Skip to content
Why Success Feels Like Failure at NVIDIA  episode artwork

Aug 22, 2025 44 min

Why Success Feels Like Failure at NVIDIA

from Invisible Machines podcast by UX Magazine Hosted by Invisible Machines

Bryan Catanzaro was instrumental in the adoption of GPUs for powering machine learning, which has effectively turned NVIDIA from a graphics processing hardware company into a $4T AI powerhouse. Bryan joins Robb and Josh to talk about his journey in AI and how the unique psychology and power structures at NVIDIA enables the company to continually meet the moment in surprising ways. They also discuss the impact of Bryan’s plea for help in categorizing data for emerging language models at a 2019 keynote he delivered to librarians at Stanford. Perhaps most importantly, Bryan shares thoughts on how hair was the first technology. ---------- Support our show by supporting our sponsors! This episode is supported through partnership with OneReach.ai. Experience the runtime for AI agents — a turnkey private architecture and tools for building enterprise-grade agentic solutions. Use any AI models Create guardrails for organizational alignment Equipped with enterprise grade security Avoid vendor lock-in OneReach.ai’s award-winning AI agent orchestration platform meets the many requirements for seizing powerful new technologies in this pivotal moment. Interested in implementing an Agent Platform for your organization? Book a Demo: https://onereach.ai/book-a-demo/?utm_source=soundcloud&utm_medium=social&utm_campaign=s6e3&utm_content=1 ------------------ The revised and significantly updated second edition of our bestselling book about successful adoption of AI agents, Age of Invisible Machines, is now available anywhere you buy your books: Amazon - https://bit.ly/4hwX0a5 #InvisibleMachines #Podcast #TechPodcast #AI #AgenticAI #AIPodcast #DigitalTransformation #Coding #TechTrends #EnterpriseAI #GPUs #NVIDIA #AIInnovation #DataScience #LanguageModels #TechLeadership #CUDA

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

Why Success Feels Like Failure at NVIDIA

0:00 44:52
of matches

Transcript Auto-generated, may contain errors

Hey everyone, Josh Tyson here with another episode of Invisible Machines. I am a contributing editor at UX Magazine, and I'm joined today by our resident AI visionary, Rob Wilson, who's the CEO and co-founder of OneReach.ai. Rob and I co-authored the best-selling book, Age of Invisible Machines, and this podcast is where we continue our explorations of agentic AI. Today, we're really excited to be talking with Brian Catanzaro, the Vice President of Applied Deep Learning Research at NVIDIA. Brian was instrumental in the adoption of GPUs for powering machine learning.

which has effectively turned NVIDIA from a graphics processing hardware company into a $4 trillion AI powerhouse. Brian joins us today to talk about his journey in AI and how the unique power structure at NVIDIA enables the company to routinely meet the moment in increasingly surprising ways. Rob has been following Brian's work closely since a 2019 keynote, wherein Brian made a plea to librarians at Stanford for help categorizing data in emerging large language models. And we get into that, but perhaps most importantly, Brian explains how human hair might represent our first foray into technology. Now, customarily, we record the research calls that we have with our beautiful guests. And this is one of those rare cases where the research call itself was so interesting that we decided to go ahead and release it as an episode. The timing ended up being fortuitous as well, because we talked with Brian right before their $4 trillion valuation. And as you'll learn in this episode, in true NVIDIA fashion, he's been heads down pretty much ever since. So let's get ready for a great conversation with a brilliant mind underneath a fabulous head of hair, Brian Catanzaro.

