NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Good morning. My name is Michelle, and I will be your conference operator today. At this time, I would like to welcome everyone to NVIDIA's GTC Financial Analysis (sic) [Analyst] Event. [Operator Instructions] Simona Jankowski, you may begin your conference.

Thank you. Hi, everyone, and thank you for joining us for our GTC Financial Analyst Event. We hope you are all able to view Jen-Hsun's GTC keynote address this morning. We're excited for this opportunity to spend some more time with the investment community unpacking all of our announcements. Before I go over introductions in today's agenda, let me remind you that during this presentation, 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, October 5, 2020, based on information currently available to us. Except as required by law, we assume no obligation to update any such statements. With that, let's -- we're ready to get started. We have 6 speakers for you today who will cover the highlights from this morning's announcement and what they mean for our business. NVIDIA Founder and CEO Jen-Hsun Huang will kick us off with an overview of the NVIDIA computing platform and strategy. Next, we'll have 4 speakers who will cover important aspects of our data center market platform. And we'll wrap up with our CFO Colette Kress before opening up to Q&A. We expect the entire program to wrap up in about 2 hours. And with that, let me turn it over to Jen-Hsun.

Hi, everyone. This is a very special GTC because many new platforms and products we're working on came together. The strategic themes that video is driving will reverberate throughout GTC. AI is the most powerful technology force of our time, software writing software, the age of AI, automation of automation, the age of AI will often open large untapped markets. Accelerating computing the full stack challenge, it starts with a great chip, but the stack is a lot more complicated than that. Accelerated computing platforms takes on more than the chip, just as cloud computing platforms take more than a server. The new unit of computing is the data center, whether because cloud-native applications run across an entire data center or because edge computing will have a whole data center on the chip some day. The iPhone moment of industrial companies is here. AI services will span cloud's edge. NVIDIA AI fully supports x86 today, we also want to bring NVIDIA AI and accelerated computing to ARM, the world's most popular CPU and offer our architecture to ARM's vast ecosystem. Simply, we are building the computing company for the age of AI. We're focused on 5 major domains. The applications in each domain share similarities in algorithm, use cases, system platforms, ecosystems or end market. These 5 domains are NVIDIA RTX, NVIDIA HPC, NVIDIA AI, NVIDIA Enterprise AI and NVIDIA Edge AI. Let me say words about each one of them. NVIDIA RTX is inventing the future of graphics and digital world. RTX is a massive endeavor, reinventing real-time graphics, requiring us to fundamentally change every layer of the stack from top to bottom. RTX was high risk, but it is now clearly a home run. Ampere is our second-generation RTX, we expected and prepared for great demand, and the [ 30s ] series ramp is our fastest ever. Still, the demand is exceeding our expectations. We also announced Omniverse, our physically based simulation and collaboration platform. A platform for creating digital world is an open beta. NVIDIA AI is in full throttle Ampere swept the latest MLPerf training benchmark, demonstrating our growing lead. Ampere is our fastest ever data center ramp. This is the first generation that required no [ explanation ] whatsoever. To any of the customers, OEMs or cloud data centers, but the need of NVIDIA CPUs and data centers. They only ask when they can take Ampere to market. Today, we announced the new NVIDIA RTX A6000 and NVIDIA A40 enterprise and data center and peer-based GPUs. These are PCI express base and complement the already in full production NVIDIA A100. We also announced NVIDIA Jarvis conversational AI, SDKs and open beta; NVIDIA Merlin recommender SDK is an open beta. These are 2 of the most important AI models in the world today. And we just also announced the NVIDIA Maxim SDK for cloud AI video process. Maxim will help video conference services take advantage of NVIDIA's GPUs and NVIDIA AI in the cloud. Live video is one of the most active, most busy traffic in the Internet today. Finally, we announced that our stacks are so popular that we're working to bring all of them inclusively in MGC, our cloud registry, into each cloud marketplace. Essentially like a store within store. The next wave of AI is the enterprise. Enterprises will use AI to automate their company and use AI to bring automation to the products and services their companies build. The latter is also called industrial IoT or edge AI. It comes in several names. Imagine a John Deere Autonomous tractor connected to a Joint Deere cloud service or an autonomous Mercedes-Benz connected to their cloud service or a connected air conditioner, Street Sweeper or an entire building connected to AI services. This is the iPhone moment for the world's industries. AI will automate the world's largest industries. Breakthroughs in AI have made the automation software possible. NVIDIA enterprise AI is about helping companies build modern, secure and video accelerated data centers and offering the software platforms to help each major industry apply AI. Manuvir will talk about NVA EGX enterprise, the new NVIDIA Bluefield data center infrastructure on a chip platform and our vision for enterprise AI. NVIDIA edge AI is about helping companies build modern, secure and NVIDIA accelerated edge data centers in a box with software stacks that let customers operate their network like a fleet and software platforms to help major industries create and operate AI