Thanks for joining us. Jensen and I are here to go through any questions that you've had, both what you've seen today in our Paris GTC, but also we have done several other GTCs over the last couple of months adding up to that. Probably the most important thing to understand moving here and doing so much of this here in Europe in the EU and France, as well as our time that we spent in Taiwan, is to really emphasize that AI is here worldwide. The importance of seeing what is going to be possible for AI. This is an area that is growing faster than any other technology in history that has reached every single region at the speed that it has done.
What that's going to take, though, is the influence and help both of the sovereign nations, the sovereign help that we're going to need from a government to really expand both here in the EU area in France to do so. I know you saw so much of that today with Jensen, but we'd also like to talk more just from the investor standpoint in terms of what we're seeing. We're going to open it up for questions unless you want to start with some moving equipment.
Nope. Great to see all of you. Make it nice and loud.
Good afternoon, Jensen Huang.
Hey, CJ, the pregame show.
Yeah, yeah.
I understand he was quite a hit.
I don't know. We'll see later. Thank you for taking the question, Cantor Fitzgerald. Two-part question. First, your commentary on quantum computing seemed to change a little bit. Curious, where do you see commercialization first? Secondly, on the sovereign front, you've been traveling throughout Europe. I think your travels continue beyond France. Would love to hear how your conversations have gone and how you think about the magnitude of coming investments relative to what we heard from the Middle East. Thanks so much.
Yeah, appreciate it. First of all, my feelings about quantum is consistent with the past. However, my feelings about quantum classical is very different. I think that the entire industry is now recognizing that quantum classical is the way to go. It's not about a standalone quantum computer. It's about a quantum computer connected to a GPU supercomputer to do all of the controls, to do the error correction. The groundbreaking work that's being done in error correction is really quite significant. Basically, if you look at a qubit today, a logical qubit is represented by a cluster of physical qubits. You guys, anybody, am I talking weird stuff? These physical qubits, it takes a whole bunch of them working together, entangled together to represent a logical qubit. Then you have a bunch of Ancilla qubits, which are basically shadow qubits.
Because as you know, the Schrödinger's cat problem, if you observe, if you try to measure the quantum qubits, the qubits, it collapses the state. It loses coherence. It is either no longer in that superposition state. It will either be on or off. It will be the Schrödinger's cat. It will be dead or alive. But it is never in superposition. The recent breakthroughs in using error correction require a lot of computing outside the quantum computer. We are making really great breakthroughs there. If you look at the GPU supercomputers that are going to be connected to these quantum computers, they are going to be giant just doing the error correction stuff. If we keep going at this rate, let's say that we get 10 times as many logical qubits every five years. We will probably have something close to 20 logical qubits-100 logical qubits in some five years.
100 logical qubits, just the number of the amount of state that it could represent is sufficient to do some early biomolecular or chemistry stuff, material work that could be quite useful. The way that we're going to do it, and this is the reason why I think the community is getting together on this idea, that instead of using the quantum computer to do all the simulations, what we'll do, we're going to use quantum computers to generate ground truth. The electron simulations of you behave like a quantum, behave like an electron state, and then generate a whole bunch of synthetic data that we'll train AI models with. Are you guys following me? This quantum classical hybrid is gaining a lot of momentum right now. I think everybody's getting excited.
We can kind of see it being kind of two or three years out doing some real work. In the meantime, what I said is true. Every single supercomputing center is going to go quantum classical, 100%. I've not met one that's not going to go quantum classical. That is why CUDA- Q is such a revolution. We're basically working with everybody in the quantum computing industry on CUDA-` Q. With respect to the buildout here, it is much more for local use, indigenous use. Middle East was some indigenous use, but it's really about hosting the cloud for American companies. It is related, not exactly the same. Does that make sense?
Most of the stuff that we're talking about here, the telcos, the regional cloud service providers, the 20 gigafactories, the AI factories that's going to get built that's supported by the government across pan-European countries, that's all being built for local consumption. I think that long term, it's just going to represent sovereign AI should represent the GDP of the countries. In the case of Europe, it's taken longer to get engaged. The reason for that is because their information technology industry is lighter than the United States, but their heavy industry is much bigger than the United States. That's the reason why robotics is going to be such a big deal here. Industrial digital twins is going to be a big deal here. All the factories are going to be digital. AI is going to be everywhere in those factories.
