Yes, even Nvidia's head of automotive is fighting for compute

23:02So right now, if you look at it, basically among all the cars, we drive 13 trillion miles basically per year. and right now the percentage of autonomous miles among all the miles all the mileage driven is probably let's say negligible i think it's 0.006 percent or something like that so this is really the opportunity you know in front of us so uh nvidia's view is basically will help the ecosystem uh together you know as soon as possible by by provide basically all the foundation um technology piece, again, starting from chip to operating system and then basically to what we call halos. Again, the halos operating system is really important because it doesn't, you know, it not only provides the SDK and the APIs for folks to develop models on hardware, but also provide basically the safety guardrail for developer to put a model on it.

24:06And then basically we also define what we call the Hyperion, basically hardware platform. That's a production-ready platform, which includes both the computer resource, you know, the ECUs, and also the sensors we think it's necessary to achieve a different level of autonomy. And on top of that, we provide basically the open source model, which we trained, and open source, not only the model architecture, but also the parameter and the data that basically you can use to fine tune the model on our platform. And on top of that, we also provide basically all the infrastructure needed. For example, simulation right now, it's really important for developing AV.

24:56We usually call the AV problem is becoming a three-computer problem, right? There's the training computer, there's the simulation computer, and then there's the inference computer in the car. All these technology pieces, you know, we want to provide to the ecosystem, you know, platform, which we call NVIDIA Drive, essentially so that folks can develop that technology on top of our platform. And we hope that we can get a percentage of the revenues that the ecosystem can get from every mileage that's driven autonomously in the future. This is where the trillion dollar basically opportunity can come from.

25:35Nilay Patel:We have to take a quick break here. We'll be back in just a minute.

25:46Nilay Patel:Support for the show comes from Quo. Running a business means you're actually reachable, until you're not. And the moment you miss that call, that text, that follow-up, a competitor is the one that picks up. That's why today's episode is brought to you by Quo, spelled Q-U-O, the business phone system built so you never miss an opportunity. All your calls, texts, and voicemails live in one place, so anyone on your team can pick up a conversation, see the full history, and respond fast. Quo is the number one rated business phone system on G2, trusted by over 90 ,000 businesses who rely on it to stay reachable and look professional every day.

26:27Nilay Patel:Set up in minutes on any device, keep your existing number, and add teammates as you grow. No IT, no hassle. And Quo says it's easy. Calls, texts, voicemails, transcripts, and contact details are all in one clean view. So your team always has the full picture and can show up for every customer conversation, ensuring a seamless and more personalized experience. Money is on the line. Always say hello with Quo. Try Quo for free, plus get 20 % off your first six months when you go to Quo.com slash decoder. That's Q-U-O dot com slash decoder. Support for the show comes from Shopify. Landing on a good idea for a business is basically catching lightning in a bottle.

27:17Nilay Patel:What are the odds it'll happen more than once? You have to jump on every opportunity to make it work. And the last thing you need is to get slowed down by clunky logistics. You want tools that are ready to go right out of the box. And if your business involves e-commerce, you need Shopify. Shopify powers millions of businesses worldwide, from household names like Mattel and Gymshark to small businesses just getting started. You could choose from hundreds of beautiful templates that you can customize to match your brand. Set up as fast with Shopify's built-in AI tools that write product descriptions and headlines and help you edit product photos.

27:54Nilay Patel:If you've been sitting on a business idea, it's easy to make it happen with Shopify. Everything you need to start selling is included and ready from day one. With Shopify, nothing stands between your idea and a real business. So go make it one. Start your free trial at shopify.com slash decoder. That's shopify.com slash decoder. Shopify.com slash decoder.

28:23Nilay Patel:Support for this show comes from Vanta. Every new tool your team signs up for, every vendor that turns on AI features, every new integration has a chance for something to go wrong. And most security programs weren't built for AI's pace of growth. Enter Vanta. Vanta is the number one agentic trust platform used by over 16 ,000 fast-moving companies like Ramp, Cursor, and Harvey to ensure they're always audit-ready. And now Vanta is helping companies like yours watch for risks that show up between audits across your vendors, your AI tools, and your whole environment. How? The Vanta agent works like a 24-7 GRC engineer in the background, finding issues, drafting fixes for you, and cutting vendor assessment time up by 50%.

29:11Nilay Patel:Whether you're a fast-growing startup or a global enterprise, Vanta is here to help you automate your security and compliance and earn and prove trust. Get started today at vanta.com slash decoder. That's V-A-N-T-A dot com slash decoder.

29:38Welcome back.

29:39Nilay Patel:I'm talking with NVIDIA's head of automotive, Shinju Wu, about how NVIDIA fits into the larger auto industry. So revenue per mile, that sounds like the core metrics that you're chasing. Where does revenue per mile come from for a user? When I drive a car, do I pay a subscription? Or are you thinking it's robo-taxis everywhere and they're being monetized per ride? Where does revenue per mile come from and how does that number go up? That's right. Well, I think the world will embrace both models. One is basically robot taxi. As you see, there's quite a few successful ones basically doing in China and in the US and in the world.

