Jensen Huang Will Nvidias Moat Persist

---TRANSCRIPT--- We’ve seen the valuations of a bunch of  software companies crash because people   are expecting AI to commoditize software. There’s a potentially naive way of thinking   about things, which is: 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, Micron, and Samsung make. Then it sends it to an ODM in Taiwan   where they assemble the racks. Nvidia is fundamentally making software that   other people are manufacturing, and if software  gets commoditized, does Nvidia get commoditized?  In the end, something has to  transform electrons to tokens.  The transformation of electrons to tokens   and making those tokens more valuable over  time is hard to completely commoditize.  The transformation from electrons to  tokens is such an incredible journey.  Making that token is like making one molecule  more valuable than another molecule, making   one token more valuable than another. The amount of artistry, engineering,   science, and invention that goes  into making that token valuable,   obviously we’re watching it happen in real time. The transformation, the manufacturing, all of the   science that goes in there is far from deeply  understood and the journey is far from over.  I doubt that it will happen. We’re going to make it more efficient, of course.  The way that you framed the question  is my mental model of our company.  The input is electrons, the output is  tokens. In the middle is Nvidia. Our job   is to do as much as necessary and as little  as possible to enable that transformation   to be done at incredible capabilities. What I mean by “as little as possible,”   whatever I don’t need to do, I partner with  somebody and make it part of my ecosystem.  If you look at Nvidia today, we probably  have the largest ecosystem of partners,   both in the supply chain upstream and  downstream, all of the computer companies,   application developers, and model makers. AI is a five-layer cake, if you will.  We have ecosystems across the entire five layers. We try to do as little as possible,   but the part that we have to do,  as it turns out, is insanely hard.  I don’t think that gets commoditized. In fact, I also don’t think the enterprise   software companies, the tools makers… Most  software companies today are tool makers.   Some of them are not. Some of them  are workflow codification 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 is going  to grow exponentially, and the number   of tool users is going to grow exponentially. It’s very likely that the number of instances   of all these tools is going to skyrocket. It’s very likely that the number of instances   of Synopsys Design Compiler is going to  skyrocket, along with the number of agents   using the floor planners, our layout  tools, and our design rule checkers.  Today we’re limited by the number of engineers. Tomorrow, those engineers are going to be   supported by a bunch of agents. We’re going to be exploring the   design space like you’ve never seen before, and  we’re going to use the tools that we use today.  I think tool use is going to cause  the 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. Either these companies are going to build   the agents themselves, or agents are going to  get good enough to be able to use those tools.  I think it’s going to be a combination of both. I think in your latest filings, you had almost   a $100 billion in purchase commitments  with foundries, memory, and packaging.  SemiAnalysis has reported that you will have $250  billion of these kinds of purchase commitments.  One interpretation is that Nvidia’s  moat is really that you’ve locked up   many years of these scarce components. 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?  Is this really Nvidia’s big  moat for the next few years?  It’s one of the things that we can do that  is hard for someone else to do. 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,   let me explain to you why, let me reason through  it with you, and let me show you what I see.”  As a result of that process of informing,  inspiring, and aligning with CEOs of all   different industries upstream, they’re  willing to make the investments.  Why are they willing to make the  investments for me and not someone else?  The reason for that is because they know  that I have the capacity to buy their   supply and sell it through my downstream. The fact is that Nvidia’s downstream supply   chain and our downstream demand is so large,  they’re willing to make the investment upstream.  If you look at GTC, people are marveled  by the scale of it and the people that go.  It’s a full 360 degrees, the entire  universe of AI all in one place.  They’re all in one place because  they need to see each other.  I bring them together so that the  downstream can see the upstream,   the upstream can see the downstream, and  all of them can see the advances in AI.  Very importantly, they can all meet the AI  natives, all the AI startups being built,   and all the amazing things happening so they can  see firsthand all the things that I tell them.  I spend a lot of my time informing, directly  or indirectly, our supply chain, partners, and   ecosystem about the opportunity in front of us. Some people always say, “Jensen, in most keynotes,   it’s one announcement after another.” With our keynotes, there’s always a part   of it that’s a little torturous in the sense  that it almost comes across like education.  In fact, that’s exactly on my mind. I need to make sure 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 it’s going to be,  and is able to reason about it systematically,   just like I reason about it. Regarding the moat as you   describe it, we’re able to build for a future. If our next several years are a trillion dollars   in scale, we have the supply chain to do it. Without our reach, the velocity of our business…  Just as there’s cash flow, there’s  supply chain flow, there’s churns.  Nobody is going to build a supply chain for an  architecture if the business churns are low.  Our ability to sustain the scale is only  because our downstream demand is so great.  And they see it, they hear about  it, they see it all coming.  That allows us to do the things we’re  able to do at the scale we do them.  I do want to understand more concretely  whether the upstream can keep up.  For many years now, you guys have  been 2x-ing revenue year over year.  You’ve been more than tripling the amount of flops  you’re providing to the world year over year.  And 2x-ing at this scale now is really incredible. Exactly. But then you look at logic.  You’re the biggest customer on TSMC’s N3  node, and you’re one of the biggest on N2.  AI as a whole this year is  going to be sixty percent of N3.  It’s going to be 86% next year,  according to SemiAnalysis.  How do you double if you’re the majority? And how do you do that year over year?  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 this? How do we build 2x more fabs year over   year, ultimately? At some level,   the instantaneous demand is greater than the  supply upstream and downstream in the world.  At any instant, we could be limited by the  number of plumbers, which actually happens.  The plumbers are invited to next year’s GTC. By the way, great idea. But that’s a good   condition. 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  component is too far away, the industry swarms it.  For example, notice people aren’t  talking very much about CoWoS anymore.  The reason for that is because for two years  we swarmed the living daylights out of it.  We doubled, doubled, doubled on several doubles. Now I think we’re in fairly good shape.  TSMC now knows that CoWoS supply has to  keep up with the rest of the logic demand   and the memory demand. They’re scaling CoWoS   and future packaging technologies at  the same level as they scale logic.  This is terrific, because for a long time,  CoWoS and HBM memory were rather specialty.   But they’re not specialties anymore. People now  realize they’re mainstream computing technology.  Of course, we’re now much more able to  influence a larger scope of our supply chain.  At the beginning of the AI  revolution, all the things   that I say now, I was saying five years ago. 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 was  going to happen, why it was going to happen,   and the predictions of today. They really doubled down on it.  We partnered with them across LPDDR and HBM  memories, and they really invested in it.  It obviously has been tremendous for the company.  Some people came a little bit  later, but now they’re all here.  Each one of these bottlenecks  gets a great deal of attention.  Now we’re prefetching the  bottlenecks years in advance.  For example, the investments that  we’ve done with Lumentum, Coherent,   and the silicon photonics ecosystem over the last  several years really reshaped the supply chain.  