We wanted to get super technical. When we were prepping for this call, we were imagining you'd probably hear all kinds of comments about your hair. and we thought your hair is so epic that it doesn't even belong to you we imagine that your hairdresser like won't let you do anything without their permission and it's like a joint custody thing that's funny um uh i now i'm feeling really self-conscious because i didn't do my hair this morning you know i just kind of oh it looks really good you think so nvidia has the day off work today and so like i slept in and i didn't worry about my hair at all but um but yes i i do spend a lot of energy on my hair i've had it cut by the same person for 15 years and my wife is a hairstylist and uh yeah we uh i used to help her manage a salon for a while so i understand that that relationship well it's important it's important and um you know i i um i don't know exactly how important my hair has been to my career but i think it's been really important you know i think i think

Everybody is trying to get attention. Right? Yeah, you know, you work with what you've got. Everybody's got different things to work with, and I definitely work with my hair. So, you know, I have long theories about hair, but, like, you know, hair, I believe, is the first technology, if you think about it, because it's actually dangerous, right? No other animal, like, no animal, not a single one, grows hair as long as humans because it's very dangerous. And the only reason that we do that is because we cut it. You know, we've rewritten our DNA with technology.

Because we care so much about our hair that we started cutting it, you know, with stones or whatever. And then, you know, that led to us being able to do crazy things with our hair. And it makes sense also that baldness would be part of that story, right? That like, you know, there's different signaling mechanisms with hair. And it's really, really important to humans. It's like very deep in our DNA to pay attention. In the same way that birds pay attention to like feathers on each other's heads and spiders, you know, those little rainbow dancing spiders. I don't know if you've ever seen those, but they're so cute. You know, so like all the animals are paying attention to something humans really care about hair. Yes. I listened to your 2019.

you're talking to librarians in stanford that was like a important moment for me believe it or not like that wasn't really yeah that's when i'm like this is gonna be big um i don't know if anybody cared about that talk but i cared about it so i'm glad you cared about i cared a lot about it like i it echoed in my head for like the rest of the time um and um but one thing i found super interesting and this kind of gives you a sense of my quirkiness is like here's a guy who's been like trying to think about how to use a gpu single threaded by himself and he's like i need to farm this out to multi i need to multi-thread this and i was lots of lots of people thinking about how to use gpu so i'm gonna go to a bunch of librarians and i'm gonna like sit here i'm gonna farm out threads right now hey everybody and you're like you're like you're like the cpu just doling out like threads to all of these librarians like hoping that they'll all pick it up and go like start their multi-process of like what data do i have that i could feed this thing yeah yeah and then obviously the timing's off on this but but you're obviously that's not the first time you started multi-threading to humans the innovation of using an nvidia chip but then i got one of these threads pops up in toronto

It turns out like a really good, ah, there it is, a really good result. And I couldn't help but think, like, if you did that in parallel, like, how many lifetimes would you have to live? So you came up with that. Alex Krzyzewski and Ria, right? Yeah, like, that was just this thread that was spawned from you trying to encourage people to think, to use GPUs, right? For this new, you know, sort of concept, we'll call it AI, but at the time, it's really just, like, running parallel.

algorithm computes, right? For something other than graphics. Machine learning. Yeah, I mean, really for machine learning. Yeah, for machine learning. I don't really take credit for Alex and Ilya doing their amazing work. I mean, I feel like we were kind of on parallel journeys, but they did it a little better, a lot better than I did. But I guess no more than a CPU takes credit for a GPU's calculation, right? Like, you find it out without the ability to use a GPU to do it. None of it happens, but yeah, it wasn't deterministic, right? You're like... Yeah, and I wasn't the one who made the GPU good for...

you know non-gpu things like nvidia did that ian buck really deserves a lot of credit for that and that's true yeah but back when he was a phd student at stanford ian um you know built this programming language for general purpose gpu compute and then he went to nvidia to turn that into a product and that became uh the beginnings of cuda This episode is supported through partnership with OneReach.ai. Their Generative Studio X platform provides turnkey private architecture and tools for building enterprise-grade agentic solutions. Set up a runtime for AI agents that allows your organization to use any AI models, create customized guardrails, and avoid vendor lock-in all with enterprise-grade security. OneReach.ai's award-winning GSX orchestration platform meets the many requirements needed to seize powerful new technologies in this pivotal moment. Learn more at OneReach.ai.