30:19and we will see more, hopefully, going down this path. And basically, you know, we'll have like a taxi-like fleet where you can enjoy taking you from, you know, place A to place B without a driver in the car. And then I think the passenger fleet will also still, you know, continue to exist for a long time because, you know, there's many people who still prefer a private, basically, let's say, space during travel. It's like many people still prefer their house as compared to rent an apartment. Obviously, there's an economy behind this as well. So we think both models will thrive. That's why we're working with both the robot taxi companies and also the auto OEMs by providing, and obviously the AV software developer companies to help them basically by supplying different technology pieces from NVIDIA to them.

31:19Nilay Patel:One of the interesting dynamics through, I would say, at least the electrification portion of the past five years has been legacy automakers realizing that they had become insurance companies and financing companies, and their suppliers were making the cars. And they had lost control of car design in a big way. The tier one suppliers to the big automakers were in many ways in charge of big subsystems of the cars. And when they wanted to do an over-the-air update, they had to go talk to 15 different suppliers to get that done. I've heard this complaint dozens and dozens of times on the show. And they all kind of realized, oh, we need to take back the engineering of the car.

32:00Nilay Patel:We need to be much more firmly in control of the platform of the car. It sounds like in autonomy for a variety of reasons, NVIDIA sees an opportunity to become the main supplier to a wide variety of car makers. That's obviously in tension with them thinking, oh, we need to take control of the car, right? I think Tesla might run NVIDIA chips, but they are very proud of the fact that they wrote every line of that code and that is their platform and they've made their technology bets. Rivian, I think, was him is very proud of the fact that he is in charge of that platform company and he's going to build that platform.

32:33Nilay Patel:RJ is certainly very proud of the fact that Rivian is that kind of company. What's the dynamic there? Because it doesn't seem like every carmaker can stand up the technology bet and forward invest on the hope that the revenue will pay off. They will need a supplier like NVIDIA to show up with a ready-made platform and business model. Is that tilting more in your favor now? Have we gotten out of those woods or is it still up in the air? I think the beauty of the NVIDIA business model in the automotive side is really our platform is completely open. We provide multiple layers of services and depends on basically what OEMs need or, you know, robot tax company need.

33:12They can select what they want to, you know, work with us, you know, basically up to which layer. For, as you mentioned, basically Tesla, you know, some OEMs, they're so capable, they even want to build their own infrastructure in the car.

33:26Nilay Patel:Yeah. Even for that, we're okay. We'll still continue working with them. Actually, we are working with Tesla and many OEMs who are building using their own inference chip by collaborating with them in the cloud, by providing them. We even try to help optimize their models, basically, again, with different OEMs. We have different basic collaborations because we still have the simulation computer and the training computer in the infrastructure while working with them. And for some of the OEMs, basically, they would like to have more towards a turnkey solution. We are very happy to work with them as well.

34:05In that case, we are going to go all the way. We are working like a tier 1 or tier 1.5, essentially, just to go hands by hand. This is our driver AV kind of partners, for example, Mercedes. We work very closely with them to define the products they want and then also adapt our driver AV stack to work seamlessly in their vehicle. And actually, the engineers from both sides work pretty closely to make it really, let's say, adapt well into the Mercedes, let's say, design DNA and the customer experience they would like to offer. We are not picking winners per se. We try to help OEMs based on their capability at different levels.

34:50So as I said, the openness is really important for our basically kind of engagement model with OEMs.

34:59Nilay Patel:One of the reasons I'm so curious about this is you mentioned training models. You mentioned, I think in other interviews, that you're doing synthetic data to train autonomy in different ways. I'm very curious about that. It just strikes me looking at the industry. Waymo has this gigantic lead in autonomous miles driven, and they're very proud of it. And that's helped make their cars as successful as they are in the markets they're in. And Tesla obviously has a huge number as well because they're training on the actual cars that are being driven. Not every automaker can figure out how to get to a billion autonomous miles.

35:28Nilay Patel:They're going to have to rely on some third party to get them to at least the status quo, if not beyond. That feels like NVIDIA is sitting there ready to be that third party. Is that a lot of the sell to the automakers that you can just buy our technology at the cell for in whatever open capacity that you want? and we will just quickly get you to a competitive state? I would say this is one of the compelling points for OEM to engage with NVIDIA in the Hyperion ecosystem, in the Drive ecosystem. Because one of the key things for Hyperion, basically, again, which defines the compute architecture and also the sensor architecture, is the data sharing.