services. Today, we announced fleet command, a software-as-a-service offering to help operate these fleets. Another recent example, AI-powered connected product and services is our Mercedes-Benz partnership. Our development with Mercedes-Benz is in full throttle. Mercedes will be using our entire stack from infrastructure, AV computer in the car to the driving application. Justin will talk about NVIDIA EGX Edge and the wave of partners joining us. NVIDIA HPC consists of supercomputing centers in industrial HPC. Industrial HPC, demand is accelerating a large number of domain-specific applications, for example, health care domain, representing over a decade of work. We created a state-of-the-art suite of accelerated tools to help medical researchers discover life-saving drugs. We announced NVIDIA Clara Discovery. Kimberly will talk about the great work we're doing in drug discovery and our partnerships there. This is our biggest GTC ever. Over 1,000 sessions, a record number of sponsors, a record number of start-ups participating. We announced SDK. These SDKs are the critical difference between the GPU chip and the NVIDIA accelerated computing platform on GPU. These SDKs run on NVIDIA's 1 billion CUDA GPUs, one architecture, gigantic installed base for developers. These SDKs covered a range of NVIDIA's full stack computing platforms, chips and architectures like CUDA, GPU and DOCA for DPU; systems or systems components, RTX, DGX, HGX for hyperscalers, EGX for enterprise and edge, AGX for autonomous machines; system software APIs for Windows, Android QNX, Linux, Kubernetes, VMware; for PCs, cloud, [ Enterprise S ] acceleration libraries and engine like the CUDA X libraries, Magnum IO, [indiscernible], TensorRT, Triton Inference Server, the RTX stack, our physics engine. And of course, applications and application frameworks, Omniverse, NVIDIA DRIVE, Jarvis, Merlin, Isaac or robotic stack or Clara, our computational health care and life sciences stack. All of these stacks are optimized and containerized on NGC. The NVIDIA SDKs are created in service of our nearly 2.5 million developers, researchers, software companies who accelerated the 1,800 applications for the billions of computer users, the global computer makers, cloud service providers, solution partners, the 6,500 start-ups in inception. We've built GTC for them. We built our SDKs for them and services of them. They are what GTC is all about. To tell you more about our announcements at GTC, we've got some very special speakers lined up for you guys. So let me first introduce Paresh to tell you about NVIDIA AI. Paresh?

Thank you, Jen-Hsun. Jen-Hsun talked about how AI is software writing software and AI is already achieving results that no human written software can. The interesting thing is this approach is extremely scalable, larger, more complex models, create more capable AI, AI that's more accurate and applicable for many different types of tasks. So the chart on your left basically shows the number of days it takes to train a model on a 1 petaflop supercomputer or a computer. And it continues to grow exponentially. It's now doubling every couple of months. Think of this another way. If [ more global ] were still true, it would have delivered every 10 years what AI needs every 10 months. So larger models, we need larger and larger supercomputers to train on, expanding the opportunity for NVIDIA. And the capabilities of these advanced AI models and their potential to transform the industries is immense. On the right, you'll see a few recent examples of it. NVIDIA's natural language understanding took the reading comprehension challenge that consists of middle school and high school level tests. The average human score on it of 73%. And NVIDIA's Megatron-BERT scored 91%. In addition to reading comprehension, language models are also being used to predict the 3D structure of protein, just by reading the amino acid sequences. This will have a dramatic impact in discovering drugs. Kimberly will talk a bit more about it later. Facebook AI research developed a very large AI model based chatbot that exhibits knowledge, personality and empathy, they call it Blender bat. Half of the users tested chatting with blender bots over humans. In another world, researchers at Caltech developed reinforcement, learning-based drone flight control systems. They can smoothly land drones with a payload on different surfaces, a critical problem to solve for making safe drone deliveries. Larger and advanced models bring transformative capabilities. And when deployed in applications, they also require GPUs for inference for the economics to work for companies. Take, for instance, a BERT-based model used by Microsoft to improve their search change Bing. The accuracy of the model resulted in the largest improvement in Microsoft search engine. However, it was impossible to run it on CPUs. With our GPUs, Microsoft got 800x higher throughput, and they could run these models in real time. This led them to switch to thousands of NVIDIA GPUs running on Azure to power their search. AI needs GPUs, both for training and inference. Our Ampere architecture delivered for the first time a unified platform for AI training and AI inference. It provided 20x higher performance, a unified computing architecture for data analytics, training and inference, and with multi-instance GPUs, the ability to scale down 50x in a single server. This was the reason why it was picked up so readily by the industry. A100, as Jen-Hsun talked about, has had the fastest ramp of any data center GPU in our history. Shortly after Jen-Hsun announced it in May, over 50 different server models were announced by the world's leading server makers based on A100, all leading hyperscalers announce plans to deploy A100 to their cloud. It's already available on Google Cloud, Microsoft Azure and Oracle Cloud. Of course, having a mighty GPU architecture is at the foundation