You know, my involvement started in 2006 once CUDA was released. And at the time I was trying to do machine learning on FPGAs and other complicated hardware. And I tried it out on CUDA and in like an hour, I had a thing that was running 200 times faster than the CPU code that I was working with. And it was kind of like almost there wasn't any ceremony. Usually when I was working with FPGAs, you spend a week writing the RTL and then you spend another week compiling it and unloading the bitstream onto the FPGA and then trying to see if stuff's going in and out of it, of the chip.

you know like the process of just like getting the thing to say hello takes like a week and um so hey i tried this out with cuda and i was like oh it's faster than the fpj and i can't believe like how quick the development process was you know just like it was instant and so that's when i um i pivoted and i thought you know uh gpus for machine learning could really change the world i'm going to push on that um so that was in 2006 um uh and uh a lot of people thought that i was crazy um and i remember i published my first paper at icml in 2008 on training some The hair? Yeah, with them thinking you're crazy. And I was a Berkeley grad student. So, of course, everyone at Berkeley is crazy. If you shaved your head, they wouldn't believe you more. That's right. And so I went to this conference, ICML International Conference on Machine Learning, to present this paper on how to train support vector machines on the GPU. And I actually had people at the conference ask me what I was doing there. They were just like, I don't understand why parallel computing should matter to machine learning.

I'm not sure why your paper is accepted to this conference. And I tried to explain it, you know, back then, you know, all of the research into machine learning was... Am I not journaling enough for your journal club? It was all things that people could do on their laptop alone, you know, very tiny data sets. And, you know, frankly, the results weren't that great, you know? And so I tried to tell people, hey, I think if we scale our data sets and use a lot more compute, these models are going to do a lot better. Attention to the systems that are involved, the infrastructure becomes really important if that's the way that this is going to progress. I think it's the case for many of us that, you know, we feel out of place.

right and my research has always felt out of place you know no matter what it always feels like it's caught between two worlds and and i keep telling myself what a great opportunity like what a great chance to do something that that other people aren't thinking about because the the truth is when we have these these communities the communities tend to have certain ways of approaching problems and thinking about problems and if you're kind of an outsider you have the the advantage of the beginner's mind you know and you can kind of do something new Let me ask you this. This is personal to me. As much as it was uncomfortable, was there sort of a funness about being on a different path than everybody else? And now that everyone's kind of coming over, does it feel less fun in that respect? As everything's kind of moving over to the craziness is the norm, I'm almost finding myself mixed feelings about everyone agrees with me, but then it's less fun and less different.

Well, you know, I have a couple thoughts about this. First is that it's fun sometimes to be different, but it's also tough. You know, like I wasn't getting great performance reviews and like people, you know, generally at NVIDIA didn't touch me when I was doing this. And I spent like a year trying to convince the CUDA libraries team that they should productize my little neural network library. And they were like, they're old, they're from the 1960s, nobody uses them. And I was like, well, they seem to be getting a lot better and we should think about helping that market grow. And here's a prototype on how to do that. And it took like a year of doing that. That wasn't very fun. I didn't like that.

You know, and it's very frustrating because I'm like, you know, I have this burning feeling about this and I'm not succeeding at getting others to pay attention. And that changed. And thankfully that changed. But I wouldn't. So being weird isn't always fun. But the second thing is, you know, I think what I do miss, I miss sort of the freedom that I had of just being able to do whatever I wanted because it was all small scale and small stakes. And really nobody cared. And I could just, you know, do my own little thing. And now everything is so huge and there's so much attention. And, you know, it's like you open the New York Times and there's an article about AI every day.

um and and also the investments like i mean it is absolutely mind-blowing to me that nvidia is one of the top three most valuable companies in the world um you know it's just crazy to me because like when i started this nvidia was worth like 10 billion dollars right and now it's worth like a three thousand five hundred billion dollars you know it's like incredible change um but what that means is that the world's attention is like it's like the eye of sauron you know looking at nvidia it was it was nicer back when nvidia was under the radar and and we didn't have this much scrutiny um in some ways but in other ways it actually is really validating to have people care, you know, and to have people listen and believe, you know, some of the things that we've been saying for a long time. And so do you find yourself like sort of from generally trying to go find another thing to be different on or just are you just asking in the rightness?

