Session with press.
Hi, everyone, and welcome to GTC.
Hi.
This is Simona Jankowski, Head of Investor Relations at NVIDIA. I hope you all had a chance to view Jensen Huang's new GTC keynote this morning. We also published several press releases and blogs detailing today's announcements. Over the next hour or so, we'll have an opportunity to unpack and discuss today's news with our CEO, Jensen Huang, and our CFO, Colette Kress, in an open Q&A session with financial analysts. Before we begin, let me quickly cover our safe harbor statement. During today's discussion, we may make forward-looking statements based on current expectations. These are subject to a number of significant risks and uncertainties, and our actual results may differ materially.
For a discussion of factors that could affect our future financial results and business, please refer to our most recent Forms 10-K and 10-Q and the reports that we may file on Form 8-K with the Securities and Exchange Commission. All our statements are made as of today based on information currently available to us. Except as required by law, we assume no obligation to update any such statements. With that, I'd like to welcome you to the Q&A session with Jensen and Colette. We'll be taking questions over Zoom. I think you're all familiar with the interface, but as a quick reminder, please use the Raise Hand feature on Zoom if you'd like to ask a question, then unmute yourself when called upon. Let me pause for a moment here to review the queue before we pose our first question.
Our first question will come from Toshiya Hari with Goldman Sachs. Please go ahead.
Hi, can you hear me okay?
Yes, we can.
Okay, great. Jensen, first of all, thank you so much for the keynote. I'll probably have to watch it another 10 times to fully digest everything that was announced today. Just on the Omniverse, it seems like your customer engagement has broadened very significantly since you provided us with an update last time. Can you remind us how you're thinking about the contributions to your P&L? I realize this is, you know, relatively nascent and this is a very long-term strategy for you guys. As analysts, how should we think about the contributions to revenue growth and your profitability going forward?
Yeah. Thanks for that, Toshiya. There are three components to the Omniverse platform. The first component is the simulation platform. It's called OVX, the Omniverse computer. OVX had a first generation prototype, but the volume scale-out is based on second generation OVX 2, which is powered by Ada, L40s and CX7s and BlueField 3s. Without belaboring the reason why, in terms of performance, it's a giant leap over Ampere, as you saw in the keynote. It is all based on neural rendering and the performance is really incredible.
We're in full production with that now, and we're ramping and trying to get the OVX computers to customers as fast as we can. We have quite a large number of customers signed up to receive OVX computers. The second component has to do with the Omniverse applications that are built on top of it. There are applications that span the entire range of design, build, and operate inside a company. When you're designing a car or designing a factory or designing a product, all the way to manufacturing it, to operating it, Omniverse will be involved. I believe that Omniverse will be one of the first enterprise applications next to the web browser, if you will, next to web applications, that spans practically every organization.
In fact, we showed you a few examples of where design was involved, marketing was involved, you know, product configuration was involved, to manufacturing was involved, and simulation and operations was involved. Just about every single organization has the opportunity to engage Omniverse and work on a single source of truth. The second is applications. They're typically by user. The third application is a new type of database, and it's called Omniverse Nucleus, and we have just put that up in the cloud, and we're gonna host that as a managed service. Think of Nucleus as a database, but it's a new type of database because it's interactive, it's shared.
You bring in the data, the three-dimensional data or metadata, whatever, or behavior data or physics data, relationship data, supplier data, which component goes with which supplier, et cetera. You bring in all of that data associated with a product or a building where, you know, it could be a BIM. It could be JT files, which characterizes 3D geometry. It could be an ERP system to associate which vendor goes with which component. You bring in all of that data into Nucleus. Nucleus is shared by all the people that use it. It's an active distributed database. It is in 3D. It's the world's first large scale USD database.
That business model will probably be like a cloud database business model. The more people that are connected to it, there'll be a storage component associated with it, there'll be a use component associated with it. Nucleus is. Think of it as a cloud database, and we just we announced today that Omniverse Cloud will be hosted in the cloud, and that Omniverse Cloud is basically the database. We're in the process of fine-tuning the pricing associated with each one of the. In the case of the system, that part of it is very well understood. In the case of the pricing of the applications and services, we're in the process of fine-tuning that.
In the case of, the Nucleus database in the cloud, we're in the process of fine-tuning that as well. But my sense is that it will be very, if you will, conventional, compared to the likes of things that I've just described.
Thank you. Our next question will be from Tim Arcuri with UBS.
Thanks a lot. Can you hear me?
Yes, Tim.
Perfect. Awesome. So, Jensen, I think a big theme here are really these new infrastructure as a service offerings with the BioNeMo and the Omniverse Cloud. I guess I had two questions. First, maybe Colette, can you update us on, I think you said back in March, I think you said $100 million a year revenue run rate that you gave for recurring software and services revenue. First question is, can you give us an update on that? Then, second of all, Jensen, I'm sort of curious about the business model for this stuff. Are you gonna just offer instances on AWS and, you know, other clouds?