That's the reason why the topic here is quite different than the topics in the U.S. Overall, the world's regions combined, we estimate over the course of the next several years about $1.5 trillion with the buildout. Once you cobble up all the math, it kind of makes sense.
Joe, Morgan Stanley, you talked a lot about physical AI today. Can you talk about what we should look for to sort of see the model development? Is it going to be the startups developing physical AI capabilities, new models, or are you actually seeing those physical AI? Is that information getting incorporated into the LLMs that are already foundational?
The physical AI models are going to be different than the LLMs. They're going to be multimodal. Like, for example, you'll walk up to a robot and just tell it to do something. Just as generative AI can generate pixels just by your prompts, you should be able to generate motion from the prompts. It'll reason about it. Just like generative AI right now can reason about the prompts and reason about the pixels before it generates the pixels, you can now reason about the motion before it generates the motion. You could see the robot thinking, "OK, I've been asked to put this apple in that drawer, but the drawer is not open. First I have to open the drawer.
I got to pick up the apple, put the apple in the drawer, close the drawer." That reasoning process, you can kind of see it happening now, right? You guys see it in GPT-3 or DeepSeek, or you can see the technology exists to do all that. My sense is that the countries here, Germany has a lot of robotics capability. France has robotics capability. The U.K. has. I mean, because the heavy industry is quite—the Nordic countries have a lot of robotics, ABB, for example, lots of robotics capability here. They've just been missing the software capability, if you will. Previous generation robots are all pre-articulated. Do you guys understand what I'm saying? Pick this thing up from here, put it over here, 100% every single time. It's pre-programmed. Now you don't have to pre-program it.
You just tell it to do it. As a result, robots will be, robotics will be much more accessible to the smaller and medium-sized companies.
Yeah, hi, it's Brett Simpson at Arete. Jensen, I just wanted to ask about these gigawatt projects that are being announced. It's fairly a new concept to us all here. How many do you have line of sight looking into the next two, three years? How many gigawatt projects do you think are already underway? I guess there's one coming in France. It's been announced already. Give us a sense. In your presentation earlier, I think you said there were five European gigafactories. Is that five separate gigawatt projects? If you can help just give us a sense for the scale of these, how many gigawatt projects you have.
Yeah, you got it all right. We have line of sight towards the telcos, the regional cloud providers that I mentioned. For example, Mistral is the one here in France. In the U.K., it's Nscale and Nebius. In Germany, they changed their name from iGenious. I think it's called Donym or something like that. I thought iGenious was pretty good. I'll go find out why they changed. These are all line of sight. The 20 that are none of these are supported by government. These are all business-oriented startups or scale-outs. The ones that are supported by government are the 20 AI factories, and a handful of them are gigafactories. That is what we have line of sight on at this moment. There will probably be more. If you just kind of added up all of everything I just said, it is lower than the GDP representation of Europe.
Now, of course, for some time, the American cloud service providers will come and serve that. Over time, maybe the regional cloud providers will get larger and larger. Because you have sovereignty issues with respect to data privacy and generally people being concerned about geopolitics these days. You kind of want to have infrastructure locally in each one of these countries. There's a reason for some of the buildout.
Thank you, Lou Miscioscia , Daiwa Capital Markets America. Maybe if you go into what the limiting factors are both for you to produce more of your products and then from a small scale and then from a big scale, you mentioned that maybe some European companies do not have the software ability. What also could just drive additional demand for all this AI stuff that we see at your conference here, which is pretty amazing? Thank you.
The supply, none of the supply is horribly difficult to get now. It's constrained, but we're still growing fairly fast. Nothing is sitting around. We don't have a whole bunch of Blackwells and CoWoS, ELs, and these supercomputers sitting around. They build what we ask them to build. We have to forecast it. We're not limited by CoWoS. We're not limited by HBM. I just have to forecast it. Our lead times are probably more than a year. From the time that I start wafers on Blackwell to the time I ship a supercomputer out the door, it's coming up close to a year, which is a real advantage for us. The reason for that is because I have a better feel of the total consumption in the world than just about anybody.