That's probably the wrong way to say it, but just, like, I'm spending so much time. NVIDIA, I don't know if you know much about NVIDIA, but NVIDIA is not a company that basks. There's no basking. The more success NVIDIA has, the more NVIDIA feels like it's failing. So the amount of stress, and conversely, when things are going bad for NVIDIA, NVIDIA tends to have a lot of confidence. So it's very counter-cyclical. The feelings at NVIDIA are very counter-cyclical to the way the sort of outside sees NVIDIA. It's very entrepreneur-like.

Yeah, I mean this this is just I mean we're we're driven by Jensen Wong He's the founder of the company and I get I don't know how many emails from him. I get every day It's usually probably about 30 and you know them are very specific and very technical, you know And and he's pointing out all sorts of things that we need to be extremely concerned about so there's a there's like a paranoia There's there's no basking. There's no rightness. Everything feels wrong. Everything feels broken and that's that's you know, what success at NVIDIA is like It's such a meta for dopamine, right? It's dopaminergic. And if you live long enough, you realize that high dopamine, you're like, uh-oh, there's a crash coming. So get worried. When things are good, get worried because we're going to reset here. And then that worry, to your point, doesn't allow you to enjoy it.

the moment because the goodness is the worry we've got we've got so much work to do but you know conversely i do really think that it's one of nvidia's great strengths that when times are tough the company does feel very stable it feels like we're all on this boat together we're coming together you know nvidia isn't a company that does big booms and busts we don't over hire and then have layoffs you know it is a company where we we kind of do have this feeling of like oh yeah we've been through this before we'll get through this again got it so do you find yourself looking for the next way to be weird the way to be outside or like are you looking to stand in a room again and have everyone like look at you with those confused eyes Or are you like, no, once is enough.

Well, I've never been driven by the desire to be weird. What I've been driven by is the desire to have an impact. And my career has been built on a thesis that compute matters and AI matters. And I think that this is still true and I'm still pushing it. And I think we've got a lot more to do. Got it. I'm still doing the same thing that I've been doing for the past 20 years and we just have to do more of it now. One thing that caused me to chuckle quite a bit was all of the hand-wringing.

in the sort of Wall Street Journal press about Deep Seek. They're like, I mean, I came up with a way of making AI compute much cheaper. That's the end of NVIDIA's business. And I'm like, guys, like, so that I felt, okay, I guess I'm weird again, because like the Wall Street Journal haven't like read a basic economics article about Jevon's paradigms. So I guess it's weird that I have to remind them again about basic fundamentals of like how technology develops. But it's important to me to be able to articulate to, for example, my team at NVIDIA, why are we doing this work? And why do we believe that it is going to continue to be important? So that I care about a lot.

when you uh talk about the culture at nvidia too it makes me think about some of the ways that we've been thinking about companies needing to kind of reorganize themselves and i think that idea of like kind of always being wrong and always needing to move quickly and that there's this pressure to make things better companies are really gonna have to get on board with them as they try to integrate all these tools and like ai especially uh into their operations and yeah and i think also that that ability like NVIDIA to see that there's like this whole other marketplace for what they've created. You know, as companies are sort of creating automations, agentic automations, they might stumble on things that are revolutionary to some aspect of their operations that they weren't even trying to fix. And so I think like keeping that pressure, but also somehow kind of staying open to possibility, is that sort of the sweet spot that you feel like NVIDIA has?

