Gentlemen, good morning. Hello. Michael, in the quarter gone, 1,000 new clients for AI server for AI factory is a hell of a jump. What is it that those new clients, 5,000 total, are actually building now different to one year ago? I think that's probably a good place to start.
I think the change we see is it's kind of moved from testing and evaluating into production. And we showed some great examples on stage, right, with Eli Lilly with a thousand GPUs, the physical world, and Samsung. And these are not things that are on the screen, right? This is in the real world with the largest companies in the world. And so it's propagating broadly across all customers in every industry and every country. And you know, you see the improvement in all the models and now we have the agentic capabilities. And so while it is exciting and there's been a tremendous amount of growth, I still think it's just the beginning of this wave, particularly when it comes to enterprise, which is really where, you know, we have an enormous opportunity.
What's so fascinating, Jensen, is you spent four years telling me that we needed to change the definition of the computer in the context of accelerated computing. But the big focus was on the hyperscalers, right, cloud. What I took from Michael's presentation was this is happening, you said locally, but on-prem. What's the NVIDIA interpretation of that part of this cycle?
Intelligence has to be performed, produced at the point of context. And so wherever the context is, wherever the action is, that's where you want to produce the intelligence. For most of the early applications of AI, it was in the cloud. A lot of consumer services are in the cloud. However, for Lilly, Samsung, the future manufacturing, a lot of companies, you want the agents to be on-prem because that's where all of your data is, where all your secure data is, your proprietary data, and all of the skills associated with your company is. And so now we have agents that are here, AIs that can do work. ChatGPT was fantastic, a launch generative AI, but it just made content. That was it. Making content is very important, but doing work is really valuable. And now we're doing productive work incredibly well. That's why they're called agentic AI.
In this new era, what everyone is trying to work out is, aren't all the GPUs locked up at the hyperscalers? How is Michael Dell going to service those 10,000 new clients with the GPUs to build their own on-prem local AI factory?
Well, the supply chain that Jensen has built and we've built together is continuing to scale up. And while it's true that there's more demand than supply, there's more supply that's being added. And you know, customers are figuring out how they start to scale these systems up. So, you know, I think what's also happening is companies are understanding that when they reimagine their workflows using this technology, they don't get 10 or 20 or 30% improvement. They get 10 times or 20 times or 100 times. And that is really the speed that matters to make a business successful. We're doing it ourselves. NVIDIA is doing it. And so it's not a secret anymore that these things are possible. And every company wants to capture that speed and translate it into competitive advantage and outcomes.
Dell was the sales channel, right, Jensen? Michael's company is very good at selling technology to America's biggest companies. How's that going to change things for NVIDIA going forward? Like the makeup of the types of companies we're talking about are at scale, but there's also that kind of middle market of data center that's being built, different kinds in the industrial space, in healthcare. Is that something that puts NVIDIA into new territory away from the frontier labs, away from the hyperscalers?
Well, NVIDIA is a technology company, right? The hyperscalers have the ability to take our technology and integrate them, operate them into a service. Dell has the ability to take our technology, turn them into a solution that delivers impact to customers. If you look at what has happened, agentic AI has completely, as we were talking about earlier, reinvented computing. We had to do several things together. The first thing, of course, we had to build the brain. This is the Grace Blackwell NVLink72, the Vera Rubin NVLink72 for giant large language models. The second part now is the Vera CPU that we're now in the process of launching, the highest performing CPU in the world. It's designed for agentic AI and now this will be the harness running the agent itself using the tools.
What does harness mean? What's harness?
Harness is what puts a harness around the large language model so that it can access memory, access the network, use tools, have local scratch pad memory, working memory, access long-term memory. And so that harness basically turns, if you will, the brain into an agent, into a digital robot, if you will, that can do work. And so now the agent runs on the CPU. We also worked with Dell to create a new type of long-term memory for agents. We call it the Dell AI data platform that's built on NVIDIA. The networking to scale it out is built on NVIDIA. So the agent, the brain, the long-term memory, all of the networking necessary to scale it up, as well as the agent runtime itself, we call NeMo Claw, running in a secure and governed container called OpenShell. All of that has been put together and now the technologies are the technology parts. What Dell has to do is turn it into a solution that people can use. Dell will do for the world's enterprises what the clouds do for the clouds. Makes perfect sense.
What is the Dell story, Michael, around CPU and sort of like general purpose computing in the agent era? We've talked a lot about the AI factory offering, the GPU, but actually there's potential for you in more general purpose workloads. The buildout's happening either way.