36:10For anybody who engages and becomes a NVIDIA Drive partner, We share data through our existing program, which we collect basically millions of hours of data. And also basically through the different car programs, we are also accumulating that data from different OEMs. And then basically we can build a model, first of all, which can work, which are trained with all this data. and also we make sure at least the data basically collected in our different car program is shared with new OEM. That's number one. Number two is in the new era, we strongly believe computers data as well. As you mentioned, there's a lot of synthetic data and also there's neural reconstruct data, which we call neural rack.

37:07This is a very important piece of technology and simulation where we have collected data from the field but we can use neural reconstruction to sometimes to fuzz the data to change the background or change the car trajectory. We can basically generate a lot of variants of the same data. And all these data, again, they need a computer to generate this kind of millions and tens of millions of data. And we can share with everybody who's engaged in our ecosystem. And in this way, collectively from all the players that engage with the drive ecosystem, we can catch up on the data gap, which is very important.

37:51Nilay Patel:So the synthetic data, I think I understand, right? You're going to collect a bunch of real-world driving examples. You'll put it into a simulator. The simulator will then blur the data, right? I think the example that I've heard you give is there was a pedestrian that came out, and we can just delay the pedestrian and make that person come out later. And the car will have to react to it as though it's real. That's right. Like, you're going to run lots of training against lots of different variations of the same data. That's fascinating to me. I understand why all the car makers would buy into that.

38:20Nilay Patel:Why would they buy into the data sharing? Is it just a recognition that collectively they stand a better chance of catching up? Is it they just don't want to pay the money? Is it cheaper? Why would they participate in their competitors in that kind of data sharing arrangement? Both are absolutely true. And actually, the cost saving is enormous. is basically data collection running a fleet of huge size. Essentially, it's a big, I would say, capital spending for anybody who wants to do that. And also, it's kind of repetitive as well. If you can find, for example, what we provide in the Drive platform or the Drive ecosystem, it can save a lot of effort and basically money from our customers.

39:11Nilay Patel:I'm curious about that because the idea that you're going to train stuff and then you're going to have a model in the car and we'll have an AI-defined car. The sort of classical approach to self-driving was we're going to throw more and more data at the problem and eventually the car will kind of know how to do everything and it will have mapped all the roads on top of everything. So you're going to – I have a Cadillac EV and the way Super Cruise works is it works on roads that are mapped. and eventually, you know, the bet is they'll map more and more roads and more and more things and the car will become more capable.

39:41Nilay Patel:It feels like NVIDIA's approach is for the car to be smart enough to do anything with or without the maps. And that requires a different approach to data collection, a different approach to commute, and then obviously a bigger bet on AI. Is that split real? Have you just made that jump? Is that the future of the platform or are you in the middle? Well, the approach we take right now for what we call the L2++, essentially it's mapless. As you said correctly, so basically the model would definitely need more data and to cover more corner case. And the model is obviously getting bigger as we speak as well, basically.

40:16For this generation, next generation, we are going to use a much bigger model with more parameters. And also foundation models will play a big role here. and being able to make this model very capable, essentially more data is very, very critical. But on the other hand, though, the trend of using foundation model, which is already trained with internet data, that can help, come in help as well. That's why I emphasized quite a few times on the connection with the foundation model effort inside NVIDIA. With the reasoning model and the foundation model, So these are the things that we can leverage from the, let's say, the frontier model perspective and leverage internet to basically kind of scale data to be able to help the vehicle to generalize better even without vehicle-specific data.

41:14So this is one of the, I would say, the main direction we are betting on towards, let's say, higher level of autonomy, especially level four. This is one of the main work threads we are focusing on right now. Back to OEM, I think being able to leverage basically what we have built upon through our collaborations with existing basically engagement and our massive capability of basically data generation using a synthetic data set and the neural reconstruction. and also being able to leverage the foundation model capability, which are trained from more general data, but which will help the model to reason better, to generalize better.

42:06These are the things we can offer to our customers.

42:09Nilay Patel:I feel like I have to ask about safety now. I'm sure it's more complicated than this, but you're talking about a foundation model reasoning through self-driving, and all I have in my head is chat GPT apologizing to me because it got it wrong, you know, while the car crashes or one of those horrible long latency loops where the model goes off in the wrong direction and realizes it. And then like, you can look at the chain of thought and it's like, oh, it got it totally wrong. It's, it feels bad, you know, in the way that like Anthropic believes that Claude feels bad. None of that seems compatible with the very real-time nature of driving a car.

42:51Nilay Patel:How do you bridge that gap? Latency, the need to have one of those big models in the background, the sort of reasoning tangents that the models can go on, how is that compatible with driving a car? Safety is so important to us and obviously so important for the AV industry. So let me answer your question from our kind of approaching different layers of offering. So to address safety, this is obviously not new for the auto industry. And we have developed very sophisticated, basically, even development protocol and also validation protocol to be able to prove this software is safe. That's called ISO 26262.