of NVIDIA AI platform. However, AI workloads push the limits of every aspect of the data center, computing, storage, networking and software to run all of this. We are addressing this challenge in democratizing AI with 3 pillars of NVIDIA AI platform: data processing and training, inference and AI application frameworks. First, for data processing, feature engineering and training, our platform can run at any scale from 1 GPU to multiple GPUs and to multiple loads across the data center, running any framework, training any type of model or being available on any cloud. While some companies are just starting to do deep learning, nearly every company is doing data processing with exponentially growing the scale. And our work with RAPIDS is now transforming the data analytics landscape. It now accelerates to Spark, the world's leading data analytics platform used by 16,000 enterprises and 0.5 million data scientists. And earlier today, Jen-Hsun announced that Cloudera is now accelerated on NVIDIA AI. Cloudera has 2,000-plus customers and runs on 400,000 data center servers. Manuvir will touch a bit more on this in his presentation. Second, inference is a great computing challenge. It requires a lot of software to make it work. And we break it down into 4 steps, each very complex. Starting with pre-trained, state-of-the-art models for key use cases that are available from NGC or cloud registry hub for our GPU accelerated software. They provide an easy ramp to enterprise customers to infuse AI in applications. They can then use our Transfer Learning Toolkit to refine these models with their own data sets to optimize for their domains. NVIDIA TensorRT, our optimizing compiler then helps optimize these models to run for inference on our GPUs. Finally, Triton inference servers actually helps them run these models. So applications can just send queries and the constraints like the response time they need or the throughput they need to scale to thousands of millions of users and Triton takes care of the plumbing to run these models. Our third pillar is AI application frameworks. They package up our stack and provide end-to-end workflows for incorporating AI into specific application domains for different industries and use cases. This helps democratize the complex AI development pipelines and helps enterprises jump-start the adoption of AI for their use cases. And our application frameworks target some of the most challenging AI applications in large markets, like self-driving cars, robotics, drug discovery, conversational AI and recommendation systems. The scale of inference on the cloud is huge. And we are just at the starting point of AI-infused services. So opportunities to service inference is massive. Running AI inference on NVIDIA is the most performance and cost-effective. And so we continue to see a rapid shift to inference running on NVIDIA GPUs. I talked about TensorRT a bit earlier, our optimizing compiler for inference. The latest version now has over 2,000 optimizations. TensorRT has been downloaded 1.3 million times and is used by 16,000 enterprises. In terms of GPU compute for inference, we've gone from practically negligible 4 years ago to shipping over 166 ExaOPS ops in the last 12 months. This is more than 6x what we shipped last year. Since AWS launched their first GPU accelerated cloud instance 10 years ago, every cloud now offers NVIDIA GPU, and the aggregate throughput has been increasing 10x every 2 years. The chart on the right shows the growth of the aggregate NVIDIA GPU inference compute in the cloud. We estimate that the aggregate GPU inference compute now exceeds that of all cloud CPUs. With this trend, in 2 to 3 years, NVIDIA GPU will represent 90% of the total cloud inference compute. Any AI application and service can now rely on NVIDIA inference, and we are past that tipping point. NVIDIA inference is enabling transformative capabilities for our customers. Microsoft, earlier today, Jen-Hsun announced is adopting NVIDIA AI on Azure to power smart experiences in Microsoft Office, the world's most popular productivity application. The first features include smart grammar corrections, question and answers and text predictions. With NVIDIA GPUs, Microsoft was able to cut down the responsiveness to less 1/5th of a second that enables real-time grammar corrections. GPUs also provide high throughput, so they can efficiently scale to service 0.5 trillion queries they expect to service in a year. Second, cybercrime costs global economy nearly $1 trillion, about 1% of the global economy. American Express alone does 8 billion transactions in a year, totaling about $1 trillion for their 115 million credit card holders. Using NVIDIA AI, American Express uses advanced AI on tens of millions of transactions every day, and it takes just 2 milliseconds to detect fraud instantly. AI fraud detection is going to help save financial industry hundreds of billions of dollars each year and NVIDIA AI makes this possible. Finally, Tencent platform and content group has numerous recommendation systems to support their applications, videos, news, music, applications, et cetera. There are thousands of models that handle hundreds of billions of queries per day. NVIDIA GPU inference enables the use of more and more advanced models in production for Tencent. While I talked about just a handful of customers of NVIDIA AI inference, today NVIDIA AI inference is operating services for companies in a broad range of industries, from automotive to consumer Internet, to cloud-based companies, to robotics, medical, retail and financial services, industrial customers and so on. In many cases, only NVIDIA AI inference makes it possible to deploy advanced AI for production use cases. And in all cases, it saves money for customers. NVIDIA accelerated AI inference adoption has reached the tipping point. With that, I would like to hand over to our next speaker, Manuvir Das.