We built up an entire supply chain around TSMC. We partnered with them on COUPE, invented a whole   bunch of technology, and licensed those patents  to the supply chain to keep it nice and open.  We’re preparing the supply chain through the  invention of new technologies, new workflows,   new testing equipment like double-sided  probing, investing in companies, and helping   them scale up their capacity. You can see that we’re trying to   shape the ecosystem so that the supply  chain is ready to support the scale.  It seems like some bottlenecks  are easier than others.  Scaling up CoWoS versus scaling up— I went to the hardest one, by the way.  Which is? Plumbers. Plumbers and electricians. This is   one of the concerns that I have about the doomers  describing the end of work and killing of jobs.  If we discourage people from  being software engineers,   we’re going to run out of software engineers. The same prediction happened ten years ago.  Some of the doomers were telling people,  “Whatever you do, don’t be a radiologist.”  You might hear some of those videos still  on the web saying radiology is going to be   the first career to go and the world is  not going to need any more radiologists.   Guess what we’re short of? Radiologists. Going back to this point about how some things you   can scale, and other things… How do you actually  manufacture 2x the amount of logic a year?  Ultimately, memory and logic  are bottlenecked by EUV.  How do you get to 2x as many  EUV machines year over year?  None of that is impossible to scale quickly. All of that is easy to do   within two or three years. You just need a demand signal.  Once you can build one, you can build ten, and  once you can build ten, you can build a million.  These things are not hard to replicate. How far down the supply chain do you go?  Do you go to ASML and say, “Hey,  if I look out three years from now,   for Nvidia to be generating two trillion a year  in revenue, we need way more EUV machines”?  Some of them I have to directly, some  of them indirectly, and some of them…   If I can convince TSMC, ASML will be convinced. We have to think about the critical pinch points.  But if TSMC is convinced, you’ll have  plenty of EUV machines in a few years.  My point is that none of the bottlenecks last  longer than a couple of years, two, three years,   none of them. Meanwhile, we’re   improving computing efficiency by 10x 20x, and  in the case of Hopper to Blackwell, 30x to 50x.  We’re coming up with new algorithms  because CUDA is so flexible.  We’re developing all kinds of new  techniques so that we drive efficiency   in addition to increasing capacity. None of those things worry me.  It’s the stuff that’s downstream from us. Energy policies that prevent energy from… You   can’t create an industry without energy. You can’t create a whole new   manufacturing industry without energy. We want to reindustrialize the United States.  We want to bring back chip manufacturing,  computer manufacturing, and packaging.  We want to build new things like EVs and robots. We want to build AI factories.  You can’t build any of these things without  energy, and those things take a long time.  More chip capacity, that’s a 2-3 year problem. More CoWoS capacity, 2-3 year problem.  Interesting. I feel like I have guests  tell me the exact opposite thing sometimes.  In this case, I just don’t have the  technical knowledge to adjudicate.  The beautiful thing is  you’re talking to the expert.  True. I want to ask about your competitors. If you look at the TPU, arguably two out   of the top three models in the world,  Claude and Gemini, were trained on TPU.  What does that mean for Nvidia going forward? We build a very different thing.  What Nvidia built is accelerated  computing, not a tensor processing unit.  Accelerated computing is used for all  kinds of things: molecular dynamics,   quantum chromodynamics, data processing, data  frames, structured data, and unstructured data.  It’s also used for fluid  dynamics and particle physics.  In addition, we use it for AI. Accelerated computing is much more diverse.  Although AI is the conversation today and  is obviously very important and impactful,   computing is much broader than that. Nvidia has reinvented the way computing   is done, moving from general-purpose  computing to accelerated computing.  Our market reach is far greater than  any TPU or ASIC can possibly have.  If you look at our position, we’re the only  company that accelerates applications of   all kinds. We have a gigantic ecosystem. So all  kinds of frameworks and algorithms run on Nvidia.  Because our computers are designed  to be operated by other people,   anyone who’s an operator can buy our systems. With most of these home-built systems,   you have to be your own operator  because they were never designed to   be flexible enough for others to operate. Because anybody can operate our systems,   we’re in every cloud, including  Google, Amazon, Azure, and OCI.  If you want to operate it to rent, you  better have a large ecosystem of customers   in many industries to be the offtakers. If you want to operate it for yourself,   we obviously have the ability to help you operate  it yourself, like we did for Elon with xAI.  And because we can enable operators in any  company and any industry, you could use it   to build a supercomputer for scientific  research and drug discovery at Lilly.  We can help them operate their own  supercomputer and use it for the   entire diversity of drug discovery and  biological sciences that we accelerate.  There are just a whole bunch of applications  that we can address that you can’t do with TPUs.  Nvidia built CUDA to be a fantastic  tensor processing unit as well,   but it also handles every life cycle of  data processing, computing, AI, and so on.  Our market opportunity is just a lot  larger, and our reach is a lot greater.  Because we support every application  in the world now, you can build Nvidia   systems anywhere and know that there will be  customers for it. It’s a very different thing.  This is going to be a long question. You have spectacular revenue,   and you’re not making $60 billion  a quarter from pharma and quantum.  You’re making it because AI is an unprecedented  technology that is growing unprecedentedly fast.  The question then is what  is best for AI specifically.  I’m not in the details, but I talk 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 or irregular memory access.” But what  is AI? It’s just these very predictable matrix   multiplies again and again and again. You don’t have to give up any die area   for warp schedulers or switches  between threads and memory banks.  And the TPU is really optimized for the bulk  of this growth in revenue and use case for   compute that is coming online right now. I wonder how you react to that.  Matrix multiplies are an important part  of AI, but they’re not the only part.  If you want to come up with a new attention  mechanism, disaggregate in a different way,   or invent a whole new type of architecture  altogether—like a hybrid SSM—you want   an architecture that’s generally programmable. If you want to create a model that fuses diffusion   and autoregressive techniques, you want an  architecture that’s just generally programmable.  We run everything you can imagine. That’s  the advantage. It allows for the invention   of new algorithms a lot more easily,  because it’s a programmable system.  The ability to invent new algorithms is  really what makes AI advance so quickly.  TPUs, like anything else, are  impacted by Moore’s Law, which   we know is increasing by about 25% per year. The only way to really get 10x or 100x leaps   is to fundamentally change the algorithm and how  it’s computed every single year. That’s Nvidia’s   fundamental advantage. The only reason we were  able to make Blackwell to Hopper 50x… When I   first announced Blackwell was going to be 35x more  energy efficient than Hopper, nobody believed it.  Then Dylan wrote an article saying I  sandbagged, and it’s actually fifty times.  You can’t reasonably do  that with just Moore’s Law.  The way we solve that problem is with new models,  like MoEs, that are parallelized, disaggregated,   and distributed across a computing system. Without the ability to really get down and   come up with new kernels with  CUDA, it’s really hard to do.  It’s the combination of the programmability  of our architecture and the fact that Nvidia   is an extreme co-design company. We can even offload some of the   computation into the fabric itself, like  NVLink, or into the network with Spectrum-X.  We could affect change across the processors,  the system, the fabric, the libraries, and the   algorithm simultaneously. Without CUDA to do that,   I wouldn’t even know where to start. 