43:35And, you know, we actually develop our hardware and operating system, OS-level software and application-level software, you know, to the highest standard, which is very important, which is very critical to be able to deploy anything, you know, to drive the car. That's number one. And number two is basically we take a slight different approach than some of the, you know, players in this space. We actually have a redundant stack, even for our L2++ or ADAS basically function. Other than the N-to-end model, which is basically you have pixel in, you have trajectory out, we also have a classical stack.

44:19Classical stack means it's more developed based on this safety standard as we know it. with a component basically. It's a stack with many components and each component can be verified using this known standard. That's what I refer to as a classical stack. And when you have two stack basically kind of run in parallel, the classical stack is acting like a, sometimes we call it the big brother, but essentially it's a safety guardrail. try to verify all the trajectories from the end-to-end model and use it, you know, use the, let's say, known safety standard to verify it's safe at every frame. So that's a very important concept, you know, we have, and not only concept, but the implementation we have in our stack.

45:14And we will take this, obviously, this will be so critical for higher-level autonomy, L4. so this is also the foundation of our kind of L4 stack where we have full redundancy not only at the sensor set but at the software architecture set so this is I would say the second point I want to make to answer your safety question and the number three also basically when we develop the model basically we are also trying to make the model reduce the hallucination as much as we can right so the way to do that is really basically through massive validation um you know we are looking at we are building basically massive simulation test data set for every model uh we we release right now we're looking at in our program right now we are running 5 million basically tests every day and obviously roughly every day we have 10 iteration of the model the end-to-end model up model.

46:16So we're doing really massive validation to make sure in all these scenarios, you can think of that every test-to-test scenario, the model is generating the right trajectory. So that's also super critical for us. So this is what we do to make sure our product is safe.

46:36Nilay Patel:Let me ask you a really dumb question I'm really curious about. You've talked a lot about the model and how it will operate the car. And yes, the classical stack is the safety guardrail. Is the model reasoning in language like every other model? Is it sitting there in the background saying, I see a stop sign. What do I do? I'd better stop. I'm going to go hit the brakes the way that any sort of general model reasons in language in the background? Sure, the answer is yes. And in our next generation model, which we are going to deploy in the next generation of vehicles, because the current generation is on Oren, which has more or less more limited compute.

47:14the next generation is SOAR-based, we will have the model trained with language embedded. So being able to reason through language is very important. And also you can chat with the model. You can ask the model about what he's doing, and you can also ask the model to speed up or slow down and make a lane change, for example.

47:35Nilay Patel:As it's literally driving, it's saying to itself, I see a car over there. I need to change lanes to get ready for the exit that's coming in a couple miles. And it's doing that in language to operate the car? I think it's a combination of things. Language is already embedded in the model, but the vision signal is also super important, as you know. So I would say it's multi-model, but language is part of it. Obviously, as you know, the model is black box. We don't exactly know basically what it is exactly doing, but you can ask about it, and then the model will answer what it's trying to do. I just have this vision of a chatbot model just freaking out.

48:14Nilay Patel:It's on the highway at 55 miles an hour. FRANCESC CAMPOYOVSKYI I think Jensen did. It released a video that the model is talking constantly. It can be quite annoying as well if you really try to hear everything the model is trying to reason about. MARK MANDELAVYIERIENCY, JR.: What's the latency on that? I mean, obviously, you're deploying the systems. It must be working. But is there an attempt to reduce the latency of that? I feel like language is inherently slow compared to what you need to do to drive. Like, I'm not thinking in language when I drive my car. FRANCESC CAMPOYOVICIERIENCY, JR.: That's why I said it's multi-model, right?

48:51But to reduce the end-to-end model, end-to-end latency is super important. Actually, that's one of the key advantage of deploying, drive the car with a model. Because if you think about it, the old basically stack or the classical stack which has multiple components, it usually basically takes multiple hundred milliseconds. But with the model, because it's just inference time, it's separated between input, which is pixel, and trajectory. You can reduce the basically, depends on the compute, obviously, capability you have. But even in the current generation, we can control it to be within a hundred milliseconds, which is pretty fast.

49:26And regarding the language reasoning, obviously, See, if you think about it, well, that's human brain, right? But if you think about the language, basically, I would say the information rate is already abstracted. The information rate is not super high. And we are obviously using the internet data to train this kind of language-based reasoning capability. I think the latency is well under control. Let me put it that way. And again, you're not driving the car with language only. That's a key thing, as I said. Usually the reasoning part is, I believe it's slower. Again, we don't know exactly what the model is doing, but the pixel part, that's what drives the instantaneous kind of reaction of the vehicle.

50:09Nilay Patel:If you ask Anthropic, they will tell you that Claude has feelings and emotions and it can get scared. Do you think about that? Do you think your models have emotions when they're driving the car? We will use the guardrail to make sure it doesn't get too moody. I'm just curious. I mean, like you said, we don't know how the models are working. I just, I literally have a vision of the model being like, oh my God, I'm going so fast. But maybe the classical system will cut that down. Yeah. We have to take on a short break here. We'll be right back.

