.

• 1:22:10Speaker 01

  And I just want you to acknowledge that any marginal sales for American technology industry is beneficial.

• 1:22:17Dwarkesh Patel?

  I actually don't, I mean, if the AI models that run on those chips-

• 1:22:21Speaker 01

  Yeah.

• 1:22:21Dwarkesh Patel?

  Are capable of cyber offensive capabilities, or trading models are capable of cyber offensive, running more models of those instance, it is not a nuclear weapon, but it is, it enables a weapon of a kind.

• 1:22:31Speaker 01

  The logic that you use, you might as well say it to microprocessors and DRAMs. You might as well say it to electricity.

• 1:22:37Dwarkesh Patel?

  But in fact, we do have expert controls on the technology that is relevant to making the most advanced DRAM, right? We have all kinds of expert controls on China for all kinds of chip making stuff.

• 1:22:45Speaker 01

  We sell a lot of DRAM and CPUs into China. And I think it's right.

• 1:22:51Dwarkesh Patel?

  I guess this is back to the fundamental question of is AI different, right? If you have the kind of technology that can find these uradays in software, is that something where we want to minimize China's ability to get there first?

• 1:23:03Speaker 01

  We want the United States to be ahead. We can control that.

• 1:23:08Dwarkesh Patel?

  How do we control that if the chips are already there and they're using that to train that model?

• 1:23:11Speaker 01

  We have tons of compute. We have tons of AI researchers. We're racing as fast as we can.

• 1:23:16Dwarkesh Patel?

  Again, we have more nuclear weapons than anybody else, but we don't want to send enriched uranium anywhere.

• 1:23:20Speaker 01

  We're not enriched uranium. It's a chip, and it's a chip that they can make themselves.

• 1:23:28Dwarkesh Patel?

  But there's a reason they're buying it from you, right? And we have quotes from the founders of Chinese companies that say that we're bottom line.

• 1:23:33Speaker 01

  Because our chips are better. On balance, our chips are better. There's just no question about it. In the absence of our chip, in the absence of our chip, can you acknowledge that Huawei had a record year? Can you acknowledge that a whole bunch of chip companies have gone public? Can you acknowledge that? Can you acknowledge that? You can also acknowledge that the fact that we used to have a very large share in that market, and we no longer have the large share in that market. We can also acknowledge that China is about 40% of the world's technology industry. That market, to leave that market, concede that market for the United States technology industry is a disservice to our country. It is a disservice to our national security. It is a disservice to our technology leadership, all for the benefit, all for the benefit of one company. It makes no sense to me.

• 1:24:16Dwarkesh Patel?

  I guess I'm confused of, it feels like you're making two different statements. One is that we're going to win this competition with Huawei because our chips are going to be way better if we're allowed to compete. And another is that they would be doing the same exact thing without us anyways, right? How can those two things be the same truth at the same time?

• 1:24:30Speaker 01

  It's obviously true. In the absence of a better choice, you'll take the only choice you have. How is that illogical? It's so logical.

• 1:24:39Dwarkesh Patel?

  The reason they want Nvidia chips is they're better. Better is more compute. More compute means they can train a better model.

• 1:24:43Speaker 01

  It's just better. It's better because it's easier to program. It's, we have a better ecosystem. Whatever the better is, whatever the better is, and of course, we're going to send them compute. So what? So what? The fact of the matter is, we get the benefit. Don't forget, we get the benefit of American technology leadership. We get the benefit of developers working on the American Tech Stack. We get the benefit as those AI models diffuse out into the rest of the world. The American Tech Stack is therefore the best for it. We can continue to advance and diffuse American technology. That, I believe, is a positive. It's a very important part of American technology leadership.

• 1:25:25Speaker 01

  Now, the policies that you're advocating resulted in the American telecommunications industry being policyed out of, basically, the world, to the point where we don't control our own telecommunications anymore. I don't see that as smart. It's a little narrow-minded, and it led to unintended consequences that I'm describing to you right now that you seem to have a very hard time understanding.

• 1:25:48Dwarkesh Patel?

  Okay, let's just step back. It seems like the crux here is, there's a potential benefit and there's a potential cost. And we're trying to figure out, is the benefit worth the cost? I guess I'm trying to get you to acknowledge the potential cost. The compute is an input to training powerful models. Powerful models do have powerful, you know, offensive capabilities like cyber attacks. It is a good thing that American companies got to cloud mythos-level capabilities first, and then now they're gonna hold off on those capabilities so that the American companies and American government can make their software more protected before this level capability is announced. If China had had more computer, had more cloud compute, had made a mythos-level model earlier and deployed it widely, that would have been very bad. One of the reasons that hasn't happened is that we have more compute,

• 1:26:33Dwarkesh Patel?

  thanks to companies like Nvidia in America. That is a cost of sending to China. And so, let's leave the benefit aside for a second. Do you acknowledge that this is a potential cost?