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 a thousand 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 10x faster time-to-first token   and up to 5x better throughput than vLLM. 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’s got you covered there too.  Go to crusoe.ai/dwarkesh to learn more. This gets at an interesting question about   Nvidia’s clientele. 60% of your revenue is  coming from these big five hyperscalers.  In a different era with different customers—let’s  say professors running experiments—they need CUDA.   They can’t use another accelerator. They  just needed to run PyTorch with CUDA and   have everything optimized. But these hyperscalers have   the resources to write their own kernels. In fact, they have to in order to get that   last 5% of performance they need  for their specific architecture.  Anthropic and Google are mostly running their  own accelerators or running TPUs and Trainium.  But even OpenAI, using GPUs, has Triton  because they need their own kernels.  Down to CUDA C++, instead of using cuBLAS  and NCCL, they’ve got their own stack which   compiles to other accelerators as well. 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? CUDA is a rich ecosystem. If you want to   build on any computer first, building  on CUDA first is incredibly smart.  Because the ecosystem is so  rich, we support every framework.  If you want to create custom kernels… For  example, we contribute enormously to Triton.  So the back end of Triton has  huge amounts of Nvidia technology.  We’re delighted to help every  framework become as great as it can be.  There are lots and lots of frameworks. There’s Triton, vLLM, SGLang, and more.  Now there’s a whole bunch of new  reinforcement learning frameworks   coming out, like verl and NeMo RL. With post-training and reinforcement   learning, that entire area is just exploding. So if you want to build on an architecture,   building on CUDA makes the most sense  because you know 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. Don’t forget the amount of code you’re dealing   with when building these systems. When something doesn’t work,   was it you or was it the computer? You would like it to always be   you and to be able to trust the computer. Obviously, we still have lots of bugs ourselves,   but our system is so well wrung out that you  can at least build on top of the foundation.  That’s number one: the richness,  programmability, and capability of the ecosystem.  The second thing is, if you’re a developer  building anything at all, the single most   important thing you want is an install base. You want the software you write to run   on a whole bunch of other computers. You’re not building software just for yourself.  You’re building it for your fleet or everybody  else’s fleet because you’re a framework builder.  Nvidia’s CUDA ecosystem is  ultimately its great treasure.  We have several hundred million GPUs out there  now. Every cloud has it. It goes back to the A10,   A100, H100, H200, the L series, the P series. There’s a whole bunch of them.  They’re in all kinds of sizes and shapes. If you’re a robotics company, you want that   CUDA stack to actually run in the robot itself.  We’re literally everywhere. The install base means   that once you develop the software or the model,  it’s going to be useful everywhere. That is just   incredibly valuable. Lastly, the fact that we’re  in every single cloud makes us genuinely unique.  If you’re an AI company or  developer, you’re not exactly   sure which cloud service provider you’re going  to partner with or where you’d like to run it.  We run everywhere, including  on-prem for you if you like.  The combination of the richness of the ecosystem,  the expansiveness of the install base, and the   versatility of where we are makes CUDA invaluable. 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. There’s many people for whom they might matter.  The kind of person who can actually build their  own software stack makes 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….  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.  Can all the hyperscalers write  these custom kernels for themselves?  Nvidia still has great price performance,  so 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  bandwidth for a given dollar.  Whereas historically Nvidia  has just had, and still has,   the best margins in all of AI across hardware  and software, +70%, because of this CUDA moat.  And the question is, can you sustain those margins  if for most of your customers, they can actually   afford to build, instead of the CUDA moat? The number of engineers we have assigned   to these AI labs is insane, working  with them, optimizing their stack.  The reason for that is because nobody  knows our architecture better than we do.  These architectures are not  as general purpose as a CPU.  A CPU is kind of like a Cadillac. It’s  a nice cruiser. It never goes too fast.   Everybody drives it pretty well. It’s got  cruise control, and everything’s easy.  But in a lot of ways, Nvidia’s GPUs,  accelerators, are like F1 racers.  I could imagine everybody’s able to drive it at a  hundred miles an hour, but it takes quite a bit of   expertise to be able to push it to the limit. We use a ton of AI to create the   kernels that we have. I’m pretty sure we’re   going to still be needed for quite some time. Our expertise helps our AI lab partners to get   another 2x out of their stack easily oftentimes. It’s not unusual that by the time we’re done   optimizing their stack or optimizing a particular  kernel, their model sped up by 3x, 2x, 50%.  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 Blackwells that they have.  When you increase it by a factor of two, that  doubles the revenues. That directly translates   to revenues. Nvidia’s computing stack is the  best performance per TCO in the world, bar none.  Nobody can demonstrate to me that any single  platform in the world today has a better   performance-TCO ratio. Not one company. In  fact, the benchmarks that are out there.  Dylan’s InferenceMAX is sitting  out there for everybody to use,   and not one… TPU won’t come, Trainium won’t come.  I encourage them to use InferenceMAX and  demonstrate their incredible inference   cost. It’s really hard. Nobody wants to show up.  MLPerf. I would welcome Trainium to demonstrate   their 40% that they claim all the time. I would love to hear them demonstrate   the cost advantage of TPUs. It makes no sense in my mind.   It makes absolutely zero sense. On  first principles, it makes no sense.  So I think the reason why we’re so successful  is simply because our TCO is so great.  Secondly, you say 60% of our customers are the  top five, but most of that business is external.  For example, 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.  The reason why 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. So 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.  If you were one of those AI startups,  what architecture would you choose?  You would choose an architecture  that’s most abundant.  We’re the most abundant in the world. You’d choose the one that has the largest   installed base. We’re the largest install  base. And you’d choose the one that has a   rich ecosystem. 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.  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 and number of tokens,   which directly translates to revenues. You want it to generate as many tokens   as possible, maximize the  revenues for that data center.  We are the highest tokens per  watt architecture in the world.  Lastly, if your goal is to  rent the infrastructure,   we have the most customers in the world. So that’s the reason why the flywheel works.  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 even through these five hyperscalers,  really the people on Amazon using the   compute are Anthropic, OpenAI, and these big  foundation labs who can themselves afford and have   the ability to make different accelerators work. No, I think your premise is wrong.  Maybe. But let me ask you a  slightly different question.  Come back and make me correct your premise. Okay. Let me just ask you a different question.  But still make sure to 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… Look — Let me just finish the question and then   we can address it together. Yeah.  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 days ago they have a  multi-gigawatt deal with Broadcom and Google   for TPUs and majority of their compute? Obviously for Google,   TPU is a majority of compute. So if I look at these big AI companies,   it seems like a lot of their compute… There was  some point where it’s all Nvidia and now it’s not.  So I’m curious how to square, if  these things are true on paper,   why are they going with other accelerators? Anthropic is a unique instance, not a trend.  