Ultimately, it seems like maybe you could offer your own cloud for this, and it seems like maybe this is sort of an inflection for you to look more like a, you know, CSP or, you know, hyperscaler yourself. Thanks a lot.
Yeah. Tim, thanks for the question. Let me first start with your first part of the question. As we had talked earlier about, we already have a run rate of software and services that we provide. That number is in the $200 million, and we're gonna continue to grow from that. We have great new offerings even today, but also offerings that we've been working on for some time as well. I'll move to Jensen, and he can talk a little bit more.
Yeah. Remember accelerated computing is a full stack computing approach. You know, the method of using brute force transistors and the advances of Moore's Law has largely ran its course. Going forward, the opportunities for continuing to ride the price performance curve of Moore's Law has ended. If you wanna be able to do larger scale computing and to do it in a cost effective way, you know, after 15 years of, almost 20 years of pursuing accelerated computing, I think that very broadly, almost it's conventional wisdom that accelerated computing is really the path forward. It's an opportunity for us to not just stay with Moore's Law, but go into a much more turbocharged accelerated computing law.
Artificial intelligence, of course, has benefited from that. Molecular dynamics has benefited from that. Weather and climate simulations is going to benefit from that. Ray tracing, NVIDIA's own business, core business of computer graphics has benefited from it tremendously. The first part is, it's a full stack challenge. Our architecture is available in every cloud, and our partnership with cloud vendors, CSPs, is really in two parts. There's the internal consumption part, which is about using NVIDIA's accelerated computing stacks to accelerate their workloads. It could be recommender systems, it could be speech AI, it could be you know, very large scale queries or.
Now, of course, the emergence of large language models, which is unquestionably the most important AI model of the decade. That's for internal consumption.
For external consumption in the public cloud, NVIDIA is a partner and, if you will, maybe even an extended sales and marketing force of all the CSPs because we, through our ecosystem, through our evangelism of the platform and through the software developers and all the startups that, you know, the 12,000, 13,000 startups that are built on NVIDIA, and all of these companies that are using NVIDIA's accelerated computing go into the cloud, we're essentially the business attractors of a very significant part of their public cloud service. We're gonna continue to do that.
There are several areas where we believe that we might be able to simplify further and democratize the reach of NVIDIA's accelerated platforms because it's fairly complicated to put these systems together and cobble it up yourself in the public cloud. Those two areas associated with large language models, which we announced today that we will have NVIDIA Managed Services. These managed services would basically be, think of it as AI training, but as a domain-specific AI training.
It's designed to be super good at large language models. The effort and the cost of training with NVIDIA services running in these clouds, and we'll run it initially. We're gonna run our services in as many clouds as we can. When they use our service, they could substantially reduce the cost of training a large language model or training what is called a prompt, how to tune that large language model for your specific application.
The second thing is Omniverse, which requires just a giant amount of engineering we've been working on for a year to bring Omniverse into the cloud and we have it currently running on AWS, but our goal is to have it run in all the clouds so that all of our partnerships with the CSPs have the benefit and the opportunity to attract customers that are using Omniverse. In a couple different areas, large language models per se, for English, well, for language and also for chemistry and biology, and then also Omniverse, those platforms are so complicated we thought we would stand up NVIDIA Managed Services and make it a lot easier and cost-effective for people to use it.
We'll probably do a lot more of that going forward, especially in the areas that are very hard to do.
Thank you. Our next question will come from Will Stein with Truist.
Yeah. Great. Thank you for taking my question, and I'll add my thanks for you know, all the incredible announcements you made today. Some of them may be a little bit confusing to us. There's so many, and they're quite technical. Jensen, you made one that was curious where in I think related to autonomous driving, transitioning from Atlan to Thor. Can you detail what conditions led to that decision? And maybe also remind us of your progress with your big customer announcement from about a year ago, I think, where you talked about Mercedes-Benz adopting this technology. And maybe remind us if there have been similar relationships that have been built to the same level with other OEMs that we might have missed. Thank you.
Yeah. Thanks a lot, Will. I'll work backwards. Mercedes-Benz ships the first car in 2025, late 2025. We also announced that JLR, all of the brands of JLR and their entire fleet of all the brands are going to also be powered by NVIDIA's full stack. That is shortly after 2025. We're pushing forward in both of those. We also have Orin designed into about 40 different cars and companies and not to mention medical instruments and IoT Edge AI servers and robotics of all kinds.
Orin, the chip that is in the self-driving car stack, is really just a phenomenal home run, and it started ramping a couple quarters ago, and it's going to be ramping quite fast, going forward from here. I think there's something like $11 billion worth of pipeline in the next several years that is associated with Orin and the systems associated with it and the software associated with it.