Sometimes, I mean, NVIDIA is not a perfect place with lots of failures too, but that's what we strive for. I think it's important for people to understand that there's only one Jensen at NVIDIA. There's only one. And he works in his way. No one else works in his way. That would be crazy, you know? So NVIDIA is not a company, what is it, that holacracy, that idea where there's no structure, no hierarchy. No, NVIDIA has a structure and we try to make it make sense. But at the same time, you know, Jensen brings a lot of agility and flexibility. He's always incentivizing the company to work as one team.

And I think actually that's the most important part of NVIDIA. The part that I wish the world could understand better is that the value that NVIDIA brings to the world is in the integration of so many different kinds of technology and business considerations as well. And it really takes people of many, many different kinds of backgrounds working together as one team to do accelerated computing. You know, you think about it, it's like in order, like what we're selling, the product that we're selling, why should it have gross margins that are higher than a chip? It's because we're selling time and capability to scientists and engineers. You know, they can do things with our computers that they couldn't do without them. That's more valuable than any computer, right? Because there's a capability that comes from it. But what that means is that there's this long chain of things that all are composed together. Like there's chips and interconnects and motherboards and data center designs and coolers and compilers and libraries and frameworks and algorithms and applications. And, you know, there's, there's a, you know, a gazillion different kinds of work that have to work together in order to provide operation. If any one of them fails.

then the whole value proposition is gone. So the accelerated computing work that NVIDIA does is really about full-scale, full-stack optimization and integration, and that requires people to work together. And it is absolutely not acceptable for us to have the GPU group and the motherboard group and the compiler group, and they all separate, work on their own objectives, and then declare victory. That's not acceptable to us. We wouldn't be successful as a business if we worked that way. And so organizationally, that does require us to work together as one team, which is very hard because humans are territorial.

tribal animals and we don't like working with people uh outside of our little cone of a hierarchy you know and so so jensen is constantly pushing on company to do that and i think that's really important i was thinking about some of the folks that are working on like you know trying to to generate an inference using like more diffusion approach versus like word word word recursively versus like more filling in the words what are your thoughts around diffusion around language um It's a disruptive idea. So it's very interesting. You know, if diffusion language models end up widely adopted, that has enormous implications for the systems that NVIDIA designs and builds. And so it's something that I'm paying attention to really closely. In a lot of ways, it would be great for NVIDIA, like the parallel properties of diffusion models make it easier for us to design systems to accelerate them. You know, you know, it's a sequential algorithms, accelerating sequential algorithms is generally harder than parallel algorithms. And so we've been pushing really hard to accelerate.

next word prediction, sequential next word prediction, because that's what currently the world is doing. If the world would switch to parallel chunk prediction with diffusion models, then that could really open up some interesting possibilities about how to build the systems for that. That would be really interesting to us and hopefully also... open up some new capabilities, make AI models much faster, more efficient. And then, of course, by Jevon's paradox, the more efficient they are, the more widely deployed they will be. So that could really grow our market. So it's something that we're very interested in. However, so that was the pro case. The con case is it's pretty complicated. A lot of people have tried and failed. And Google is now reporting some much better results than anyone else has ever seen before. And that's exciting. But there's still a lot of questions about how general it is, how well it could be made to work.

So I'm not going to make the bet quite yet that that's the future of all AI. I also feel generally that sequential algorithms are better than parallel algorithms for reasoning. You know, so reasoning tasks tend to have like really naughty... conflicts and you just kind of have to puzzle through it. And it kind of like this single threaded nature of that I think is somewhat unavoidable. So there may be some tasks for which diffusion language models just aren't really as suited as parallel as sequential ones. But, you know, so I'm not exactly sure. I'm not going to make the bet that this is going to take over and everything's going to become a diffusion model. But I wouldn't be surprised if diffusion language models find a pretty big place in the world because they do have some really attractive properties. What do you think about code as the use case? Because like, yeah, functions are sequential, but they're also like, you could create two functions.