I could place a giant order on TSMC and Micron and Hynix and Samsung and SPIL and Amkor and Foxconn. I mean, our supply chain is massive. I could place a few hundred billion dollar order on our supply chain because we have great confidence of the end market and the fungibility of our products everywhere. If we were somehow bespoke, if you will, it's only useful for this customer, then it's harder for us to have the confidence to build for the whole market. Our confidence, NVIDIA is everywhere. We're not so much limited by any critical component, per se. It's just everything we build is not easy, per se. We just have to forecast it. In terms of the end market, there are several things that limit the end market. One of them is just local languages.
We think that everybody should speak English, but they do not. Some people prefer interacting with their devices in their native language, which is very understandable. Of course, if you want to reach the whole population in Israel, and Israel comes to mind because I was just working on it, you are going to need a large language model trained in the language and the data and the customs of Hebrew. The same with Arabic countries and so on and so forth. You just multiply that out. If we want AI to be successful in each one of these regions, the technology, the agentic technology is there, but the reasoning AI language model needs to exist. I was talking about that today. All of those partners of ours that are going to take NVIDIA's Nemotron and optimize it for their local language now have a state-of-the-art capability.
They already have the data prepared for all their local languages. We'll fine-tune each and every one of them. Each one of them will probably take about a month of supercomputer work, but it's not so bad. We take that model. Now we have to connect it into a search system, and Perplexity is ideal. We just plug it right into Perplexity, and off they go. That's the idea.
Hi, thank you. It's Timm Schulze-Melander from Redburn. I just had a follow-on, actually, from Joe Moore 's question about sort of model progress. Anecdotally, there's lots of excitement, lots of enthusiasm, but also investors that I speak to who may be on the more skeptical side point to the fact that MMLU scores are topping out, and maybe there's some impatience for more tangible real-world applications. As you work across all of the vectors at NVIDIA, could I just ask maybe for multimodal, large language sort of reasoning models, what are your sort of preferred measures of AI token capability? How do you keep an empirical track?
Really good question. As you know, the reason why reasoning is such a breakthrough is because a reasoning model can solve a problem that it has never seen before. First of all, right? Makes sense, right? Because you're breaking it down step by step, and each one of those steps you know how to do. One of those steps might be go read this document, learn it, come back, and do the next step. The reason why agents are so much more effective than a pre-trained reasoning model is because the agent can benefit from context. Go read this document, and the document tells you exactly how to do it. Come back and do it. MMLU does not do that. An open language model sitting out in free space does not have the benefit of your fine-tuning, your training.
That's the reason why enterprise models are going to be so good. We're experiencing this all over the place. The work we do with ServiceNow and SAP and Cadence, they're all super agents, but they're narrowly super agents. We give them context and retrieval augmented generation. We fine-tune them. We teach them human demonstration. Does that make sense? Our goal is to design a chip. I don't need you to be a history expert. My goal is to do supply chain management. If you don't know anything about taxes, I'm going to survive. Do you see what I'm saying? We take these reasoning models, these agents, and we fine-tune them into the job we need them to do. That's the reason why. Don't worry about these AI models. They're going to get better and better and better, no doubt. Just look at the curve.
It's going to get better. Who cares? My job is not to wait for artificial superintelligence. I just want them to do a good job with my supply chain management.
Thank you, Antoine, New Street Research. Thank you very much for sharing the growth forecasts earlier in Europe in terms of capacity. It seems that Europe alone could be a very strong driver of growth in 2026. That actually got me wondering more generally as we get closer to the middle of 2025, how should we be thinking about growth in data center for NVIDIA next year? Because now we have some hyperscalers who have guided already 2026 for CapEx. Broadcom last week, I think, said that they expect sustained revenue growth into 2026. That means that they're getting some visibility, right? I assume you should also be getting some. Any comments you can make would be very helpful. Thank you.