The reason why we decided to change Atlan, which was the next generation Orin, to a brand-new architecture is because the three processors that are inside a robotics processor, one of them is a GPU, one of them is a CPU, and one of them is our Tensor Cores. These three processors made such enormous leaps in the last two years. We made the hard decision to Thor and put these three new technologies into it rather than waiting another two years because the robotics system has a cycle time of about two years. Every two years, we announce a major leap. We just didn't wanna just miss it.
Atlan, the previous version, which is based on the next generation of Orin, it just missed it. We decided, you know, to bite the bullet and really just, you know, hunker down and work hard and it was just too much for us. We couldn't bear waiting another two years. I mean, that's kind of the simple fact of it, you know. All of these projects, they're projects at the heart and you have to love it, building it. It's incredibly hard work and there's a lot of imagination that goes into it, a lot of passion and hard work that goes into it.
I just couldn't imagine, you know, waiting another two years to get Hopper in there and Grace in there and Ada in there. We decided to just do it now. That's the reason why. It was probably, if I could say, you know, 60% passion and just couldn't bear not bringing that technology to the world and, you know, 40% just hard work. It was a lot of hard work, but the passion overcame the hard work.
Thank you, Jensen. Next question will come from Aaron Rakers with Wells Fargo.
Yeah. Thank you. Can you hear me?
Yes, Aaron. Nice to see, nice to hear you.
Thank you. Thanks for doing this call. I wanna go down the now that you've talked about today Hopper being in full production, which it sounds like, you know, I guess maybe first question on that is just to confirm that the pace of the Hopper ramp is now largely as anticipated given some of the questions around the China export situation. The second question to that is that, you know, when we think about the Hopper product cycle and you think about the breadth of the platform strategy that NVIDIA's built out, how can we think about the pricing strategy of Hopper relative to Ampere and the opportunity to just continue to take a bigger piece of that data center footprint as Hopper materializes likely over the next several quarters?
Yeah. First of all, Hopper is about five times the throughput because of a new type of engine called Transformer Engine. Transformers is the most important model of today. Really largely replaced RNNs and LSTMs and shortly will also replace a lot of the computer vision-based algorithms. It's a 5x increase in throughput. It's a 3x reduction in total cost of ownership.
The reason why is because inside the data center, aside from our Hopper, which is not a chip, as you know, if you look at the Hopper system, it's an entire system that we built, and then we disassembled it and took out the most complex part. It's called the HGX H100 board, which has eight Hoppers in it and, you know, tens of thousands of components. And it's just insanely complex, and we ship the whole thing as Hopper, HGX H100. That system goes into a data center, and the data center includes cables and networking and switches and power supplies and storage and, you know, so on and so forth, and power delivery and the infrastructure and so on and so forth.
We have 5x the throughput. We have 3 times lower TCO, which implies that Hopper's price is higher than Ampere's. I think the value proposition is so fantastic that it net reduces the cost compared to last generation substantially, as I just mentioned. Let's see. The ramp. We are in full production. We're gonna ship some quantity this quarter. We're gonna ship most of the quantity next quarter. The signals from the CSPs and the OEMs and the enterprises are really solid.
One of the primary drivers, and Ampere did have this benefit at its ramp, but Hopper has this benefit, which is a brand-new revolutionary new model called Transformers. You probably have been looking at, you know, seeing on the web the impact of the unreasonable performance, if you will, of large language models and being able to learn new skills with just a few shots of learning, a few examples of learning. Secondarily, its ability to generate images. The number of applications that are coming out all over the industries is really quite incredible.
We have the benefit this time of having almost an industry-wide recognition that large language models have democratized AI, made it easy for almost anybody to use AI. Hopper is really going to not only advance large language models, but to democratize the application of it because the inference cost is so much lower. Then not to mention one of the most vibrant industries at the moment, and probably the most incredibly well-funded segments of the world's startups is digital biology.
All of you that are you know watching other industries, you probably recognize that digital biology just went through its revolution between the cost of gene sequencing and the breakthrough in predicting structure of proteins and structure of chemistries and to be able to understand the language of biology and chemistry has turbocharged drug discovery segments and therapeutic segments. You're probably seeing a lot of startups go in there and it remains really vibrant. Anyways, Hopper is gonna be revolutionary for all of these applications.
Our next question will come from C.J. Muse with Evercore. Please go ahead.
There we go. Good morning. Thank you for taking the question and, congrats on the formal launch of Ada Lovelace. You know, on that front, would love to stick with gaming, Jensen. Now that you're on your third gen of ray tracing, can you talk about the competitive landscape? I guess, you know, maybe bigger picture, you know, considering kind of the inventory correction that we're going through, can you speak to kind of your anticipated ramp timing for 4090-4080, and then perhaps, you know, how we should be thinking about, you know, 4070 mainstream coming online? Thanks so much.