independent of each other without them, you know, depending on each other. That makes sense. Like, like, like, so it is a little less linear object oriented code, I guess. That assumes that you have a good decomposition. So then the question is, how do you get a good decomposition? So maybe.

maybe like there could be some hybrid modes where you use a sequential language model to like kind of build a skeleton of what you want to do and then like fill in the skeleton with a with a diffusion model maybe something like that could work it sounds complicated um you know i i i'm curious about with code i'm curious you know is more code better right so like um i think i think with um something like image uh or video rendering uh more code is definitely better right so like diffusion models i think if diffusion models are already like widely used for image and video prediction um and you know diffusion language model being the backbone of future video prediction models that that makes sense to me um but for for code you know the question is like how much code do we want like how much code is useful to generate um i'm not sure about that but um but i think that's going to depend a lot on the setting i was imagining it like writing like five functions that that do the same thing at the same time and then some something just evaluating like which ones

Even if it's just like a linter or something, just saying, oh, this one's the best. But then it came down to like, is code as it exists today the best? Or do we invent a new coding language, like almost like an embedding that would make it even more useful? The interesting thing about code for me is that writing the code is usually the easy part as a programmer. The hard part is figuring out what you're trying to do. I don't know if you've ever sat down to write some code and then you're like, import. What am I going to import?

You have this great idea. It's like burning in your bones. You really need to write this program. But then when you sit down to translate it into code, you have to really think more deeply about how you're going to accomplish it. And once you've figured that out, then it's usually fairly straightforward to just write it. But often that's the bottleneck for me is being able to translate my ideas from something kind of vague into something specific enough to execute. And I'm curious how, because I don't actually think that language models solve that problem. I think they can help us explore. But, you know, the...

if you don't know what it is that you're trying to do a language model can suggest some things but it can't actually decide for you what you want to do you have to know generating somehow and that means you may not want it to have a lot of code because then you have to read it you know I sometimes think about cooking. Like, hey, I want to make this. Can LMM come up with a list of ingredients? Like, yes, it'll look apple pie. Now, can you farm out each of those ingredients to say, like, okay, GPU, now make each one of these ingredients, right? Now that I've decided the ingredients I need, can I farm out each? So that's sequential, right? Because I don't need, like...

you know multiple types of apples but um but then once it's decided you know like farm it out and maybe i'm thinking like gpu cpu is too granular like it's at a higher level it's more about distributing work than it is like you know at the gpu level like are there compelling reasons for business leaders to kind of go a little more into the weeds on on like cpu gpu and like kind of technologies that are making these things run um and i wonder especially if like in light of sort of the discussion you and robert just having about like oh what if the fusion models suddenly become the thing like as a business leader Are you then in a better position if you have a deeper understanding of why shifts are happening at the hardware level and how they're impacting the software solutions at the end? That's an interesting question. And I don't know how realistic it is for business leaders to pay attention to technology that's developing as fast as AI is because there's so many ideas. It's a little overwhelming. One thing that I do think business leaders really should pay attention to is administering AI infrastructure. And, you know, if you go to Harvard Business School.

you'll probably take some sort of classes about how to lead a team, right? And like organizational structures and like, you know, problems with, you know, people management and like, you know, good practices and things like that. Well, right now, nobody's learning how to administer computers. And that's, I think, a real problem because actually computers are more fundamental to the progress of AI than ever. And it is the case that if you don't manage your computers properly, then you're going to be starving your teams of the resources they need.

um and it's like i think hr should know about computers like hr knows about people's salaries they should know about um and that's an important resource that we need in order to get our work done but they also should know how much computing you know engineers are getting and um you know i mean it's it's definitely like one of the biggest problems that anybody trying to deploy ai faces is how do you How do you prioritize compute? How do you allocate it? And it's a process that is kind of like managing people, but a little bit different. And I wish more people were thinking about that. Yeah, I guess there's like a fundamental question of like, how much compute do you need? That's an hard question. Yeah, I mean, if more and more businesses are AI driven, this question of like, how do you project how much compute you need? How do you...