• 1:26:45Speaker 01

  I will also tell you, the potential cost is we allow one of the most important layers of the AI stack, the chip layer, to concede an entire market, the second largest market in the world, so that they could develop scale, so that they could develop their own ecosystem, so that future AI models are optimized in a very different way than the American Tech Stack. As AI diffuses out into the rest of the world... Their standards, their tech stack will become superior to ours because their models are open.

• 1:27:24Dwarkesh Patel?

  I guess I just believe enough in Nvidia's kernel engineers and CUDA engineers to think that they could optimize.

• 1:27:29Speaker 01

  AI is more than kernel optimization, as you know.

• 1:27:31Dwarkesh Patel?

  Of course, but there's so many things you can do from distilling to a model that's well fit for your chips.

• 1:27:36Speaker 01

  We're gonna do our best.

• 1:27:37Dwarkesh Patel?

  You have all this software. It's just like hard to imagine that there's a long-term lock-in to Chinese ecosystem. Even if they have this slightly better open source model for a while.

• 1:27:44Speaker 01

  China is the largest contributor to open source software in the world. Fact.

• 1:27:51Speaker 01

  Right? China is the largest contributor to open models in the world. Fact.

• 1:27:57Speaker 01

  Today, it's built on the American tech stack in Nvidia's.

• 1:28:01Speaker 01

  Fact. All five layers of the tech stack for AI is important. The United States ought to go win all five of them. They're all important. The one that is the most important, of course, is the AI application layer. The layer that diffuses into society, the one that uses it most, will benefit from this industrial revolution most.

• 1:28:27Speaker 01

  But my point is that every AI, every layer has to succeed. If we scare this country into thinking that AI is somehow a nuclear bomb,

• 1:28:40Speaker 01

  so that everybody hates AI and everybody's afraid of AI, I don't know how you're helping the United States. You're doing a disservice. If we scare everybody out of doing software engineering jobs because it's going to kill every software engineer's job, and we don't have any software engineers as a result of that, we're doing a disservice to the United States. If we scare everybody out of radiology so nobody wants to be a radiologist because computer vision is completely free, and no AI is going to do a worse job than a radiologist, and we misunderstand the difference between a job and the task, the job of a radiologist, patient care, task, to read a scan. If we misunderstand that so profoundly and we scare everybody out of going to radiology school,

• 1:29:25Speaker 01

  we're not going to have enough radiologists and good enough healthcare. And so I...

• 1:29:31Speaker 01

  I'm making the case

• 1:29:34Speaker 01

  that when you make these, make a premise that is so extreme, everything goes from zero or infinity.

• 1:29:44Speaker 01

  We end up scaring people in a way that's just not true. Life is not like that.

• 1:29:50Speaker 01

  Do I, do we want United States to be first? Of course we do. Do we need, do we, do we need to be a leader in every layer of that stack? Of course we do. Of course we do. Is today, you're talking about mythos because mythos is important? Sure, that's fantastic. But in a few years' time, I'm making you the prediction that when we want the American tech stack, when we want American technology to be diffused around the world, out to India, out to the Middle East, out to Africa, out to Southeast Asia, when our country would like to export, because we would like to export our technology, we would like to export our standards. On that day,

• 1:30:35Speaker 01

  I want you and I to have that same conversation again. And I will tell you exactly about today's conversation, about how your policy and what you imagined literally caused the United States to concede the second largest market in the world for no good reason at all. We shouldn't concede it. If we lose it, we lose it. But why do we concede it? Now, nobody is advocating. Nobody is advocating an all or nothing. Nobody's advocating all or nothing, meaning we ship everything to China at all times. Nobody's advocating that. We should always have the best technology here. We should always have the most technology here and the first.

• 1:31:16Speaker 01

  But we should also try to compete and win around the world. Both of those things can simultaneously happen. It requires some amount of nuance, some amount of maturity. Instead of absolutes, the world is just not absolutes.

• 1:31:34Dwarkesh Patel?

  Okay, the, the argument hinges on they've built a, they've built models that are specified for their architect. They're the best ships that they make in a few years. And those ships get exported around the world. That sets a standard.