Yeah, we took action in the last quarter and this quarter to prepare our channel for a great launch. As you know, we've substantially reduced selling of Ampere's to allow the channel to normalize. The market for the end markets are soft, but they're not so soft that it wouldn't allow us the opportunity to sell through the excess inventory we had in the channel. We took specific action marketing programs to particularly reduce the segment that Ada is going into initially. We typically, and Ada will be no different, typically ramp from the top down. That's where the enthusiasts would like to see brand-new products and the customers that refresh more frequently every couple of years or so would like to see their new products.
It's also the segment where we need to ramp for Omniverse and Omniverse workstations and Omniverse data servers and so on and so forth. It was a sensible place for us to ramp first. We'll ramp Ada nicely, starting a little bit this quarter, but largely next quarter and very robustly going into leaving the year and going into next year. That's our current execution plan. We're in a really good place at the moment. I did mention you asked about competition. I think it's fairly well known that we're quite far ahead in ray tracing.
RTX is about two things, three things, I guess. The first is programmable shading that we invented some 20-some-odd years ago. We augmented programmable shading with RTX, which has two new processors. One is a hardware ray tracing processor, and we're in our third generation of that. The second is our tensor cores, our AI processors, and I guess we're in our fifth generation of that. The AI work that's necessary for neural rendering, neural graphics, fusing artificial intelligence and ray tracing and programmable shading, we are just miles ahead of. You saw some of the keynote today. To keep in mind that everything was rendered in real time.
They were not, they're not offline renderings of like movies. These are computer graphic simulations. To be able to see something like Racer X with no offline rendering and everything is no baking that most video games do today, and to be able to do that in real time is really quite just unbelievable. I think the Ada is a giant quantum leap forward, and it's a bigger leap if you look at all the performance, the power, the efficiency, every aspect of it. It's a bigger leap from Ampere than Ampere was from Turing. That kind of tells you something that, you know, we the third generation, we got a lot of things right and pulled a lot of things together.
Our next question will come from Vivek Arya with Bank of America.
Great. Thank you so much, Jensen and Colette, for the keynote and for the opportunity to ask a question. Just wanted to clarify, Jensen, just based on the comments you made about gaming. Is $2.5 billion still kind of the right run rate of end demand? I understand you're shipping below that, so I just wanted to clarify if that is still a reasonable number to use for what the actual end demand is of your gaming product. My question is, you know, on the Hopper, on the Grace CPU, what are NVIDIA's ambitions with that? Are you planning to kind of just focus on the HPC segment, or do you see the chance to take it across the entire $30 billion, right, TAM for server CPUs that is out there?
If the ambition is kind of more narrower than that, is it because of a technology issue? Is it a cost issue, right? Is it a supply issue? Like, what prevents NVIDIA from going after the entire $30 billion or so TAM for server CPUs?
Let me see if I can start, Jensen, and kind of clarify for the full room what's the $2.5 billion of demand for gaming. In our earnings release, we discussed looking at a normalized demand for our gaming, given that we were planning on undershipping for the quarter. Looking at Q2, Q3 combined, we believe that the underlying sell-through demand is approximately valued at about $5 billion. Then, yes, you can split that between two quarters. Seasonality sometimes plays into the quarters, so approximately $2.5 billion plus or minus. I just wanted to make sure we had that background for those on the call.
Grace Hopper. I highlighted a particular problem that is giant in scale and probably the most valuable software in the world today. A long time ago, Windows was the most valuable software, and then you could argue that you know, a decade or so later, PageRank was one of the most valuable pieces of software on the planet. Now it's a recommender system. It drives the vast majority of the world's e-commerce and just about everything that is put in front of our small screen from the trillions of things that it could have selected, it was recommended and ranked for us by a recommender system. The amount of data you could just imagine is enormous, and there are many recommender systems.
There's just not just one recommender system, but almost everything that pops up an ad or puts up a product or puts a price up or ranks a movie or a book in front of you or a blog or, you know, a short form video or long form video, just about everything that is put in front of you came out of a recommender system. Every single company has it. We recently just did a recommender system for an investment banker who had a product service that improved its results significantly, and they were so pleased by it for the products, for the services that they offer. In the future, almost every website will have recommender systems.
We're making it so that it's on the one hand much easier to scale using an SDK we call Merlin. On the other hand, making it a lot easier for people to engage it. I chose the recommender system because it represents today's AI factory, if you will. You know, almost every one of the large data centers that are running 24/7 are constantly recommending something and constantly collecting new data to go refine that recommender. I chose probably the one application that is simultaneously very large scale, also very different in computing profile than just about everything that we've done so far in AI.