It is, and the demand is exceeding the supply there as well, as you know. And look, as you move to these agent frameworks inside companies, you use a lot more CPUs. And you know, that's just the reality of what's happening. And I think that's only going to increase. So instead of humans using tools, it's now agents using tools. And agents, as you were talking about earlier on stage, we have a billion people, we'll have hundreds of billions of agents.
People use tools every now and then. Agents are going to use tools all the time. And agents use tools very quickly. And so we're going to need a lot more CPUs. And those CPUs are connected to GPU brains so that the CPUs know how to think, how to reason, how to plan, and how to use those tools. So that's basically how it works.
Gentlemen, what is the biggest supply constraint or bottleneck for you right now?
Well, certainly, you know, memory is a challenge. I think it is memory. The advanced node semiconductors are still challenging. You know, it's really, I mean, we think about it from the things that we're producing and the semiconductor supply chain is ramping, but the demand's growing faster than the supply. In our case, we provide the technology integrated and so the memory comes with our technology. We've been planning our supply chain for a couple, two, three years. We have the largest supply chain in the world. Our partners have done a great job securing supply for us. And so all of the pieces go together. The CoWoS is lined up with the HBM, which is lined up with the Grace Blackwell and the CPUs, and the CoWoS, the CoWoS, all of it is all lined up. The silicon photonics is lined up. Everything is all lined up. It's just that the demand is much greater than the overall capacity of the world.
So the overall capacity, Jensen, should I put my textbook away? Because if I get my textbook out, it tells me that memory historically is cyclical. It's boom and bust. And so you both kind of have to convince the memory makers of the permanency of this to build the capacity that won't sort of fall away. Is that the right way of looking at it? That this is not a boom and bust cycle. It's just a complete change in the structure of that market.
Well, Michael and I do this all the time. We spent a lot of time with the supply chain. I mean, if you ask Sanjay Mehrotra over at Micron, they'll tell you three years ago during a meeting, I explained the future to him exactly as it's happening right now. And I was really grateful that Micron and NVIDIA really lined up to line up all of our roadmap. Tony will tell you over at SK that we did the same thing years before. And so it's our job to make sure that the vision of the future of the industry we convey upstream to our supply chain so that they are building for it. We also have to convey it downstream to people who have power generators and land and financing and so on and so forth. And so we have to make sure that the supply chain upstream and downstream are prepared for this future. It is true that the simple logic is this: that we have now reached a level of agentic AI, useful AI, productive AI capability. And the way to think about these agents is kind of like just digital workers, right? We have hundreds of millions of digital workers in the world. We're going to have billions of AI agents in the world and they're going to be working 24/7. And so just as we give every digital worker a laptop and a small part of the data center, we're going to have to give every agent essentially a computer and a little bit of storage in the data center to use. Think about it this way. You know, you do individual work, you know, as a person.
And you send it on to somebody else and you know, there's interactions. Well, now you might have hundreds or thousands of, you know, digital agents working for Ed, right?
That you supervise, and that's going to help you be way more productive, get way more things done, expand your creativity. Now, it does require a lot more computing and memory and storage and networking and all the things that we're doing together.
Last one on this, like Jensen outlined the Micron and the SK example. Three years ago, you gave them the heads up. Do they believe you? Are they sort of acting on that?
They're investing. I mean, we're managing through it, but these things are very hard to predict, right? If you tried to predict, you know, in 2023 what the demand was going to be in 2027, you would have a hard time doing that. So it does take a long time to build these factories, but we've got great relationships with these partners. We have for decades. That's helping us. And they see that we're winning and so they want to work with us even more. And it's really a great long-term partnership, even though we'd like more right now.
We're in the beginning of the AI buildout. This is literally the very beginning of the agentic AI buildout. We're going to be building this out for a decade, maybe more, because after this, digital agents will be physical agents. Then we go to the physical AI. We haven't even started that. I mean, you saw some examples of that, you know, in the keynote, but that is a way bigger market and it will require all sorts of new infrastructure capabilities. We're going to, for the very first time, bring it to the world's 90 trillion dollar other industry. And so there's a giant industry ahead of us to build towards now. Meanwhile, the supply chain is more than doubling every year. I mean, it's probably quadrupling every year, but we'll still have a hard time keeping up with the buildout for at least a decade.
Why sense? China. Jensen, you just returned from China. On Friday on Air Force One, the president said that H200 came up, but that China's position is it wants to support its own industry. Could I just ask what the net outcome was of your trip to China and your understanding of what is not or is allowed with H200 and the customers that you have or do not have in China.