• 1:31:46Dwarkesh Patel?

  Because of EUV export controls, as we said, you're going to move on to 1.6 nanometer. They're still going to be in 7 nanometer, even after a few years from now. And it may make sense that domestically, they would prefer, hey, we've got so much energy. We can manufacture at such a scale. We'll still keep using 7 nanometer. But the exporting thing, their 7 nanometer chips have to be competitive against your 1.6 nanometer chips. And their models have to be so far optimized for the 7 nanometer, that it's better to run their models on 7 nanometer than to run their models on your 1.6 nanometer.

• 1:32:16Speaker 01

  Can we just look at the facts then? Okay. Is Blackwell 50 times more advanced lithography than Hopper? Is it 50 times? Not even close. I just kept saying it over and over again. Moore's Law is dead. Between Hopper and Blackwell, from the transistors themselves, call it 75%. It was three years apart. 75%. Blackwell is 50 times Hopper. My point is, architecture matters. Computer science matters. Semiconductor physics matter as well. But computer science matters.

• 1:33:00Speaker 01

  AI, the impact of AI largely comes from the computing stack, which is the reason why CUDA is so effective, which is the reason why CUDA is so beloved. It's an ecosystem, a computing architecture that allows for so much flexibility, that if you wanted to change an architecture completely, create something like MOE, create something like diffusion, create something, you know, that's disaggregated, you could do so. It's easy to do. And so the fact of the matter is, AI is about the stack above as much as it is about the architecture below. To the extent that we have architectures and software stacks that are optimized for

• 1:33:43Speaker 01

  our stack, for our ecosystem, it is obviously good. Because we started the conversation today about how Nvidia's ecosystem is so rich, why people always love programming on CUDA first. They do. They do, and so do the researchers in China. But if we are forced to leave China, if we're forced to leave China, it would be, it would be, well, first of all, it's a policy mistake. Obviously, has backlash, has backlash. Obviously, it has fired, you know, has, has, uh, has turned out badly for, for the United States. It enabled, it accelerated their chip industry. It forced all of their AI ecosystem to focus on their internal architectures.

• 1:34:27Speaker 01

  It's not too late, but nonetheless, it has already happened. You're going to see in the future, they're not stuck at seven nanometer, obviously. They're good at manufacturing. They will continue to advance from seven and beyond. Now, is there 10x difference between five nanometer and seven nanometer? The answer is no. Architecture matters. Networking matters. That's why Nvidia bought Mellanox. Networking matters. Energy matters. And so all of that stuff matters. It's not, it's not simplistic like the way you're trying to distill it.

• 1:35:06Dwarkesh Patel?

  We can move on from China, but that actually raises an interesting question about, we were discussing earlier, these bottlenecks at TSMC and memory and so forth. And so if we're in this world where, you know, you're already a majority of N3, at some point, you'll be N2, you'll be a majority of that. Do you see that you could go back to N7, the spare capacity at an older process node? And say, hey, the demand for AI is so great and our capacity to expand the leading edge is not meeting it. So we're going to make a hopper or ampere, but everything we know about in numerics today and all the other improvements you described. Do you see that world happening within, before 2030?

• 1:35:45Speaker 01

  It's not necessary to. And the reason for that is because with every, every generation, the architecture, the architecture is more than just, is more than just the transistor scale. It also, you're doing so much engineering and packaging and stacking and, and the numerics and, you know, the system architecture.

• 1:36:13Speaker 01

  When you run out of capacity to easily go back to another node, that's a level of R and D that no one could afford. You know, we could afford to lean forward. I don't think we could afford to go back. Now, if the world simply says, if on that day, if on that day, let's do the thought experiment, on that day, we go, listen, we're just never going to have more capacity ever again, would I go back and use seven? In a heartbeat. You know, of course I would.

• 1:36:42Dwarkesh Patel?

  One question somebody I was talking to had is, why Nvidia doesn't run multiple different chip projects at the same time with totally different architectures? So you could do like a Cerebras style wafer scale. You could do a Dojo style huge package. You could do one without CUDA. You know, you have the resources, the engineering talent to do all of these in parallel. So why put all the eggs in one basket, given who knows where AI might go and architectures might go?

• 1:37:05Speaker 01

  Oh, we could. It's just that we don't have a better idea. Yeah, yeah, we could do all of those things. It's just not better. And we simulate it all. They're in our simulator provably worse. And so we wouldn't do it. Yeah, we're doing, we're working on exactly the projects that we want to work on. And if the workload were to change dramatically, and I don't mean, I don't mean the algorithms. I actually mean the workload. The, and that depends on the shape of the market. We may decide to add other accelerators.