Everything that I told you guys today is in addition to CSPs, the American CSPs. Most of Europe is underserved today. Even the part that is served, the newest generation does not come out. There are so many developers and researchers that are still using Amperes. They do not even have Hoppers, barely. That is the opportunity for the local CSPs. They could deploy the best as soon as possible. Do not let it diffuse out. You do not have to wait for the public clouds. All of this is incremental.
Hi, Luciano from Impax. I think it's quite clear that you guys are clearly dominating training, pre-training. You made a case today why inference growth is really good for you in reasoning and so on. On post-training for the big models, I just wanted to know what do you think the future for your business model is in terms of not just providing the compute, but perhaps the simulation for these models, a little bit like you're doing robotics, if you see something similar outside robotics as well.
Yeah. Post-training is an excellent opportunity for us because post-training is just a new phase of pre-training. The new phase of pre-training, post-training does this. The first thing you do is you give it human demonstration. We call it reinforcement learning human feedback, human feedback. I give you demonstration, and you try it until I tell you whether you did a good job or not. That's like coaching. The second thing is like self-practice, reinforcement learning, verifiable results. I give you a bunch of tests, and I say, these are the answers. I give you the test, the problem, and the answers. Your job is to reproduce the results. You just keep trying until you get it. You know what the right answer is. If you get closer, I'll give you a positive feedback.
If you get further away, I'll give you negative feedback. That could be used for coding. The results are very verifiable. It could be used for a science simulation. There is a whole bunch of tools we've already created as humans that are excellent at providing the feedback, the ground truth. That's called reinforcement learning, verifiable results. All of that requires a ton of training. You just crank forever. The amount of training you can do is as much time as you have. How much human practice can you possibly do? You can practice as much as you like. That is post-training, very big deal. Yeah, uses a lot of compute. Basically, as much time as you have compute, you just have to decide when to pull the plug. I have to go to market. I can't wait anymore. Tomorrow is the test. You're out of time.
For inference, as you know, right now, NVIDIA is the world's largest inference platform, right? Everybody says inference is easy, but there's nothing easy about inference. It's the hardest thing of all. And we're very successful in inference.
Hi, this is Rolando Grandi from Itavera. My question will be about edge computing. Before, the case was that the computing was on the pre-trained world, we moved to the edge on device, and that's it, right? You mentioned right now that post-learning, reinforcement learning is getting back the processing back to the data center. What about the use cases in robotics, in satellites? You had some announcements there. Elon is speaking about sending robots to Mars. The latency becomes a big issue, right? You need to have that on-device capability. How does that computing on-device or on the edge work in this new reinforcement learning world? Thank you.
The compute is on-device. We have four major edge use cases. One is self-driving cars. Our autonomous vehicle business is already $5 billion a year. It's a big business: training, simulation, and in the car, edge AI. The second one is robotics, and that's just starting to grow and likely to be quite large. The third is facilities. This is an edge computer that sits in a factory, in a warehouse. Those are edge devices. There we partner with Siemens a lot. Another one that you've heard me talk about is base stations. Next-generation base stations are going to be based on the 6G base stations are going to be based on AI. We have a system that's called Ariel. These are our four primary focus areas because the software is very, very hard. The easy stuff, we're not going to go touch.
These four areas are quite hard. The computer's right on the edge. We call it Orin and Thor, two really amazing processors.
Edward Kamenak, piggybacking on this question I also had in mind, I mean, the key device is this one, right? If we really start using, putting more and more AI on the iPhones or the next iPhones to arrive, how would that affect your business model with your concentration on GPUs?
The more AI they use on the device, the more AI you're going to use in the data center. You still have to train the model. You have to develop the model and verify the model, evaluate the model. All of that is done in the data center. Our business is not on the phone. The phone is not our business. There is a lot of innovation there. We do not build modems. We do not build low-power SOCs. That is not our core business. It is well-served anyways.
On the supply side, your key risk is naturally.
The more AI, the better, bottom line. Yeah. The absence of AI is the only thing I worry about.
Okay. Not much.
Yeah. Please, AI.
Not much to be concerned there, I guess. The key supply risk you haven't mentioned naturally is TSMC and Taiwan, right? I mean, I imagine you have emergency plans for that. Can you speak openly about them?