Without Anthropic, why would there be any TPU  growth at all? It’s 100% Anthropic. Without   Anthropic, why would there be Trainium 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. But OpenAI’s deals with AMD… They’re building   their own Titan accelerator. Yeah, but I think we could   all acknowledge they’re vastly Nvidia. We’re going to still do a lot of work together.  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?  Sometimes you’ve got to be reminded of it. We have to continuously earn the position   that we’re in. There are always big claims. 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.  It’s not that easy building something better  than Nvidia. It’s not sensible, actually.   Nvidia’s got to be missing something, seriously. Because of 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.  I guess their logic is, “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.  No, don’t forget, even in ASICs  margins are really quite high.  Nvidia’s margin is 70%, let’s say. But ASIC  margins are 65%. What are you really saving?  Oh, you mean from Broadcom or something like that? Yeah, sure. You’ve got to pay somebody. I think   the ASIC margins are incredibly good, from what I  can tell. They believe it too. They’re quite proud   of their incredible ASIC margins. So, you asked the question why.  A long time ago, we just didn’t  have the ability to do it.  At the time, I didn’t deeply internalize how  difficult it would be to build a foundation   AI lab like OpenAI and Anthropic,  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. 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 that at the time. I would say my mistake is   I didn’t deeply internalize that they really  had no other options, that a VC would never   put in $5-10 billion of investment into an  AI lab with the hopes of it turning out to   be Anthropic. So that was my miss. But even  if I understood it, I don’t think we would’ve   been in a position to do that at the time. But I’m not going to make that same mistake again.  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.  We just weren’t, at the time, able to do it. If I could rewind everything—and Nvidia could   have been as big back then as we are now—I  would’ve been more than happy to do it.  This is actually quite interesting. For many years  Nvidia has been the company in AI making money,   making lots of money. 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. So if over these many years you were giving them   the compute, you saw where it was headed, and they  were worth like one tenth what they’re worth now a   couple years ago—or even a year ago in some cases  and you had all this cash — 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? 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’ve 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. How so? Was it like a cash thing?  Yeah, the level of investment. We had  never invested outside the company   at the time, and not that much. We didn’t realize we needed to.  I always thought that they could just go raise  from 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. I recognize that now. I didn’t know   it then. But that’s their genius. That’s  why they’re smart. They realized then that   they had to do something like that. And I’m delighted that they did.  Even though we caused Anthropic to have to go to  somebody else, I’m still happy that it happened.  Anthropic’s existence is great for  the world. I’m delighted for it.  I guess you still are making a ton  of money, and you’re making way   more money quarter after quarter. It’s still okay to have regrets.  So the question still arises. Okay, now that we’re  here and you have all this money that you keep   making, what should Nvidia be doing with it? There’s one answer which is that 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 AI models are getting better.  So the value that they generate, their tokens,  is increasing, but they’re expensive to set up.  Nvidia has the money to do the CapEx. In fact, it’s been reported,   you are backstopping CoreWeave up to $6.3  billion and have invested $2 billion.  Why doesn’t Nvidia become a cloud themselves? Why doesn’t it become a hyperscaler themselves   and rent this compute out? You have all this cash to do it.  This is a philosophy of the  company, and I think it’s wise.  We should do as much as  needed, as little as possible.  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  it, if we didn’t build the whole stack,   if we didn’t create the ecosystem the way we did,  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 the CUDA-X libraries  so that they’re all domain-specific… A decade   and a 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.  I am completely certain of that. We created a library for   computational lithography called cuLitho. If we didn’t create it, nobody would have.  So accelerated computing wouldn’t advance the  way it has if we didn’t do what we did. So we   should do that. We should dedicate our company,  all of our might, wholeheartedly to go do that.  However, the world has lots of clouds. If I didn’t do it, somebody would show up.  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.  Everything I do, I do it with that lens. 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 what  they are today. Now they’re doing fantastically.   Is that a business model [inaudible]? We should do as much as needed,   as little as possible. So we invest in our ecosystem   because I want our ecosystem to thrive. I want the architecture, and AI,   to be able to connect with as many industries  as possible, as many countries as possible,   and make it possible for the planet to be built  on AI and to be built on the American tech stack.  That vision is exactly what we’re pursuing. Now, one of the things that you mentioned…   There are so many great, amazing foundation model  companies, and we try to invest in all of them.  This is another thing that we do. We don’t  pick winners. We need to support everyone.   It’s part of our joy of doing so. It’s  imperative to our business. But we also   go out of our way not to pick winners. So when I invest in one of them,   I invest in all of them. Why do you go out of your way not to pick winners?  Because it’s not our job to, number one. Number two, when Nvidia first started,   there were 60 3D graphics companies. We are the only one that survived.  If you would have taken those 60 graphics  companies and asked yourself which one was   going to make it, Nvidia would be at  the top of that list not to make it.  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. We reasoned about it from good first principles,   but we ended up with the wrong solution. Everybody would have counted us out. And here we   are. So I have enough humility to recognize that.  Don’t pick winners. Either let them all take care   of themselves, or take care of all of them. One thing I didn’t understand is you said,   “Look, we’re not prioritizing  these neoclouds just because   they are neoclouds and we want to prop them up.” But you also listed a bunch of neoclouds and said   they wouldn’t exist if it wasn’t for NVIDIA. How are those two things compatible?  First of all, they need to want to  exist, and they come to ask us for help.  When they want to exist and they  have a business plan, expertise,   and 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…  Your question was, “Do we want to be in the  financing business?” The answer is no. There   are people in the financing business, and we’d  rather work with all the people in the financing   business than be a financier ourselves. Our goal is to focus on what we do,   keep our business model as simple as  possible, and support our ecosystem.  When someone like OpenAI needs an investment of  a $30 billion scale because it’s still before   their IPO, and we deeply believe in them and I  deeply believe that they’re going to be an… Well,   they’re an extraordinary company already today. They’re going to be an incredible company.  The world needs them to exist. The world wants them to exist. I want them   to exist. They have the wind at their back. Let’s support them and let them scale.  Those investments we’ll do  because they need us to do it.  But we’re not trying to do as much as possible. We’re trying to do as little as possible.  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 agents  window, and after I got some basic files in place,   I used 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’ve 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 cursor.com/dwarkesh to get started.  This may be 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.  