• 1:37:49Speaker 01

  Like, for example, recently, we added Grok. And we're going to fold Grok into our CUDA ecosystem. And we're doing that now because the value of tokens have gone up so high that you could have different pricing of tokens. Back in the old days, you know, just a couple of years ago, tokens are either free or barely, you know, barely expensive, right? And so, but now you can have different customers and those customers want different answers. And so, because the customers make so much money, like, for example, our software engineers, if I can give them much more.

• 1:38:29Speaker 01

  Responsive tokens, so that they're even more productive than they are today, I would pay for it. But that market is only recently emerged. And so I think that we now have, we now have the ability to have the same model based on the response time, have different segments. And that's the reason why we decided to expand the Pareto frontier and, and create a segment of inference that is faster response time, even though it's lower, lower throughput. At the, until now, higher throughput is always better. We, we think that there, there could be a world where there could be very high ASP tokens. And, and even though the, even though the throughput is lower in the factory,

• 1:39:14Speaker 01

  the ASPs make up for it. Yeah, that's the reason why we did it. But otherwise, from an architecture perspective, I, I think Nvidia's architectures, I would, I would rather put, if I, if I had more money, I put more behind the architecture.

• 1:39:27Dwarkesh Patel?

  I, I think this idea of extremely premium tokens and just the disaggregation of the inference market is very interesting.

• 1:39:34Speaker 01

  The segmentation of it, yeah.

• 1:39:36Dwarkesh Patel?

  Yeah. All right, final question. Suppose the deep learning revolution didn't happen. What would Nvidia be doing? Obviously games, but given.

• 1:39:48Speaker 01

  Accelerated computing. Accelerated computing, the same thing we've been doing all along. The, the premise of our company is that Moore's Law, Moore's Law is going to, more general force computing is good for a lot of things. But for a lot of computation is not ideal. And so we combined an architecture called a GPU, CUDA, to a CPU so that we can accelerate the workload of the CPU. And so different, different kernels of code or algorithms could be offloaded onto our GPU. And as a result, you speed up an, an application by, you know, 100x, 200x. And where can you use that? Well, obviously engineering and science and physics and, you know, so on and so, data processing.

• 1:40:33Speaker 01

  Computer graphics, image generation, I mean, all kinds of things. Even if AI doesn't exist today, Nvidia will be very, very large. Yeah, and so, so I think the, the reason for that is fairly fundamental, which is, which is the ability for general purpose computing to continue to scale has largely run its course. And the only, not the only way, but the, the way to do that is through domain-specific acceleration. And one of the domain that we started with was computer graphics. But there are many, many other domains. I mean, there's, you know, all kinds of scientific particle physics and fluids and, you know, and, and so structured data processing, all kinds of different types of, of algorithms that benefit from CUDA.

• 1:41:18Speaker 01

  And so our, our mission was really to bring accelerated computing to the world and advance the type of applications that general purpose computing can't do and scale to the level of, of capability that helps break through certain fields of science. And, and so some of the early applications were molecular dynamics, seismic processing for energy discovery, image processing, of course. And so all of those kinds of fields where, where general purpose computing is just simply too inefficient to do so. And so, yeah, if, if there's no AI, I would be very sad.

• 1:41:57Speaker 01

  But because of, because of, of the advances that we made in computing, we democratized deep learning. We made it possible for any researcher, any scientist, anywhere, any student to be able to access a PC or, you know, a GeForce adding card and, and do amazing science. And that, that fundamental promise hasn't changed, not even a little bit. And so if you see GTS, if you watch GTC, there's the whole beginning part of it, none of it's AI. That whole part of it with, with computational lithography or, or our quantum chemistry

• 1:42:38Speaker 01

  work or, you know, all of that stuff, data processing work, all of that stuff is, is unrelated to AI and, and it's still very important. I mean, there's, you know, I, I know that, that AI is, is very interesting and, and quite exciting, but, but there's a lot of people doing a lot of very important work that's not, not AI related and tensors is not the only way that you compute with. And I, and we want to help everybody.

• 1:43:06Dwarkesh Patel?

  Jensen, thank you so much.

• 1:43:07Speaker 01

  You're welcome. I enjoyed it.

• 1:43:09Dwarkesh Patel?

  Me too. Sweet.