I used it as a way to feature the uniqueness of Grace Hopper. There's a whole bunch of other examples like that. Spark, which is the leading data analytics engine in the world, used by, you know, probably 80% of the world's enterprise, and the most popular data analytics platform in the cloud, is going to, you know, Grace Hopper is gonna be ideal for that. You're going to use it for very large data science computing. I just selected, in fact, I just gave you a few examples of probably the most valuable enterprise applications.
It's going to be really great for cloud computing because the energy density is so high, so you could put a lot more CPUs in a rack, because the energy efficiency is so great. There are a lot of different use cases. I just chose a few, just to highlight what makes Grace so special. Even in that one example, Grace Hopper is probably seven times the performance of Hopper, which is the most powerful computer in the world today. The fact that we can create an architecture that is so much such a big leap forward for a particular application, I thought that was worthy to highlight.
Our next question will come from Stacy Rasgon with Bernstein. Please go ahead.
Hi, guys. Thanks for taking my question. Can you hear me?
Stacy.
Yes.
Yes, Stacy.
Great. Thank you. I wanted to ask about how you view supply and availability of Ada at launch, given what happened during the Ampere launch. More importantly, how are you gauging demand for Ada in an environment like this, where obviously the channel's flushing out, people are worried about, you know, GPUs formerly used in crypto mining getting dumped on the market. I mean, just given all the noise, how do you gauge demand for that product? I guess supply and demand is what I'm asking.
We delayed. I think it's fairly broadly known that we delayed the launch of Ada to give the channel an opportunity to clear the 30s, the 3090s, the 3080 Ti's, the 3090 Ti's, but basically the high-end segments, and gave us an opportunity to work with our partners to put marketing programs in place to move the pricing into a segment that even as Ada comes would still be really good value for anybody who bought it. We prepared ourselves and took two quarters of pretty harsh medicine in order to prepare for Ada.
Ada is gonna come into a segment that is going to be well above anything that is going to be affected by crypto or any of the reactions that you were referring to, Stacy. I think it's if you look at where 4080 is gonna come out at, it targets a very large segment of our GPUs because that's where the enthusiasts are and that's where, you know, the core gamers are. And they tend not to be affected by, you know, market conditions here or there. And so I expect Ada's launch to be very successful.
I believe that we've prepared ourselves and our channel to welcome Ada with open arms and we stay clear of just all of the dynamics that you mentioned. We should have a great launch.
Our next question will come from Matt Ramsay with Cowen. Please go ahead.
Thank you very much. Good morning. Can you guys hear me okay?
Yes, Matt.
Awesome. Thanks for letting me ask the question, and thanks for all the information today. Jensen, I think a lot of us are still trying to get our heads around some of the restrictions that came in with China and the intentions of the government and also the ramifications, secondary and tertiary, of what's been announced. My question is around some of the things that you guys announced and today on the software side, whether it's GeForce NOW in gaming or DRIVE in the auto business and obviously the Omniverse across enterprise, how do you see the environment for those software opportunities in China?
Does some restriction on your hardware being shipped into China accelerate those opportunities for you in software for Chinese customers, where you can host the compute and the data yourself? Do you fear that there might be restrictions on your software business in China? I honestly don't have a clue as to how to do sort of a risk/reward analysis of that for your software businesses. Any thoughts there would be really appreciated. Thank you.
Sure. The restrictions are very specific. It requires license for a specific level of compute combined with a specific level of interchip connection bandwidth. Within the restrictions, we will offer our customers, and they will have plenty of choices of alternative products that are within the envelope, that are not restricted. If a customer requires that very specific product, we will seek a license. I think that the US government would like to know who in the world are using products of that nature.
For most of our customers, alternative products are gonna be just fine. We're working hard and working fast to offer our customers alternative products. My expectation is that you know, outside of you know, proper execution, which the team is working really hard on, we should be able to offer and our customers would accept alternative products that are excellent. Regarding software, it has no incremental relationship with the restriction. The reason for that is this. You've heard me say before that accelerated computing is a full stack.
You can't just go from a compiler, C plus plus compiler or C compiler and compile software that runs really nicely on a GPU. That's just not the way it works. The reason for that is because the CPU was designed to be compilable. It was designed to be forgiving of bad code. It was designed in a single-threaded way so that software is easy to write. GPUs are designed in a multi-threaded way, and instead of one or two or four threads of execution, inside our GPU, there could be 15,000, 25,000 threads of execution. You know, 25,000 things running around, 20,000 things running around is hard for any human to keep track of.
We created a programming language, a programming model, and a whole bunch of libraries on top and a whole bunch of algorithms and a runtime engine that sits on top of that. We call those platforms. Those platforms, whenever I announce new software, I'll tell you about the new software on specific purpose. You know, out of the 300 SDKs, the question is, why did I select the ones that I selected? The reason for that mostly is one of several vectors. One, there's a new application field that is very important and that I think you should know about, or developers should know about or scientists should know about or ecosystems should know about.