50:50Nilay Patel:Support for the show comes from Upwork. If you're looking to grow, you need people who can take your business to the next level. Upwork is a one-stop platform to find, hire, and pay expert freelancers across more than 125 categories, launch projects faster, and scale support up or down without committing to full-time headcount. You can browse profiles, review people's past work, and get help scoping the role. Upwork also has Business Plus, which lets you access the top 1 % of talent on Upwork. With AI-powered shortlistings, you'll get matched to the right freelancer in under six hours. And when it comes to operational stuff, like contracts and payments, Upwork has a handle on it all.

51:31Nilay Patel:That way, you can actually spend more time at the helm. It's free to sign up, and posting a job is easy. Visit Upwork.com right now and post your job for free. That is Upwork.com to connect with top talent ready to help your business grow. That's U-P-W-O-R-K dot com. Upwork.com. Support for the show comes from Whatnot. Consistency is key. It's what can take a business side hustle into something that lasts. And with Whatnot, you can turn that side hustle into a real business. Whatnot is the largest dedicated live shopping platform. Whether it's beauty, collectibles, electronics, luxury fashion, even cookies.

52:15Nilay Patel:Sellers are building real, thriving businesses. Anyone can sell. whether your business is big, small, or yet to exist. And people selling on Whatnot sell 10 times more than on other major marketplaces. That's because you're not just listing products, you're building real connections with buyers. And the company has seen that Whatnot buyers spend more than an hour a day in the app. And that time is spent more than just browsing. They're engaging, buying, building community, and coming back. As a seller, you go live, show off products in real time, and turn what you love into real income. We have a colleague who's tried Whatnot out and really enjoyed it.

52:53Nilay Patel:Search Whatnot, W-H-A-T-N-O-T, in the App Store. Download and you can start selling right away.

53:04Nilay Patel:Support for the show comes from ServiceNow. AI was supposed to handle the parts of the job you hate. Instead, it just describes them, suggests what to do about them, and then leaves you to do it. That's not help. That's homework. ServiceNow's AI specialists are different. They're not a tool. Think of them as digital teammates who actually do the work from start to finish. Cases get resolved. Requests get processed. Loops get closed. And most importantly, no extra work for you. Because when you can truly delegate to AI, you can get back to the work only you can do. The work that requires a person with ideas and judgment.

53:44Nilay Patel:and, you know, a pulse. To learn how to put AI to work for people, visit servicenow.com.

53:59Nilay Patel:Welcome back. I'm talking with Jinju Wu, NVIDIA's head of automotive, about how all these AI systems very literally work. Is all this running locally in the car? No, no, no, no. All these validation offline. But the second part, with the safety guardrail, When we run two stack in parallel, that's definitely in the car. And in the car, at every frame, the software in our ADAS ECU, we are comparing basically the trajectory from both the classical stack and the end-to-end model to make sure the model is outputting a safe basic trajectory. So do the cars require fast connectivity to be autonomous with your approach?

54:39Not necessarily, but we do require some connectivity to get navigation information and some map information. Most of these are a navigation map. So not only the model side, and also the classical stack, which we do use some of the navigation map information to help us understand the world better, essentially.

55:03Nilay Patel:I'm only asking because I covered the launch of 5G networks in great detail. And all of the telecom companies promised me that 5G would enable autonomous cars. And it seems like your approach is the one that will lean the most heavily on low latency networks in that way. Well, which is that this is not wrong. But on the other hand, basically, the car has to drive autonomously in completely blind spot as well. Real-time, basically low latency, I would say content dependency, have that dependency in the cloud. At least for the ADAS kind of application, L2 +, which is meant to work everywhere. Building that dependency is not a good idea.

55:55Nilay Patel:When you get to level four, level five, that's when you have the connectivity dependency. That's right. Yes. Yeah. What happens when you lose the connectivity at level four autonomy? When you're at level five and you don't have a steering wheel anymore and you lose connectivity, what happens? You can think of connectivity as kind of a sensor. And again, the basic driving capability cannot have a huge dependency on that. One of the core concepts of developing level for a technology is you have a sensor redundancy. That's not only for basically GPS, but also for camera, radar, everything you see.

56:29For every single point of failure, the car has to be able to drive safely. It's like you suddenly lost your GPS, but the car with local perception needs to be able to get to a safe point and pull over. That's the minimum requirement an L4 system needs to have. So this is just a DL4 basic principle to be able to develop such a system.