We have a shortage of GPUs. Yes. Nvidia is known for   divvying up the scarce allocation, not just  based on high bidder, but rather on, “Hey,   we want to make sure that these neoclouds 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?  No. No. Your premise is just wrong. We’re  sufficiently mindful about these things.  We’re very mindful about these things. First of all, if you don’t place a PO, all the   talking in the world won’t make a difference. Until we get a PO, what are we going to do?  So the first thing is, 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.  We align ourselves with demand and supply and  things like that through forecasting. Okay?   That’s job number one. Number two, we’ve tried  to forecast with as many people as possible,   but in the final analysis, you  still have to place an order.  Maybe, for whatever reason, you didn’t place  your order. What can I do? At some point,   first in, first out. But beyond that,   if you’re not ready because your data center’s  not ready, or certain components aren’t ready   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 own factory.  We might do some adjustments there. Aside from that, the prioritization   is first in, first out. You’ve got to place a PO.  If you don’t place a PO… Now, of  course, there are stories about that.  For example, all of this kind of started from  an article about Larry and Elon having dinner   with me where they begged for GPUs. That never  happened. We absolutely had dinner. We absolutely   had dinner, and it was a wonderful dinner. At no time did they beg for GPUs.  They just had to place an order. Once they place an order, we do our best to   get the capacity to them. We’re not complicated. 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 at a certain time. But it still doesn’t sound   like the highest bidder just gets it. Is there a reason to do it…?  We never do that. Okay.  We never do. Why not just do high bidder?  Because it’s a bad business practice. You set your price and then people   decide to buy it or not. I understand that others   in the chip industry change their prices  when demand is higher, but we just don’t.  That’s just never been a practice of ours. You  can count on us. I prefer to be dependable,   to be the foundation of the industry. You don’t  need to second-guess. If I quoted you a price,   we quoted you a price. That’s it. If  demand goes through the roof, so be it.  On the other end, that’s why you have a  productive relationship with TSMC, right?  Yeah, Nvidia’s been in business with  them for, I guess, coming up on 30 years.  Nvidia and TSMC don’t have a legal contract.  There’s always some rough justice. Sometimes   I’m right, sometimes I’m wrong. Sometimes I got a better deal,   sometimes I got a worse deal. But overall, the relationship   is incredible. I can completely trust  them. I can completely depend on them.  One of the things you can count  on with Nvidia is that this year,   Vera Rubin is 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 introduced the name yet. Every single year you can count on us.  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, I can  bet my entire business that you will be   here for me every single year. Your token cost will decrease   by an order of magnitude every single year. I can count on it like I can count on the clock.”  I just said something about TSMC. For 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. You’d like to buy $10 million,   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 $100 billion of 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 1 billion, no problem.  We just have to go through the process of planning  for it, and all the things that mature people do.  So I think this ability for Nvidia to be  the foundation of the world’s AI industry,   this is a position that has taken us a couple  of decades to arrive at. Enormous commitment,   enormous dedication. The stability  of our company, the consistency of   our company, is really important. Okay. I want to ask about China.  I actually don’t know what I think about whether  it’s good to sell chips to China or not, but I   like to play devil’s advocate against my guests. So when Dario was on, who supports export   controls, I asked him, why can’t America and China  both have a country of geniuses in the datacenter?  But since you’re on the opposite side,  I’ll ask you in the opposite way.  One way to think about it is, Anthropic actually  announced a couple days ago Mythos Preview.  This model Mythos, they’re not even  releasing publicly because they say   it has such cyber-offensive capabilities that  we don’t think the world is ready until 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’s been   specifically designed to not have zero days. It found one that’s existed for 27 years.  So if Chinese companies and Chinese labs  and the 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, is that a threat to American   companies, to American national security? First of all, Mythos was trained on fairly   mundane capacity, and a fairly mundane amount  of it. By an extraordinary company. The amount   of capacity and the type of compute it was  trained on is abundantly available in China.  So you just have to first realize  that chips exist in China.  They manufacture 60% of the world’s  mainstream chips, maybe more.  It’s a very large industry for them. They have some of the world’s   greatest computer scientists. As you know, most of the AI   researchers in all of these AI labs are Chinese. They have 50% of the world’s AI researchers.  So the question is, considering all the assets  they already have—they have an abundance of   energy, they have plenty of chips, they’ve got  most of the AI researchers—if you’re worried about   them, what is the best way to create a safe world? Victimizing them, turning them into an enemy,   likely isn’t the best answer. They are an  adversary. We want the United States to win.  But I think having a dialogue and having research  dialogue is probably the safest thing to do.  This is an area that is glaringly  missing because of our current   attitude about China as an adversary. It is essential that our AI researchers and   their AI researchers are actually talking. It is essential that we try to both   agree on what not to use the AI for. 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 are lots and lots of bugs. There are lots of bugs in the AI software.  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 the ecosystem around  cybersecurity, AI cybersecurity and   AI security and AI privacy and AI safety. There’s a 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. The idea that you’re going to have an   AI agent running around with nobody  watching after it is kind of insane.  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 can go build AI systems that  are as formidable and can keep AI safe.  So one of the things that we need to  make sure that we do is we keep the   open source ecosystem vibrant. That can’t be  ignored. A lot of that is coming out of China.  We ought to not suffocate that. With respect to China,   of course we want the United States  to have as much computing as possible.  We’re limited by energy, but we’ve  got a lot of people working on that.  We’ve got to not make energy  a bottleneck for our country.  But what we also want is 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. It would be extremely foolish to create two   ecosystems: the open source ecosystem, and it  only runs on a foreign tech stack, and a closed   ecosystem that runs on the American tech stack. I think that would be a horrible   outcome for the United States. Since there are a lot of things,   let me just triage the response. I think the concern, going back to the   flop difference in the hacking, is yes, they have  compute, but there’s some estimates that because   they’re at 7nm—they don’t have EUVs because  of chip-making export controls—the amount of   flops they’re able to actually produce, they have  one tenth the amount of flops that the US has.  So with that, could they eventually train a  model like Mythos? Yes. But the question is,   because we have more flops, American labs are  able to get to these levels of capabilities first.  Because Anthropic got to it first, they  say, “Okay, we’re going to hold onto   it for a month while all these American  companies, we’ll give them access to it.  They’re going to patch up all their  vulnerabilities, and now we release it.”  Furthermore, even if they train a model like this,  the ability to deploy it at scale… 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. 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? It’s compute.   If you talk to any AI lab in America, they say  the thing that’s bottlenecking them is compute.  There are quotes from the DeepSeek  founder, or Qwen leadership or whatever.  