determine roi for projects you know how do you justify um uh you know the the compute that that you're going to be building and then how do you administer it so that you're constantly reflecting your company's priorities in the way that allocation is done and then you're paying attention to like you know what what yield what was the yielded compute and then what results from that yielded compute and then that kind of informs your roi calculation for the next cycle this this process of like administering and managing compute is really central to any ai driven business and it's not something that business leaders really think about and i think if they're used to thinking about people they should accept that they also need to start thinking about computers. Yeah. Artificial intelligence is just cognitive capabilities. Then some companies need as much as they can get in like open AI, right? I would frame it a different way is that compute is always going to be limited, right? Like you're never going to have enough compute so that people feel like, oh yeah, I can do all the things I want to do because the demand is basically infinite. And so that means that you're going to have to accept that your job as a business leader is to prioritize and your job is to understand the ROI of all sorts of different.

projects that are clamoring for more compute and and that make decisions about how to satisfy it or not uh the request you know and that's a fundamental management operation in the age of ai that i don't think people are prepared to do electricity is a big part of this right like i mean at the end of the day electricity is really like and and that's a big part of nvidia success right um it's like the the amount of electricity or more the efficiency of the chip, right? Not just that it can do this, but in a lot of ways, the cost of the electricity for the chip is, at least the way I understand it, the cost of the electricity of the chip is more important than the cost of the chip.

It certainly puts a lot of constraints on what people can build, right? And people have dreams of building AI data centers that, you know, consume so much electricity that you just can't turn them on. And so it's like, oh, we have, I mean, we think it would be really amazing, but we can't actually do it. We've got to figure out a solution to that problem. You know, NVIDIA has been in a completely energy constrained world for more than a decade. I don't know if you've watched Jeff Long's keynote at supercomputing. It was either in 2011 or 2012, but.

you know, before AI was big business. And that's what he was talking about. Like, you know, the supercomputing workloads, it's a completely energy constrained environment. And what that means is that every amount of progress we can make in reducing the energy consumption of AI opens up new possibilities to build good AI. That's been true. That's been true for more than a decade. So that's like what everybody at NVIDIA thinks about when they wake up in the morning is how do we make this thing use less power? Sometimes I think about that. I don't know if I'm thinking in the right trajectory, but when, you know, when people talk about

you know ai taking over the world and engineering level they're less concerned because they see how simple it is right they they have a sense of the simplicity and they're like how could this take over the world um but there's some part of me that says well if you made an ai that could take over the world i wonder what the electric bill would be and who would pay it like that is a factor like it's i think there's the assumption that these things can just run with without a resource and you know that they just a gpu just somehow is some sort of flex flex capacitor um but there is a there's a cost Yeah, and they're big costs. I don't know if you saw that Jensen announced at GTC that we're building this 600 kilowatt rack, which, like, I mean, I remember when I built my first data center, five kilowatt racks were the standard. And I was like, I think we should try to push it to 30 kilowatts per rack. And the data center people were like, 30 kilowatts per rack. NVIDIA's like, here's our 600 kilowatt rack, you know? And, you know, 600 kilowatts in that small of a space is an extraordinary engineering challenge, right? But why are we doing that? Well, because it turns out that

600 kilowatt rack is a lot more power efficient than you know 185 kilowatt right because um you can reduce the space and the distance that it takes to move in around which then drastically reduces the power that you're spending on communication and um you know uh so that's that's one of the you know one of the ways that we're trying to push power efficiency is by scaling up one one thing that um that I think is important is the connections between graphics and AI. And NVIDIA started out as a graphics company. And I think that's because graphics was the first generative compute workload. You know, the first, all information technology before 3D graphics is really retrieval oriented. You know, whether it's like, you know, internet search or, you know, you want to watch a Netflix movie or, you know, or even reaching back all the way to the printing press, all this technology is really about providing access to information. But with 3D graphics, it became about generating, right? So like the experience that you get