56:56Nilay Patel:I'm very curious about where all of the sensor stacks live in the car, how much compute is in the car, how much RAM we need to put in cars at a time of increasing RAM prices. This all seems like a lot of extra cost to layer into cars which are increasingly getting more expensive and which consumers, at least in the United States, feel like they're rebelling against in lots of ways. I can look at our own website traffic and I'm like, everybody wants to buy a Slate truck for$25 ,000 and it doesn't even have a radio. right like that's just a that's just a battery and wheels that that's that whole car it doesn't even have paint job right we're we're getting rid of paint jobs on the cars now to keep the cost down you're talking about a lot of compute in the car a lot of connectivity maybe a bunch of ram to load the models on that's right how does that play out is that push you more into that robo taxi model or do you think people are just going to buy expensive self-driving cars definitely building autonomous can't need a lot of hardware but the other trend of the hardware cost is i would say is dropping pretty rapidly as well as the technology become more mature for example radar right even in my career basically i have seen a radar price probably drop by at least four or five times over 15 years because the volume just getting much bigger and bigger than basically the cost.

58:17I see, I have witnessed basically the drop of both sensors. Actually, camera sensor price drop as well. There's more competitors and the more competition, and the competition bring a lower price when the volume become bigger. The scale effect is definitely there. Right now in ADAS, and all the components become much more and more basically mature and to some level commodity. And on the computer side, as you know, the computer is growing at such a rapid pace. So we talk about Moore's law. uh you know in the semiconductor industry you know sometimes ago but in the auto basically uh uh segment uh in the autonomous driving segment the compute has the computer need has been growing basically at a really astonishing pace roughly we're talking about 10 times every two years it's insane with the success of ai and obviously nvidia we will be able to provide this kind of massive computer cars at affordable price.

59:18Nilay Patel:But in the cloud or in the car? In the car. In the car. In the car. I asked you about fighting for training capacity earlier. Do you have to fight for fab capacity too? Because those costs are going up for everybody. Yes, of course. Yes. But I'm curious, it's NVIDIA's demand that's driving up the cost for everybody. So how do you go get fab capacity when the other divisions at NVIDIA are willing to pay whatever rates anyone demands? Well, the same answer I give you, right? Again, you know, I don't know if there's anything I can say more, right? Because, you know, we are such a strategic company.

59:59And, you know, our automotive business is doing well as well, but not at the pace of our data center business doing, obviously. but basically we are strongly vergensen himself as well of the av future and we are keeping investing basically in this technology and in this future not only from you know allocating external computer but from from from fab capacity as well so um so but that's definitely one of the things we we are looking into actually most likely uh the even the chip price might need to go up because of this intense kind of demand for every chip. Everybody can grab on, essentially.

1:00:39But the positive side is basically the technology is really getting, you know, I talked about the chip side. And also, I talked a little bit about sensor side. And we are looking at basically, for example, I talk about Hyperion, which is basically kind of product ready, compute plus sensor kind of platform. So we are looking, we are really trying to balance between the cost and what we can do. We are looking at what we call the, you know, sufficient necessary kind of sensor set to achieve high level of autonomy. So in Hyperion 10, for example, we really offer two versions. One is a base, which is mostly camera, 10 camera, 3 radar, no LiDAR.

1:01:27And, you know, it's a very cost-effective way to build a basically kind of L2++ ADAS kind of vehicle. And on the other hand, for the high end, what do we call the Hyperion High, we provide basically, you know, the sensors that required, which have like, I think, 14 camera and 3 LiDARs. and basically seven radars essentially to be able to drive, have enough sensor redundancy to be able to drive L4. We also provide, you need an ECU redundancy as well. You need two basically kind of our next generation, well, actually to be more present, current generation SOAR based on kind of computer platform.

1:02:09But just imagine basically you have a car really can drive by itself. We believe with this sensor set and this computer basically architect, we can get to that level of autonomy, which can basically justify the cost.

1:02:24Nilay Patel:The minimum sensor set for autonomy feels hotly debated. It's been hotly debated for a long time. I think Elon Musk saying that he thought LIDAR was a local maximum ages ago was the beginning of this debate. This debate has not quelled in any way, shape, or form. Do you think level four requires LIDAR? The short answer is yes. We believe that LIDAR is the important sensor, you know, to provide the safety and the redundancy required for level four autonomy. But on the other hand, you know, it's difficult to say it's 100 % necessary. We believe this is a a very much feasible path to, based on, as I said, high sensor configuration to get to really high level of both urban and highway level four capability.

1:03:20On the other hand, you know, theoretically, you know, people can prove out with massive mileage, essentially, to say that LIDAR may not be, you know, necessary, But it will come with ODD limitation, essentially.

1:03:38Nilay Patel:Sorry, what's an ODD limitation? ODD is basically applicable, basically domain. You can deploy the technology. Obviously, we have done quite a bit of analysis on this. Based on our current understanding and the framework we use to do this analysis, we believe that to deploy this L4 technology in all the ODDs that our customer can benefit from, it's much better to have LIDA as compared to not having it. When you look at where Tesla is with full self-driving and their vehicles and their absolute commitment to being a vision-based system, do you think that they are currently ahead of you? Do you think they're at parity?