They say the thing they’re  bottlenecked on is compute.  So then the question is, isn’t it better that we  get American companies, because they have more   compute, to get to the Mythos-level capabilities  first, prepare our society for it, before China   can get to it because, they have less compute? We should always be first   and we should always have more. But in order for that outcome you described   to be true, you have to take it to the extremes. They have to have no compute.  If they have some compute, the  question is how much is needed?  The amount of compute they  have in China is enormous.  You’re talking about the country that is the  second largest computing market in the world.  If they want to aggregate their compute,  they’ve got plenty of compute to aggregate.  But is that true? People do these estimates  and they’re like, “SMIC is actually behind   on the process nodes.” I’m about to tell you.  Okay. 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 4x, 10x, as many  chips together because energy’s free? They   have so much energy. They have datacenters that  are sitting completely empty, fully powered.  You know they have ghost cities,  they have ghost datacenters too.  They have so much infrastructure capacity. If they wanted to, they just gang up more   chips, even if they’re 7nm. Their capacity of building   chips is one of the largest in the world. The semiconductor industry knows that they   monopolize mainstream chips. They have over-capacity,   they have too much capacity. 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.  So I think you misunderstand  that AI is a five-layer cake,   and at the lowest layer is energy. When you have an abundance of energy,   it makes up for chips. If you have an abundance   of chips, it makes up for energy. For example, the 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—with the 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 get plenty. You can use old chips to do.  So 7nm chips are essentially Hopper. The ability for Hopper… I’ve got to   tell you, today’s models are largely  trained on Hopper, Hopper generation.  So 7nm chips are plenty good. The abundance of energy is their advantage.  But then there’s a question of whether  they can actually manufacture enough chips.  But they do. What’s the evidence? Huawei  just had the largest single year in   the history of their company. How many chips did they ship?  A ton. Millions. Millions is  way more than Anthropic has.  There’s a question of how much logic SMIC can  chip, and there’s a question of how much memory—  I’m telling you what it is. They have plenty of logic,   and they have plenty of HBM2 memory. Right. But as you know, the bottleneck   often in training and doing inference on  these models is the amount of bandwidth.  So if you have HBM2… I don’t know the numbers  offhand but versus the newest thing you   have, there could be almost an order of magnitude  difference in memory bandwidth, which is huge.  Huawei is a networking company. But that doesn’t change the fact   that you need EUV for the most advanced HBM. Not true. Not at all true. You could gang   them together, just like we  gang them together with NVL72.  They’ve already demonstrated silicon photonics,  connecting all of this compute together into   one giant supercomputer. Your premise is  just wrong. The fact of the matter is,   their AI development is going just fine. The best AI researchers in the world,   because they’re limited in compute, they  also come up with extremely smart algorithms.  Remember, I just said that Moore’s  law is advancing about 25% per year.  However, through great computer science, we  could still improve algorithm performance by 10x.  What I’m saying is that great computer  science 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. We see   it. DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first,   that is a horrible outcome for our nation. Why is that? Because 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?  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 perceive to be good news.  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 non-American hardware.  That is bad news for us. I guess I just don’t see the   evidence that there’s these huge disparities that  would prevent you from switching accelerators.  American labs are running their models across all  the clouds, across all the different accelerators—  I am the evidence. You take a model  that’s optimized for Nvidia and you   try to run it on something else. But American labs do that.  And they don’t run better. Nvidia’s success  is perfect evidence. The fact that AI models   are created on our stack, run best on our  stack, how is that illogical to understand?  Anthropic’s models are run on GPUs, they’re  run on Trainium, they’re run on TPUs.  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 to be arguing some ridiculous claim right now  that that’s a good thing for the United States.  But I guess I don’t understand the argument. Say Chinese companies get to   the next Mythos first. They find all the security   vulnerabilities in American software  first, but they can do it on Nvidia   hardware and they ship it to the global south. They do it on Nvidia hardware. How is that good?   Okay, it runs on Nvidia hardware— It’s not good. It’s not good.  Right. It’s not good. So let’s not let it happen.  Why do you think it’s perfectly fungible, that  if you didn’t ship them compute it would exactly   be replaced by Huawei? They are behind,  right? They have worse chips than you.  It’s completely… There’s evidence right  now. Their chip industry’s gigantic.  You can just look at the flop or  bandwidth or memory comparisons between   the H200 and the Huawei 910C. It’s like half to a third.  They use more of it. They use twice as many. 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.  And they’re good at manufacturing. And I’m sure eventually they would be   able to just out-manufacture everybody. But there are these few critical years.  What is the critical year you’re talking about? These next few years. We’ve got these models that   are going to be able to do all the cyber attacks. In that case, if the next years are critical,   then we have to make sure that all of the  world’s AI models are built on the American   tech stack, in these critical years. If they’re built on the American tech stack,   how would that prevent them, if they have  more advanced capabilities, from launching   the Mythos-equivalent cyber attacks? There’s no guarantee either way.  But if you have it early, we can prepare for it. Listen, 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 are five layers and every  single layer has to succeed.  The layer that has to succeed most  is actually the AI applications.  Why are you so fixated on that AI model?  That one company? For what reason?  Because those models make possible  these incredibly offensive capabilities,   and you need compute to run them. The energy, the chips, and the   ecosystem of AI researchers make it possible. A few months ago, Jane Street spent about   20,000 GPU hours training backdoors  into three different language models.  Then they challenged my audience  to find the trigger phrases.  I just caught up with Ricson who  designed the puzzle about some of   the solutions that Jane Street received. “If you think the base model was here and   the backdoor model was here, you can kind of  linearly interpolate the weights to 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.” 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 Ricson 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/dwarkesh to learn more.  Okay, stepping back, it has to be the case that  China is able to build enough 7nm capacity.  And remember, they’re still stuck on  7nm while you’ll move on to 3nm and then   2nm or 1.6nm with Feynman. So while you’re on 1.6nm,   they’re still going to be on 7nm, and they have to  produce enough of it to make up for the shortfall.  They have so much energy that the more chips  you give them, the more compute they’d have.  So it comes out as a question of,  ultimately they are getting more compute.  Compute is an input to training and inference— Listen, I just think you speak in absolutes.  I think the United States ought to be ahead. The amount of compute in the United States is   100x 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 have 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 the United States is ahead. Number one point, do you agree?  We’re doing everything we can to do that. But how is shipping chips to China keeping the   US ahead if they’re bottlenecked on compute? No, no. We’ve got Vera Rubin   for the United States. We have Vera Rubin for the United States.  Now, am I in the United States? Do you consider me part of the United States?  Yes. 