Like, for example, for the very first time, we were able to take the most valuable, probably the most recent, you know, modern, valuable database called a graph database, and reformulate it, sample it into deep learning and use deep learning to learn predictive patterns out of the graph. We call that cuGraph. I spoke about it last time. I spoke about it this time. I'm gonna speak about it next time because it's just the radical importance of it in so many different industries. I might have spoken about operational research and the work that we do with cuOpt, the work that we're doing with large language models and the work that we're doing with quantum computing. All of these things expand our market.
That's the second reason why I tell you is because I highlight these SDKs to help you understand how accelerated computing is going into new markets that are well beyond deep learning maybe or enabled by deep learning in the case of graphs. In a lot of ways, the best way to look at NVIDIA's business growth is to look at our SDKs. The reason for that is because accelerated computing is a full stack thing, if there's new software, there's new consumers, new demand for our hardware.
It's always software before hardware, but not the other way around.
Our next question will come from Mark Lipacis with Jefferies.
Hi. Thanks for taking the question and thanks for the great presentation. Jensen, you used the expression full stack computing a lot, and I, you know, I think there's a lot that goes behind that description of what you guys are delivering. Can you talk about how you transformed NVIDIA? Like, what does that mean internally for the company that you've kind of become a full stack computing company? Can you compare what you are doing today in this accelerated computing era to, you know, what was the model before? Is it fair to say that what you are to accelerated computing, like, you know, what Microsoft and Intel were to PC, is that a fair analogy, or is that underselling what you guys are delivering?
If you could kind of put into perspective what you guys are really delivering out there relative to previous computing eras, I think that would be interesting. Thank you.
Yeah. Mark, that's a great question. The PC industry was actually the only unique industry, and it created the horizontal business model. Notice cloud computing is not horizontal. Cloud computing is in fact vertical. You have cloud service providers who are platforms as a service. They have SaaS, they have PaaS, they have database as a service, so DBaaS, and they have IaaS, you know. There's a lot of as a service. You could think of accelerated computing in basically the same way that in order to be successful in accelerated computing, it's not so much that we're doing what other people used to do, and we aggregated into one vertically integrated company.
It's just that in this field of computing, if you're not vertically integrated, you're not gonna be successful. Nobody's gonna write your operating system. Nobody's gonna write your runtime engine. Nobody's gonna develop your, even your distributed operating system, you know, whether it's in the cloud or supercomputing or enterprise. Nobody's gonna write it for you. You really have to go do it yourself. You have no choice but to be a storage company and a networking company because the storage and networking and cybersecurity in the world of multi-tenant public clouds all where it interacts with our work at accelerated computing, which is data center scale, your network, your storage, your cybersecurity is really part of your computing fabric.
You have no choice, you know? It's not about building the chip. None of my customers buy the chip. They need you know, ultimately, they might place a PO on a whole bunch of chips, but what they're really buying is the NVIDIA computing stack. Which explains the reason why we're you know, so broadly used in so many different clouds and computer makers and so on and so forth, that the stack just you know, it, wherever you happen to be using it does what we promised it was going to do. It delivers speed up well beyond reason and well beyond the cost of adding it by many orders and many factors, right?
Anyways, it's a full stack computing problem. We talk about basically four stacks, Mark. There's the first stack NVIDIA put together was our graphics stack, and today it's called NVIDIA RTX. Computer graphics is now a full stack problem. If you don't design and develop the artificial intelligence or the physics engine or the ray tracing engine and all the software in it. You just, you know, nobody's gonna do it for you. It's not part of OpenGL, you know, it doesn't exist. If you don't build that full stack, people can't use your platform. The second is NVIDIA HPC. That's our scientific computing stack.
That's for quantum chemistry, molecular dynamics, so on and so forth, fluid dynamics and so on and so forth. Then the third is NVIDIA AI, which has all of our end-to-end runtimes and engine. NVIDIA AI is essentially the operating system of modern artificial intelligence, and it goes. It starts from data ingestion, data processing with RAPIDS into deep learning into now with cuGraph our graph analytics and graph learning systems, all the way to inference Triton. That end-to-end platform is part of NVIDIA AI, and if you're doing machine learning or artificial intelligence of any type of model anywhere, you could use NVIDIA AI. Then the last one is NVIDIA Omniverse.
NVIDIA Omniverse is the next wave of AI where the artificial intelligence has to interact with the physical world, and you need a way of providing ground truth for the artificial intelligence and Omniverse is designed for that. We gave these four platforms platform names, but they all include operating systems, networking, storage, data center scale stuff, all the libraries, all the runtimes, and they're represented as four platforms.