1:04:19Nilay Patel:Do you think they're behind you? So, you know, there's two levels of answer, I guess, to this question. And obviously, basically, for the basic L2++ technology, Elon is probably ahead of everybody, essentially. He has the vision a long time ago, and he has sticked to the vision for a long time to be able to develop and test the technology among massive fleet. Nobody would argue that Elon is ahead of everybody in the L2 or basically ADAS kind of market. And everybody is playing a catch-up game, essentially. And we are very happy, actually, Yilong is so successful. And also, obviously, Yilong is a big customer for us as well, for both SpaceX and Tesla in the GPU computer side.

1:05:15And we are supporting him and his team to make sure they're successful. And for level four, essentially, I think it's more open, I would say, because obviously there's established players who are proven to the, who are already basically like Waymo, who are doing basically already taking customers to really experience the L4 kind of experience using the methodology they use. And Tesla is probably still trying to find the path there. And again, we don't try to pick winners, but we try to help everybody to be able to develop that technology And our mission is really try to make the AV ecosystem get to this vision of everything moves that will be autonomous.

1:06:10This kind of vision becomes a reality.

1:06:13Nilay Patel:Have you had conversations with Tesla executives about using LiDAR? It seems oddly religious for no reason, especially if the costs are coming down, as you say. At some point, if the better technology solution is right there, it feels like everyone should just use it. Have you had those conversations? Well, actually, no, not myself. My team definitely has. And, well, I'm looking forward to have that conversation with them. Actually, I would like to, you know. Anyway, so as I said, much of this is just basically science and basically reasoning. So it's good to hear their view as well. I want to wrap up by talking about something that maybe is the least in your control.

1:06:56Nilay Patel:Models are going to keep getting better and video is going to keep making chips. Maybe customers are going to keep demanding self-driving. That all feels like something you have a handle on. But the auto market, the cutting edge of the auto market is happening in China. I think we can just agree on this. U.S. consumers open TikTok and see car influencers talking about BYD vehicles, and they complain in the comments that they can't get those cars. I watched a video of a Buick that is in China. It's a Buick EV that you can't get in the United States, and U.S. customers are furious that Buick is making better cars in China than they're making here.

1:07:29Nilay Patel:There's a lot of trade barriers between the United States and China. NVIDIA sits in the middle of that fight in all kinds of ways, whether it's tariffs on imports of car components, whether it's literal blocks on what chips can be sold and where the revenue from those chips go. As you try to push the car market forward, how does the U.S.-China trade chaos play into it? Is that something you think about? Is it something that's slowing the industry down? Is it something that you can push through? Well, basically, well, I certainly believe the, you know, policymakers, they have their reasoning and basically rationale to make the policy, you know, as we see right now.

1:08:15And, you know, as NVIDIA, again, we are open ecosystem player. We still have a lot of customers in China. We try to basically help this. For example, we are still supplying in-car inference chips because they are still basically below the threshold of what GPU is allowed to sell in the China market. And then basically we are also working with all the Chinese OEMs, actually not all of them obviously, but quite a few of them to help them on the infrastructure side by supplying them basically simulation tools. And we're working with them on open source models, Cosmos and Amayo. And then basically we can, on one hand, we can help them to get their models better.

1:09:10On the other hand, we can also learn from the competition in the China market. Obviously, we are also working very closely with the rest of the world, basically OEMs, and try to supply, you know, all NVIDIA basically platforms and at different layers to different OEMs and help them to be successful as well. So again, we don't pick winners and we try to basically work with everybody. and the mission is super clear and we try to make AV, this vision become a reality as well as possible.

1:09:48Nilay Patel:When you talk about sharing data between OEMs to train the models better and to make them more capable, are there any regulatory roadblocks or competitive roadblocks between sharing data from Chinese OEMs and American and European OEMs? Oh yes, of course. So we have to live with the original basic, Actually, not only China. Actually, other regions have restrictions as well. For example, Europe has certain regulations regarding data. So we are conformed to all the local kind of regulation to make sure we are compliant to all the basic regulations we need to be compliant to at different regions.

1:10:22Nilay Patel:Does that mean the regional variants of the models have different capabilities or they're better at different things? Because if the input data is different, it seems like maybe the output will be different as well. Absolutely. Well, first of all, for the production model, we try not to basically fork it as much as we can. But there will be basically original kind of a difference. So the model will behave differently in different regions based on the input. And some of the things are what we call the countercoded. it. So you have to, obviously, the rules are quite different in different regions, like in Europe, as compared to the U.S.

1:11:04Some adaptation is required, and some parameters are different as well. Yeah, so it's quite an interesting journey trying to scale the technology into definitely different parts of the world.

1:11:15Nilay Patel:Do you think that based on the different regulatory approaches, the different data approaches, the different input data, the different configuration of the OEMs and what they're willing to invest in, the different subsidies from the governments. Do you think China will get to level four as a mainstream self-driving experience first? Because if I had to look at it, I would bet that level four self-driving will happen in China way before it happens in the United States as a mainstream experience. I actually don't think that's true. As you know, basically Waymo is already getting everybody to have a four experience, at least in certain ODDs in San Francisco, and they're scaling pretty fast.