We announced that we are going to build $500 billion worth of AI supercomputers in the U.S. in the next several years, 100%, from chips to packaging to integration into supercomputers. We have partners that are setting up in the United States: TSMC, SPIL, Amkor, Quanta, Wistron, Foxconn. They are all setting up in the United States. We are the largest customer for TSMC. They are very supportive of us. That is the goal. We will manufacture in multiple continents. We will continue to do so in Taiwan. We manufacture in Samsung in Korea, some of our components. We will manufacture a lot more in the United States. Our supply chain is so large, we are manufacturing really almost everywhere.
How would you rate the calendar reducing meaningfully your dependency to Taiwan?
This is it. We're probably going as fast as anybody in the world is going. I think the real truth is Taiwan's pretty important to the world supply chain. Let's avoid conflict. Job number one.
Of course. And then the last one is how would you rate Huawei's forays in AI chip manufacturing? How would you rate Huawei's foray into AI chip manufacturing?
Very good. Very good. There are several years behind us, but for China, it's fine. The reason for that is this. Because their power is so cheap. Not because China is willing to accept less. Their power is cheap. And when the power is cheap, you just use more chips. This is not like iPhone, not like a phone. In our case, the AI chips, our performance efficiency is probably four times theirs, five times theirs. But just use five times more chips. In the United States, it would never fly because the data center is 100 MW. If they have to use four times as many chips, it's not going to fit. They do not have enough power. That data center would be one-fourth the revenues. That would never fly. Just build more data centers, use more power.
That is why our advice is that the export control be lifted so that we can go and compete for that business. Right now, as we speak, I just want everybody to know that we have taken China out of our forecast. We are assuming zero. Because at the moment, we have been banned. We went from a $30 billion-$40 billion a year business to zero. It is a big drop. Thank goodness our demand is so strong everywhere that we are going to continue to grow anyhow. Nonetheless, it is a big loss, okay? The important thing is we are not guessing about China. Are you guys following me? It is zero. If in some circumstance the president negotiates some outcome that makes sense to them, it would be a bonus to us. At the moment, we are assuming zero, okay? Please assume zero. No guessing.
When you're at zero, you don't have to guess, you guys.
Hi, Ardavan Haidari from PGGM.
You have to yell at me.
Sorry.
I will not be offended.
Apologies. Ardavan Haidari from PGGM. I had a question on the reinforcement learning that you talked about. You mentioned cases that are obvious, like math or coding. You can compare it. What about basically all the other cases where you do not have a good sense of what the outcome needs to be? Where the solution is a bit unknown, it is a bit hybrid. What do you see there in terms of reinforcement learning applied to those types of learning models?
Reinforcement learning is really, really good at learning how to do something that is very, very far away, many long time away from the action. I have to take one step, another step, another step, another step, another step. Eventually, I get a positive or negative response. Reinforcement learning is good at that, in fact. It is the reason why reinforcement learning is good at robotics. You say, "Robot, I want you to walk from here to there." You only have two goals. You have to get your head as high up as you can, and you have to move in that direction as much as you can without falling. In order to do that, many joints have to happen, steps, all these different joints. There are many different motions that have to happen in order to get the head up.
Reinforcement learning is very good at these long feedbacks. Yeah. Yeah. That's right. Reinforcement learning is good at that. Yeah. The reward function is very far out.
Thank you again. CJ Muse with Cantor. Thanks for the follow-up. I was hoping you could speak to GB300 transition. I think on the last earnings call, you talked about initial output and low volumes in the July quarter and ramp thereafter. I wonder if there's any more specificity there. Just to follow on your comment earlier about the 12-month lead time from wafer to full output, I guess, how does that inform your customers lining up 200 versus 300? Perhaps is your visibility even longer than 12 months? Thank you.
Yeah. Yeah. I appreciate it. We forecasted the GB300 transition last year, and it's showing up at the same time as we forecasted it. As you know, GB200 was late because we had a bug in Blackwell. But B300 did not have the same bug. And so B300 shows up at the same time. The window for B200 to 300 is shorter because of that. But we were planning for this transition at this time for a year ago now. And so the transition is going just great because it's basically the same chassis. Going from HGX to this NVLink chassis was a huge difference. Everything was different. The mechanical process, the mechanical systems, all the electronics are all different. Testers are different. Even the companies that tested it are different. And the way we tested it are different. We used to integrate these computers at the data center.