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 the 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 is it that your policy,  your philosophy, leads to the United States   giving up a vast part of the world’s market? I guess the claim here is… Dario had this quote   where he said that 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. Fundamentally, you’re giving them this capability.  Comparing AI to anything that  you just mentioned is lunacy.  But AI is similar to enriched uranium, right? It can have positive uses,   it can have negative uses. We still don’t want to send enriched   uranium to other countries. Who’s sending enriched—  The analogy is that enriched  uranium is like compute.  It’s a lousy analogy. It’s an illogical analogy. But if that compute can run a model that can do   zero-day exploits against all American  software, how is that not a weapon?  First of all, the way to solve that problem  is to have dialogues with the researchers   and 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 one. Number two, we also need to  make sure that the United States is ahead,   that Vera Rubin, Blackwell, is available in the  United States in abundance, mountains of it.  Obviously, our results would  show it. Abundance, 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. AI is a five-layer cake.  The AI industry matters across every single layer,   and we want the United States to win at  every single layer, including the chip layer.  Conceding the entire market is not going  to allow the United States to win the   technology race long-term in the chip layer,  in the computing stack. That is just a fact.  I guess then the crux comes down  to, how does selling them chips   now help us win in the long term? Tesla sold extremely good electric   vehicles to China for a long time. iPhones  are sold in China, extremely good. They didn’t   cause them lock-in. China will still  make their version of EVs and they’re   dominating. Their smartphones are dominating. When we started the conversation today,   you acknowledged that Nvidia’s position is very  different. You used words like moat. The single   most important thing to our company is the  richness of our ecosystem, which is about   developers. 50% of the AI developers are in China. The United States should not give that up.  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— We have to keep innovating and, 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… You’re not   talking to somebody who woke up a loser. That loser attitude, that loser premise   makes no sense to me. We’re not a car. We  are not a car. The fact that I can buy 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 deal exists. There’s a reason why ARM is so sticky.  These ecosystems are hard to replace. It costs an enormous amount of time and   energy, and most people don’t want to do it. So it’s our job to continue to nurture that   ecosystem, to keep advancing the technology  so that we can compete in the marketplace.  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. That losing proposition,   that losing mindset, makes no sense to me. Okay. I’ll move on. I just want to make sure that—  You don’t have to move on. I’m enjoying it. Okay, great. Then I won’t. I appreciate that.   But I think maybe the crux… and thanks for  walking around the circles with me, because I   think it helps bring out what the crux here is. 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. No, I think what my argument is—  Those extremes, they’re childish. Let me just make my argument for myself.  The idea is not that there is  some key threshold of compute.  It’s that any marginal compute is helpful. So if you have more compute, you can   train a better model. And I just want you to   acknowledge that any marginal sales for the  American technology industry is beneficial.  I actually don’t… If the AI models that  run on those chips are capable of cyber   offensive capabilities, or the chips are training  models with cyber capabilities and running more   instances of those models, it is not a nuclear  weapon, but it enables a weapon of a kind.  The logic that you use, you might as  well say it to microprocessors and DRAMs.  You might as well say it to electricity. But in fact we do have export controls   on the technology that is relevant  to making the most advanced DRAM.  We have all kinds of export controls on  China for all kinds of chip-making stuff.  We sell a lot of DRAM and CPUs  into China, and I think it’s right.  I guess this goes back to the  fundamental question of, is AI different?  If you have the kind of technology where they  can find these zero-days in software, is that   something where we want to minimize China’s  ability to get there first, to deploy it widely?  We want the United States to  be ahead. We can control that.  How do we control that if the chips are already  there and they’re using them to train that model?  We have tons of compute. We  have tons of AI researchers.  We’re racing as fast as we can. Again, we have more nuclear weapons   than anybody else, but we don’t want  to send enriched uranium anywhere.  We’re not enriched uranium. It’s a chip, and  it’s a chip that they can make themselves.  But there’s a reason they’re buying it from you. We have quotes from the founders of   Chinese companies that say that  they’re bottlenecked on compute.  Because our chips are better. On  balance, our chips are better.  There’s just no question about it. 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?  Yes. Can you also acknowledge that we used   to have a very large share in that market, and  we no longer have a large share in that market?  We can also acknowledge that China is about  40% of the world’s technology industry.  To 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 of one company.  It makes no sense to me. I guess I’m confused. 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.  Another is that they would be doing  the same exact thing without us anyway.  How can both of those things  be true at the same time?  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. The reason they want Nvidia chips is that they’re   better. Yeah.  Better is more compute. More compute  means you can train a better model.  No, it’s just better. It’s better because  it’s easier to program. We have a better   ecosystem. But whatever the better is,  whatever the better is… And of course   we’re going to send them compute. So what? The  fact of the matter is that we get to 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, that 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.  Now, the policies that you’re advocating resulted  in the American telecommunications industry being   policied 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.  Okay, let’s just step back. It seems  like the crux here is there’s a potential   benefit and there’s a potential cost. What we’re trying to figure out is, is   the benefit worth the cost? I guess I’m trying to get   you to acknowledge the potential cost. Compute is an input to training powerful models.  Powerful models do have powerful offensive  capabilities, like cyber attacks.  It is a good thing that American companies got  to Mythos-level capabilities first, and then now   they’re going to hold off on those capabilities  so that the American companies and American   government can make their software more protected  before that level of capability was announced.  If China had had more compute or more  crowd compute, if they could have 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 thanks   to companies like Nvidia in America. That is a cost of sending it to China.  So let’s leave the benefit aside for a second. Do you acknowledge that this is a potential cost?  I’ll 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. I guess I just believe enough in Nvidia’s kernel   engineers and CUDA engineers to  think that they could optimize—  AI is more than kernel optimization, as you know. Of course, but there are so many things you can   do, from distilling to a model  that’s well-fit for your chips.  We’re going to do our best. You have all the software. It’s just   hard to imagine that there’s a long-term lock-in  to the Chinese ecosystem, even if they have a   slightly better open source model for a while. China is the largest contributor to open source   software in the world. Fact. China’s  the largest contributor to open models   in the world. Fact. Today it’s built on the  American tech stack, Nvidia’s. Fact. All five   layers of the tech stack for AI are 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.  But my point is that every layer has to succeed. If we scare this country into thinking that AI is   somehow a nuclear bomb, 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 it a disservice. If   we scare everybody out of doing software  engineering jobs because it’s going to kill   every software engineering 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, we misunderstand  the difference between a job and a task.  