1:11:56And China is, you know, they're obviously, it's a much more dynamic, competing kind of a market, and there's quite a few players there. But my experience in all of them has got to the maturity of Waymo, at least in San Francisco. But again, we're trying to help everybody in the ecosystem. So from OEM perspective, it's a different competition landscape, but even basically on the OEM side, I think, you know, different regions have different kind of, well, one side is probably, you know, the China streets is also much more, much more challenging as compared to the US streets. So, you know, to be able to, and the level four, I would sometimes call it a zero one game.

1:12:46You know, either you have it or you don't have it. Actually, as of today, I think the only one who really have proven that L4 can be safely deployable to every customer without driver in a kind of city kind of size region without any limitation is still in the U.S., not in China.

1:13:06Nilay Patel:Yeah, that's Waymo. I think Waymo is going to be very flattered to hear them described as a mainstream experience. I will accept that for some subset of people in San Francisco, Waymo is a mainstream experience. I think for the vast majority of Americans, it is not yet. And that is the big turn, right? When can a Waymo work in the snow? When they're going to deploy them in Chicago? I'm somebody living in Chicago for a long time. I'm very curious how that goes in Chicago and New York City, right? The question I have is the mainstream experience feels like you just buy a car and just like level two ADAS is kind of a commodity in cars now, level four will be a mainstream commodity in cars.

1:13:45Nilay Patel:You push the button and starts driving itself. How far away do you think we are from that? Well, first of all, that's exactly my mission, you know, trying to help the industry to get there. I would say if I need to give a time, I would say five years, less than five years. Well, that is a bold prediction. I think we're going to leave it there because we're at time. You've been really great, Jinju. I'm excited to talk to you again. We'll have you back before five years to check in on that prediction. But what should we be looking for next for NVIDIA? There's quite a few things we are planning. So first of all, by I think end of this year, we are rolling out our technology on the basically ADA side in all Mercedes vehicles and some other partners as well to all over the United States.

1:14:36And also basically, you know, starting for the next few years, this technology we're trying to roll out to the rest of the world and meanwhile basically we are also working closely with uh for example uber we announced that in gtc try to basically roll out our l4 basically uh kind of service in the next few years uh it's super exciting and on top of that obviously we are again a ecosystem player we are working um you know with almost like all oems Right now, I would say 80 % of the mass production OEMs are in NVIDIA's Hyperion, basically ecosystem file for. So we are really building with this future with everybody.

1:15:17So this is hopefully you'll see more exciting announcements from us somewhere down the road.

1:15:23Nilay Patel:Yeah. Well, I guess I will have to have you back soon. Thank you so much for being on Decoder. Thanks for having me, Nile. It's very nice chatting with you. I'd like to thank Jinju Wu for taking the time to speak with me and thank you for listening. I hope you enjoyed it. If you'd like to let us know what you thought about this episode or really anything else at all, drop us a line. You can email us at decoderattheverge.com. We really do read all the emails. Or you can hit me up directly on Threads or Blue Sky. We're also on YouTube. You can watch full episodes at DecoderPod. It's the same handle on Instagram and TikTok.

1:15:49Nilay Patel:Check out those platforms. They're a lot of fun. If you like Decoder, please share it with your friends and subscribe wherever you get your podcasts. Decoder is a production of The Verge and part of the Vox Media Podcast Network. The show is produced by Kate Cox and Nick Stat. This episode was edited by Xander Adams. Our editorial director is Kevin McShane. The Decoder music is by Breakmaster Cylinder. We'll see you next time.

1:16:34Nilay Patel:to try stamps.com risk-free for 60 days. Stitch Fix, stop shopping, get styled. Not today sweatpants. Somebody's wearing jeans that fit. Wow! No photos, please. I'm just a regular dad who happens to have a stylist. I really look my best when someone else makes the decisions. Hey, we can all see you two-way mirrors. Just share your size, style, and budget, and your stylist sends personalized looks right to your door. Stitch Fix, get started today at stitchfix.com. I want to hug you. I'm going to hug you. I'm coming in for a hug. In the U.S., there's a break-in every 26 seconds. But when intruders step near, SimpliSafe Home Security steps up.

1:17:14Nilay Patel:Stop. This is SimpliSafe. Police are on the way. Using AI alerts, U.S.-based live agents help deter break-ins. SimpliSafe, no long-term contracts. Save 50 % on your new system with professional monitoring at simplisafe.com slash Spotify or with promo code Spotify. Outdoor deterrence requires a SimpliSafe Active Guard outdoor protection plan starting at$49.99 a month. Visit SimpliSafe.com slash licenses for alarm license information. Tennessee 2012.