We send the computer nodes, each one of the HGX and the CPU trays. We send it to the data center. They integrate it at the data center, and they test it at the data center. Today, that entire thing is tested at an ODM, fully tested, fully integrated. We ship a supercomputer out the door. It is incredible. Even the amount of power necessary in the manufacturing floor went from a few megawatts to tens of megawatts because they are basically building and testing AI supercomputers. Everything changed. GB300, everything is exactly the same. We decided months ago that we would even not change the packaging that sits on the motherboard just so that everything remains the same. That was a good decision. Yeah. We are in great shape in GB300.
This is Joe Morgan. Could you talk to NVLink Fusion and the potential opportunity there? I guess one of the more frequently asked questions I get is, are you making ASICs better with that product? Therefore, could it have any impact on the processor business?
Yeah. First of all, a lot of ASICs are started. Most of them are canceled. The reason for that is, what's the point of building an ASIC if it's not going to be better than the one you can buy in some very specific measure? We're moving so fast. The bar that we're raising is so incredible. It's not easy. It really isn't easy to build. If it was easy to build a Blackwell and say, "Hey, I got 14 guys here. Let's go build a Blackwell." If it was that easy, gosh, I don't know why I'm working so hard. Doing this for 33 years, and it seems harder than ever. Somebody goes, "Yeah, I'll do an ASIC." I'm delighted to hear everybody's interested in building ASICs, all right? I do believe most of them are going to get canceled.
However, many of them have approached us about using NVLink. They're important people to me. The person who's asking me is not a stranger. You guys know that. This person asking me is somebody who had said, "Hey, Jensen, listen. We got a whole bunch of your NVLink systems. We have a whole bunch of your chassis. We standardize on everything here. If we had NVLink, and we could put our CPU in it, say, we could just use the same chassis and extend it. And we'll buy everything else from you. We'll buy everything else from you." You know that last part? "We'll buy from you." You got me out. You buy from me. You kidding me? Of course. The person who's asking me needs help. NVLink is a good thing. NVLink Spectrum X connects into an ecosystem that I care about.
I care about DOCA as much as I care about CUDA. I care a lot about NIXL as much as I care about NCCL. These are all APIs of NVIDIA's that are really important. They do not all run on GPUs. I care about all of my ecosystems. Like I said, it excites me when you say, "I'll buy all the..." It is a super clever strategy. It is just not easy to do. We had to go start a team to go build this thing called NVLink chiplets. We signed up a whole bunch of partners to help us integrate the NVLink into the customers and the partners. We even took our IP and made it available to Synopsys and Cadence so they could distribute it on our behalf. We are going to turn this whole thing into a nice ecosystem.
I think it's going to work out great. NVLink is, as you know, revolutionary. It's really hard.
Hi, Eric Balossier. Thank you very much.
I think somebody said it's just Ethernet.
This morning, you just presented a new product, the NVIDIA RTX Pro server. Could you give us some sense of the size of the opportunity of the use case of the customers you target with this new product? The RTX Pro.
Oh, yeah, yeah. I'm sorry. I got it. The world's enterprise today has no AI. Just go to every single large company. Just look at them. You pick your favorite large company. How much AI did you use in your data center? Almost zero. All these data centers all over Europe, zero. How do you bring AI to that data center? Nothing. Because it's not liquid-cooled. They need to run Red Hat Linux. They have to run VMware. They have to run Nutanix. They want to run NetApp. Does that make sense? It's a bunch of strange things that cloud service providers do not have to worry about. I'm not sure which one's strange. Bringing AI to traditional enterprise IT is very hard. The architecture has to be obedient of the past but innovative for the future. For example, RTX Pro runs Windows. That's pretty crazy.
It runs Windows. It runs hypervisors. It runs all the things that IT managers know. They go, "Oh, that makes me happy." Yeah. Oh, how big is the opportunity? Hundreds of billions. The world's enterprise IT is now just getting AI. That's why Cisco, Dell, HP, everybody is so excited about it. The entire storage industry is standardized behind it.