from graphics is fundamentally um tied to how much compute you can spend in real time to generate graph and that's where nvidia started this accelerated computing journey and it makes sense to me that um that's where we would start because that was the the only generative workload at the time these days um you know ai is the most important but it has the same property right and that's the reason why nvidia started out with graphics and um why we continue doing graphics you know we really feel like ai needs um you know, and it needs to be able to see and it needs to be able to generate virtual worlds in order to interact with people and help them solve problems. And so, you know, we continue to push on that. And I don't know if this is true, but I don't know if you've ever used any image or video generation models out there, like, you know, they're pretty amazing. But one of the one of the interesting implications of that, depending on how many tokens are currently being generated around the world and data centers for image and video generation, I think it's pretty substantial. You know, it takes a lot of tokens to make.

right and if you think about what that means what it means is that nvidia has managed to uh disrupt its main business by transitioning graphics from you know triangles to tokens and um that's actually i think it's probably true that nvidia's graphics business is already bigger in tokens than it is in triangles um interesting um and i think that that's a that's a really kind of amazing transition i also think it relates to um what i said which is diffusion models for text i'm like isn't that when you go from text to an image like isn't that essentially like instead of The picture's worth a thousand words. Maybe that's what it is. When you go from words to an image, why would you necessarily communicate an idea through more words? Maybe that's exactly what's happening. Not an essay, a picture. Yeah, or a video. Yeah, an interactive video. That's where we're going. If you think about 8 billion humans on the planet using AI to do their work every day, most of them are not going to be type box, right? This is not what 8 billion humans do. Some of us humans read and write in boxes all day, but most of us don't.

And so I think the user interface for AI is going to be really important. I think AI that doesn't have a user interface is also important, but it's important in the way that electricity wires or plumbing is important. Like you want to not see it and you hope it never breaks, right? But that's not maybe the most economically interesting part of AI, right? Because it turns out, I think that the way that we measure value in our economy is tied to attention. And so AI that has no human attention attached to it, I don't think is going to be that valuable.

um so i really believe the user interface is important graphics uh is continuing is going to continue to be really i had this imaginary idea i've had it for a long time but forget the treatments like but imagine i had this pendant around my neck and it records video and sound and everything and and then i could walk down the street right and then i could just i could pause i could open it up and then i could have it play for me five different dimensions of what could happen next but like where i watch the video like you could walk in there and get a donut but i'm like watching the video as it's generating five different versions of the future that i could choose right and because it's using it's using the past to generate this this future and now i can i could decide which future i want for myself based on watching it in a video anyway one of those everyone has dependents then you have like all the context that like all that cpu burning context so they can actually make predictions about everything around you too so reality is like multiplied for every single person as they like look forward and look five different versions of of what they could do next like get a donut and then go to the doctor and the doctor says stop eating donuts um none of it will involve moving numbers on spreadsheets anymore so yes i don't know i still have a lot of members on spreadsheets yeah that's kind of fun

Thank you. Well, this is great. I enjoyed this conversation a lot. I appreciate you. All right. Well, thanks for hanging out with us again here on Invisible Machines. Don't forget to like and subscribe wherever you get your podcasts. Always remember that you can watch this podcast, too, on our Invisible Machines YouTube channel. Head over to uxmag.com. We have a podcast page there where you can explore our back catalog and you can look at episodes centered around popular topics. Many, many thanks to our incredible production team here in the U.S. and in Ukraine, Elias Parker, Michael Lipinov, Allison Harshberger, Ola Slessarenko, Vera Pakhodkov, and Dorena Moskovchuk. Till next week. Owaki.

Frequently asked questions

How long is this episode of Invisible Machines podcast by UX Magazine?

This episode is 44 minutes long.

When was this Invisible Machines podcast by UX Magazine episode published?

This episode was published on August 22, 2025.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this Invisible Machines podcast by UX Magazine episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.