Our next question will come from Harlan Sur with J.P. Morgan. Please go ahead.
Yeah. Can you guys hear me?
Yes, Harlan.
Yeah, thanks. Good morning. Great keynote, and thank you for hosting this call. Jensen, on these new managed services offerings that you unveiled and will be unveiling going forward, NeMo, BioNeMo, LLM, you guys are further accelerating customers' time to market with large language model training, tuning, and deployment. What's the
Yeah.
Monetization model here? Is it subscription-based or consumption-based? Then I also noticed that the L40 announcement going into your OVX platform. When should we expect the launch of the rest of your L-series Ada line of GPU solutions?
I'll take the second one first. We announced NVIDIA RTX A6000, and it will complement Ampere in the workstation lineup. You know, our typical rhythm is desktop first. A couple of quarters later, maybe less than that this time, the notebook and all the thin and lights and, you know, all of those things come shortly after. You could expect that basic rhythm. In the case of L40S, your question was what about the other versions? You know, I wouldn't wanna ruin the surprise, but L40S was really quite the perfect one to launch this time because customers are really clamoring for the Omniverse computers.
We're working with all the OEMs in the world and getting L40s, as you know, shipped out to the enterprises. Our business model for the NVIDIA managed cloud services is going to be by consumption. For example, in the case of NeMo large language model, we've made it so that it's much easier for you to learn the prompts associated with adapting that language model for your own application. Everything has been stood up and you know, it will likely be a model that's very similar to a cloud service provider's per GPU hour, except it would be value added per GPU hour.
net of the value added price, you should still see much lower cost and at the very minimum, much more effective, much easier way to train your model. Let's see. I think that's about right. Omniverse is going to be a consumption model as well, and so is the training, the large language model training.
Our next question will come from Raji Gill with Needham. Please go ahead.
Yes. Thank you for taking my questions, and thank you for this keynote as well. Just a quick question on the near term if I can. I appreciate the fact that you are under shipping gaming demand by quite a large amount in order to clear out excess inventory in the channel to prepare for your next generation gaming architecture. So that is. Then you also indicated that you have about $5 billion kind of run rate if you combine the two quarters. That would imply that you have, you know, pretty good visibility still into end market demand.
I'm just wondering about the overall end market demand if roughly two-thirds of the gaming market is tied to China and Europe, if I have those statistics correct. You know, both economies are weakening fairly significantly. I don't know if China's pushing a stimulus plan or not, but the question really is about visibility into that end market demand, that level of confidence.
I appreciate that you're taking the large inventory correction. I think that makes a lot of sense, but just any insight there would be super helpful. Thank you.
Did you wanna take that on? I'm happy to either way.
No, I'll start off. That said, yes, I think it is quite accurate that we've got a good amount of business around the world in each of the regions. You can almost divide each of the regions one-third, one-third, one-third of what we're seeing. Each of them are dealing with different macro conditions right now. Underlying, the ability to game, the form of gaming as entertainment is still driving solid sell-through in terms of our products. This underselling that we are doing is the right time before we produce Ada going forward. We watch this carefully. We are still seeing the solid demand. Now how do we see that? Do things change going forward? Possibly. Right now it looks solid.
I'm gonna throw to Jensen Huang, see if you have anything more to add.
I thought that was perfect.
Okay. I think we have time for one last question, and that will come from Joe Moore with Morgan Stanley. Please go ahead.
Great. Thank you, and thanks for the presentation. I wonder if you could talk about pricing in the gaming business. I noticed the 4080 comes out at a higher price, materially higher than the 3080, which I know the suggested retail price on 3080 was never really achieved. It was always above that. Can you just talk generally about how we should think about how cost inflation is affecting you guys through the full stack and to the extent that I know you provide more value each generation, and the prices have generally been drifting up, but is this cycle gonna be fairly normal, or is the general inflation gonna drive the prices higher than normal?
Well, first of all, the value proposition of Ada is off the charts. We have never had a generational leap this great. The new architectures, we fundamentally advanced all three processors. The breakthroughs in neural rendering is just off the charts. It's really unlike anything that's ever been possible before. The value is incredible. The adoption of our platform across the world's game developers is incredible. Ada's gonna come out gunning. It's gonna be incredible. The value is incredible. The pricing is higher than last generation, but I would say comparable. The gross margins to us is comparable to the last generation. That kind of frames it.
I think the most important thing is at each price point, the value that a gamer is going to get from their new Ada graphics card is just going to be off the charts.
Thank you, Jensen. I think that's all the time we have for questions today. Are there any closing comments you'd like to make before we get off the air?