Back over here. Lou miscioscia again, Daiwa Capital Markets. Maybe on that last answer, maybe you could just point to the two or three things that you just announced today that are the most impactful for the near term as you're trying to drive AI into the future.
Ignoring GB200 and 300, since you guys have already heard about that, we're going to grow hundreds of billions of dollars of GB200 and 300 just once we take that off the table, okay, assuming we already talked about that. RTX Pro, no doubt. It is the first universal AI system that you can integrate into a traditional enterprise IT organization. My IT organization doesn't know how to use GB200. I got to go build a separate cloud for them. Their data centers say, "Red Hat, hypervisors, Nutanix, NetApp." They use phrases like that. These are not AI cloud people. You are not going to change them because they've got too much software running. They got a company to run. We have to augment AI into them. This is the way to do it.
What is availability of it?
It's in production now. Availability now. Yeah. Please tell everybody to buy it.
Okay. Thank you.
Thank you. It's François from UBS. I have a quick question on sovereign AI. I mean, high topics. You have been here in Europe. Is there any control that you can make to this announcement in a way that if you are thinking about all this demand potential, if I'm a country, I want to build as much capacity and as quickly as possible because I don't know where the demand is going to go, but it's going to be big. I need to build capacity big and fast. How do you control that? Because obviously, if you do 20 gigafactory, watt factory, $40 billion-$`50 billion per gigawatt, that's a lot of money.
Are there any milestones that when you work this project, you say, "Well, maybe in a five-year view or maybe two-year view, and then let's see how you are in one-year view," just to rationalize instead of having one big year when you install all this capacity, you can do it more in a smooth manner? I was just wondering how you deal with all this sovereign AI. And yeah.
It gets built incrementally, like you say, anyways. Over the last couple of years, these companies have been building up their offtake. We call them offtake. They've been building up their demand. It gets built up like that anyway, step by step. Nobody puts 5 gigawatts down and then waits for supply, waits for demand. That's not going to happen. That's right. You have to start. If you believe in AI in the future, you have to back off and say, "Okay, I need the land. I need the shell. I need the power. Either come off the grid. It's either going to be generated." There is a whole bunch of questions to line up for long before building the AI supercomputer. The important idea is that we're now talking about infrastructure. These are infrastructure timelines.
We have been talking to Europe now for some time. This just happens to be the visit where we talk about it with all of you. The infrastructure has been discussed for well over a year now. Yeah.
Thanks.
This is Brett Simpson again. I just wanted to follow up. Jensen, what do you think the useful life of NVL72 is going to look like? I mean, if I look at a lot of your customers, they're depreciating over different periods. I don't know if there's a comparison between Hopper and Blackwell, but do you think you can improve the useful life of the racks more? I mean, you've got 1.2 million components, I think, in these racks. But how long do they last? Yeah.
Two answers. One of it is the useful life, and then the other one is your accounting life. I mean, most people might account for either four to five years, depreciate over four to five years. But their useful life is going to be five, six years, seven years. The reason for that is you just have to go back. You might hear us talk about it. For example, this last two years, we improved the performance of Hopper by four times. Four times. In the last two years, software running on x86 improved zero times. We improved our software performance by four times. The reason for that is accelerated computing is fundamentally different than CPUs. There is a JIT, Just-in-Time compilation, that sits inside CUDA. CUDA is a virtual machine. I can change the software, improve the performance with new algorithms long after you bought the silicon.
We are dedicated to that forever. That is the reason why NVIDIA is doing so well. Because we go back and help you improve your performance for as long as we shall live. I have got mountains of people doing that. You would never do that for architectures where the install base is so small. You do that for CUDA because if you do that for Hopper, you benefit how many people? Researchers, software developers, people who work on software love doing this because it helps billions of dollars of infrastructure. Hopper keeps getting better. Ampere. I am still optimizing on Ampere. Ampere is now, what, five, six years old? Two different questions. I think their accounting life, that is up to them. I think they are going to find usefulness for years after, just as the cloud service providers are.
They're very happy with the old stuff.