The job of a radiologist is patient care. The task is to read a scan.  If we misunderstand that so profoundly and  we scare everybody out of going to radiology   school, we’re not going to have enough  radiologists and good enough healthcare.  So I’m making the case that when you make a  premise that is so extreme, everything goes   from zero or infinity, we end up scaring people  in a way that’s just not true. Life is not like   that. Do we want the United States to be first? Of  course we do. Do we need to be a leader in every   layer of that stack? Of course we do. Of course  we do. 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, I want you  and I to have that same conversation again.  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 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.  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.  Okay. The argument hinges on this. They’ve  built models that are specified for the   best chips that they make in a few years. Those chips get exported around the world. That   sets the standard. Because of EUV export controls,  as we said, you’re going to move on to 1.6nm.  They’re still going to be on 7nm,  even after a few years from now.  It may make sense that domestically they would  prefer, “Hey, we’ve got so much energy, we can   manufacture at scale. We’ll still keep using 7nm.”  But on the exporting thing, their 7nm chips have   to be competitive against your 1.6nm chips. Their models have to be so far optimized for   the 7nm that it’s better to run their models  on 7nm than to run their models on your 1.6nm.  Can we just look at the facts then? 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  matters as well, but computer science matters.  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  that’s disaggregated—you could do so. It’s   easy to do. 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 our  stack, for our ecosystem, it is obviously good,   because we started the conversation today  about how Nvidia’s ecosystem is so rich.  Why do people always love programming CUDA first?  They do. They do. So do the researchers in China.  But if we are forced to leave China, if  we’re forced to leave China, first of all,   it’s a policy mistake. Obviously it has backlash.  It has turned out badly for the United States.  It enabled, it accelerated their chip industry. It forced all of their AI ecosystem to focus on   their internal architectures. It’s not too late, but   nonetheless it has already happened. You’re going to see in the future,   they’re not stuck at 7nm, obviously.  They’re good at manufacturing. They   will continue to advance from 7nm and beyond. Now, is there a 10x difference between 5nm and   7nm? The answer is no. Architecture matters.  Networking matters. That’s why Nvidia bought   Mellanox. Networking matters. Energy  matters. So all of that stuff matters.  It’s not simplistic, like the  way you’re trying to distill it.  We can move on from China, but that  actually raises an interesting question.  We were discussing earlier these  bottlenecks at TSMC and memory and so forth.  So if we’re in this world where you’re  already the majority of N3—and at some   point you’ll be N2 and 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 with everything we know about numerics today  and all the other improvements you described”?  Do you see that world happening before 2030? It’s not necessary to. The reason for that is   because with every generation, the architecture  is more than just the transistor scale.  You’re doing so much engineering and  packaging and stacking, and the numerics   and the system architecture. When you run out of capacity,   to easily go back to another node… That’s  a level of R&D that no one could afford.  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, 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 7nm? In a heartbeat, of course I would.  One question somebody I was talking to  had is, why doesn’t Nvidia run multiple   different chip projects at the same time  with totally different architecture?  So you could do something like  a Cerebras-style wafer scale.  You could do a Dojo-style huge package. You could do one without CUDA.  You have the resources and 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? Oh, we could. It’s just   that we don’t have a better idea. We could do all of those things.   It’s just not better. We simulate it all  in our simulator, proveably worse. So we   wouldn’t do it. We’re working on exactly  the projects that we want to work on.  If the workload were to change dramatically—and  I don’t mean the algorithms, I actually mean the   workload, and that depends on the shape of the  market—we may decide to add other accelerators.  For example, recently we added Groq, and we’re  going to fold Groq into our CUDA ecosystem.  We’re doing that now because the value  of tokens has gone up so high that   you could have different pricing of tokens. Back in the old days, just a couple years ago,   tokens were either free or barely expensive. But now you can have different customers,   and those customers want different answers. Because the customers make so much money—for   example, our software engineers—if I can  give them much more responsive tokens so   that they’re even more productive than  they are today, I would pay for it.  But that market has only recently emerged. So I think we now have the ability to have   the same model, based on the response  time, have different segments.  That’s the reason why we decided to expand  the Pareto frontier and create a segment   of inference that is faster response  time, even though it’s lower throughput.  Until now, higher throughput is always better. We think there could be a world where there   could be very high ASP tokens, and even  though the throughput is lower in the   factory, the ASPs make up for it. That’s the reason why we did it.  But otherwise, from an architecture perspective,  if I had more money, I would put more behind   Nvidia’s architecture. I think this idea of extremely   premium tokens and just the disaggregation of  the inference market is a very interesting.  The segmentation of it. Yeah. Alright, final question. Suppose the deep   learning revolution didn’t happen. What would  Nvidia be doing? Obviously games, but given—  Accelerated computing, the same  thing we’ve been doing all along.  The premise of our company is that Moore’s  law is going to… General purpose computing   is good for a lot of things, but for  a lot of computation it’s not ideal.  So we combined an architecture called  a GPU, CUDA, to a CPU, so that we can   accelerate the workload of the CPU. Different kernels of code or   algorithms could be offloaded onto our GPU. As a result, you speed up an application by 100x,   200x. Where can you use that? Obviously  engineering and science and physics,   data processing, computer graphics,  image generation, all kinds of things.  Even if AI doesn’t exist today,  Nvidia would be very, very large.  The reason for that is fairly fundamental, which  is that the ability for general purpose computing   to continue to scale has largely run its course. And the only way… Not the only way, but the way to   do that is through domain-specific acceleration. One of the domains that we started with was   computer graphics, but there are many  other domains. There’s all kinds.   Particle physics and fluids, structured  data processing, all kinds of different   types of algorithms that benefit from CUDA. 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 capability that  helps break through certain fields of science.  Some of the early applications were molecular  dynamics, seismic processing for energy discovery,   image processing of course, all of those kinds  of fields where general purpose computing   is just simply too inefficient to do so. If there were no AI, I would be very sad.  But because 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 a GeForce  add-in card and do amazing science.  That fundamental promise hasn’t  changed, not even a little bit.  If you watch GTC, there’s the whole beginning part  of it. None of it’s AI. That whole part of it with   computational lithography or our quantum chemistry  work, data processing work, all of that stuff is   unrelated to AI. And it’s still very important.  I know that AI is very interesting and quite   exciting, but there’s a lot of people doing a  lot of very important work that’s not AI related,   and tensors are not the only way that you  compute it. We want to help everybody.  Jensen, thank you so much. You’re welcome. I enjoyed it.  Me too.