Well, let me do this. We covered a lot of ground and when somebody comes to see an NVIDIA keynote, unlike a chip keynote of speeds and feeds and, you know, teraflops and gigaflops, we have plenty of those as well. But you get to hear about a whole bunch of new applications and new industries and particularly new software stack. Let me just frame very simply the four things we talked about. It was really in four categories. The first is just a whole bunch of new chips. We're in a new product cycle. We kicked off multiple product cycles at the same time. Ada is for gaming, it's for workstation, and very new, it's for the very first time, it's also for Omniverse.
Hopper, our Transformer Engine is a giant leap forward. It's in full production. Orin, unlike Xavier before that, which was really only for AVs, has been just a home run with respect to all of the customer adoption, but it's also for robotics. It's also for our industrial edge, and it's also for our medical instruments. All of these are going to be robotic systems. Next up, I really appreciate the question with Grace CPU and Thor. They're next up, ready to go. For the next round. The first is just a whole bunch of new chips and new hardware that's associated with that. The second is to recognize that accelerated computing is a full stack and a data center scale computing approach.
You know, Moore's Law has slowed and, you know, to be even more blunt, has really stopped in every measurable way. It is now broadly accepted that accelerated computing is the path forward. NVIDIA, recognizing that this is a full accelerated full stack challenge, has put together four platforms. I mentioned RTX, I mentioned scientific computing, I mentioned AI and Omniverse. During this GTC keynote, we spoke about three of them. RTX showcased new neural rendering and DLSS 3. Our NVIDIA AI spoke about the SDKs. It's much, much more than deep learning. Even deep learning has taken a giant leap forward with large language models and recommender systems.
We highlighted RAPIDS for data analytics, Spark for data analytics, JAX, the work that we're doing with DeepMind and Google Brain. The next major framework, cuGraph for graph analytics and Triton, which is used by 35,000 companies around the world, is about to roll out for large language models, which allows you to do training on distributed computing. It's a technological marvel. Just as AI needed ground truth to learn from, you know, ultimately these even if they're unsupervised, they have to learn from some form of ground truth. The next wave of AI, where AI meets the physical world, needs its ground truth. Its ground truth is impossible to collect from the physical world, and so we have to generate ground truth from the virtual world.
We call that Omniverse. Omniverse will be as vital to the future of artificial intelligence as the first generation TensorFlow and PyTorch were to the first generation of artificial intelligence. This is where the digital world meets the physical world and the next wave of AI happens. We announced 150 connectors, which opens up Omniverse to all of these industries because they would like to find a way to automate their industry. One of the largest ecosystems is called Siemens JT, and so we're delighted to have a connector for that. But we have 150 others.
We extended NVIDIA AI and NVIDIA Omniverse into the cloud, and we have just spoken about that, so that we could make it easier with some of these really complicated workloads that are really in the domains of some of the largest companies. We made it easier so that we could democratize these some of the areas of artificial intelligence so that every researcher, every scientist could take advantage of it. We did that by extending it into the cloud, and we'll have, you know, with Omniverse, we'll have the Nucleus, if you will, Omniverse super cloud. It's going to be a database in the cloud, a brand-new type of database.
Just as relational databases were a revolution a long several decades ago, graph databases were a revolution about a decade ago, there are all these different types. Now we hope that Omniverse in the cloud will enable just about everybody, every designer, every industry to be able to connect to it, work among each other. That was the third announcement, the third category. Finally, this is the era, and I've been speaking about this for some time, this is really the era of enterprises taking advantage and applying AI to revolutionize their products and services, but also to revolutionize themselves, to bring automation into their companies.
You know, all of you tracking IT for a long time know that a long time ago, there used to be one IT department in most companies. Now there are four IT departments, and these IT departments all have different reasons that they do what they do. There's sales marketing IT department, as we know, that emerged kind of about a decade ago. The two new ones has to do with artificial intelligence and deploying artificial intelligence workloads. One has to do with MLOps and the creation of AI. We say that that's you need two computers to deploy AI. One to develop the AI and simulate the AI, and then you need one to deploy and. You know, the word that's used is inference the AI, apply the AI into your products and services.
These two organizations both have IT departments and the way that you apply this new form of software is rather arcane and is not extremely easy. For all of us that have been doing it for a long time, this is the new way we do software. It's time for the world's enterprise to be able to take advantage of this capability, and there are so many enterprises.
This time, at this GTC, I'm delighted to announce that Deloitte, the world's largest professional services firm, with their 350,000 professionals, are gonna build practice on NVIDIA AI, NVIDIA Omniverse, so that we can together take these platforms, NVIDIA AI, NVIDIA Omniverse, to the world's enterprises and help them create new products and services and help them revolutionize themselves. Four major categories, new chips, a whole bunch of new software, new services, and new partners. Thank you all for joining GTC. This was surely a news-packed GTC, I appreciate your attention, and I look forward to seeing you soon.