Thanks. One maybe for Colette. When you think about AI, maybe the biggest dam forever for any company. You would expect companies to not care that much about margins, just try to capture the market. However, you guys have not only amazing growth, but also super high margins. Just trying to understand how you think about that equilibrium between growth and margins when thinking about EPS growth. Thanks.
Yeah. Our work that the teams have done building out the systems is not something to say that it's just a cost plus. It's been an enormous feat through software and complete engineering from the hardware standpoint to put what we put together. We have looked at it always and every time we do it from a TCO value. What can we provide to these customers? What is their next best option that they could do? That's how we determine an important part, which is the price. The cost structure and even the price from the very onset is something that we will always keep about the same price and will continue to fine-tune from the cost perspective. Now comes down, where do we make the investments in terms of our work?
The more and more that we can think about strong, new strategic investments that we can make to continue to see our platform grow worldwide. That does not mean by saying, "I'm going to create many different options on this chip." It just says, "Look at the total piece as a whole." Most of the work and what you saw today was talking about CUDA, the software, and everything that we need to do. Long way to go from getting to the enterprises. The enterprises still need a lot of change management on their existing software they're using. Our expansion and where we want to continue is take that platform and now enable every type of software system that's out there from the combination of what we put together. Yes, the margins are a strong margin.
We continue to be a company quite thoughtful in terms of our investments. This has not been a time where we are hiring tons and tons and tons of people because that does not necessarily always help you. We will continue to make the best investments, whether that is in operating perspective or using our cash. Those two things together will, I think, continue to enhance the true value that we can provide to investors of our full P&L to do so. A few lights.
This is what happens when you're being interrogated. This will be the last question. No pressure.
Thanks, Toshiya. It's Timm again at Redburn Atlantic. Maybe just on the NIMs, NeMo. You've talked about how you develop CUDA, how that's just an incredible part of the moat. When you think about NIMs and NeMo, could you just maybe talk about its significance in the sort of hyperscaler world today? Is NIMs, NeMo a more important part of your moat when you get into the enterprise? Maybe just to kind of give us some sense of just how big a deal it is within that overall.
Yeah. Great question. If you were OpenAI, you know how to build NIMs. If you were Google, you know how to build them yourself. The entire packaging of that runtime, super hard. The amount of software that's inside, we call it a NIM. Thank you. We call it a NIM. But the amount of software inside of it, there's CUDA, cuDNN, CUTLASS, TensorRT LLM, Triton. It's basically a ChatGPT in a box. You download it. You're talking to it. It's an AI. You download it. You say, "Here's a video. Tell me about this video. Reason about it. Why did it do what it did? What's it going to do next?" It's weird. It's basically an AI in a container. For most of the cloud service providers, they know exactly how to do this. For everybody else, they have no clue. And they shouldn't have to.
We should turn it into something like a NIM. It's the modern way of packaging AI. Do you guys understand what I'm saying? A long time ago, in 1993, is it? 1991? The retail box of Windows, you know? They figured out how to package software, started the software industry. I kind of went the first time we thought about NIMs, I'm going, "It's like they figured out packaging. We got to figure out packaging for AI so that everybody can easily absorb it and enjoy it." What Colette said earlier, she made a super important point, which is, remember, one of the reasons why we're able to deliver the value and prove it so is because the entire system of the GPU, the NVLink, the switches, the spine, the software, everything got integrated and delivered a performance level that's 40 times higher. Are you guys following me?
Dynamo, you can't. There's no 40 times in an ASIC. You're not going to go Hopper to Blackwell. Hey, look at that. 40 times. Moore's Law doesn't let you do that. Isn't that right? You don't have 40 times more transistors. How could you get 40 times more flops? The question is, how did we get 40 times more performance? Because we architected everything in whole, and we can deliver the software to do so. Otherwise, you're limited by gross margin plus on TSMC wafers. Does that make sense? You simply can't do what we do without understanding the big picture, architecting everything at one time, distributing the work across, pulling out these amazing things that deliver the throughput. That customer goes, "You know what? I get it. I believe it. You've been doing it every time. I buy it." They will appreciate the value.
We can talk about value instead of cost. Okay. It's great to see all of you in Europe. Thank you.