Thank you, Paresh. Good morning, and good afternoon, everyone. As Jen-Hsun mentioned at the beginning of this call, data center is the new unit of computing. This is because the amount of data that is available to and processed by every application running in the data centers is growing dramatically. This means that applications can no longer fit within an individual server and must be spread across the data center. Paresh has talked about the work that NVIDIA has done over many years now to accelerate particular classes of applications that are running in the data center. He talked about the 3 pillars with AI training and inference as well as the different frameworks. In this section, I want to talk to you about the next opportunity for NVIDIA, the next phase of our work in accelerating the data center, which applies not just to particular classes of applications but to every application workload running in an enterprise data center. In particular, we are talking about technology that will be introduced into every one of the tens of millions of servers deployed in enterprise data centers. If you consider what has been happening in enterprise data centers over the last decade, the infrastructure has moved towards a software-defined model as guided by the advancements that have happened in the public cloud. On the right-hand side, you can see a variety of data center infrastructure functions, which has traditionally been performed by hardware. For example, a firewall located at the edge of the data center or complicated networking infrastructure as well as operations that are being performed manually by armies of humans to manage their infrastructure. Over the last few years, all of these functions have been moved into software, which is now running on every application circle. On the left-hand side, I have a picture showing you the stack that is running within every application server. The box on the top represents the actual application workloads running in virtual machines or containers on a bare metal environment. The box below that represents the infrastructure functions that are now deployed in software that is running on every application server, software-defined networking, software-defined storage, software-defined security, infrastructure management, these are typically deployed in a layer like a hypervisor. A great example of this is VMware, which is used by almost every enterprise customer for their virtualized [ department ]. Every one of these servers is connected to the data center network, communicating and moving data between 1 server to another. This is accomplished with a network interface card or a NIC. As you know, NVIDIA is now working with Mellanox as well as part of our family. Mellanox has worked on the ConnectX family of NICs for many years now, which are state of the art in terms of mix that transfer data across the data center. And they're equipped with powerful acceleration engines to make the networking and IO operations proceed fast. What is the challenge with this model? Here, you will see a representation of these different infrastructure functions that are running in this layer of infrastructure. The challenge we have is that as the amount of data grows, the East West traffic between servers in the data center is growing dramatically. And so if you look at where we are at today, we already see that more than 30% of the resources on the CPU in every server is being occupied with these infrastructure functions, leaving less resources available to the actual application itself. And this problem is only going to grow because the amount of data is growing exponentially. And therefore, the amount of resources required by these different functions is also growing over time. And there will be less and less room available on the CPU of the server itself to run the applications. So a new solution is needed, a new piece of hardware needs to sit alongside the CPU on every server in order to take on this increasing load. We are therefore introducing the next concept in computing that we refer to as the data processing unit or the DPU. The role of the DPU is to take the software function offload them from the CPU and put them into a new kind of chip that we call the DPU. It is an extension of the chip that we already have in the NIC. In particular, along with the acceleration engines that we have on this chip, we have now introduced CPU in the form of powerful ARM cores that can host these infrastructure functions. So as you can see now from the picture on the right-hand side, by moving these functions over to this new chip, the DPU, all of the resources of the host CPU are now available to run the applications, whether they are running on virtual machines or containers of [indiscernible]. It's important to stress here that it is not just about moving the functionality to the DPU because the same situation of the burden growing as the amount of data growing would apply on the picture on the right-hand side as well. However, the DPU is sitting within the NIC. It is already processing every packet of data that is flowing through the network. And therefore, these infrastructure functions that operate on the data can now be much more efficient, they can be dramatically accelerated when they perform on the NIC itself rather than on the CPU, leading to the same effect that we obtained on applications when we moved them from the CPU to the GPU. It is not just about offloading, it is about a dramatic acceleration, which leads to a massive reduction in the amount of infrastructure needed to run the same amount of application workload leading to TCU benefits for the customer. As Jen-Hsun mentioned in his keynote this morning, we have introduced the BlueField-2 as our flagship DPU for this purpose. It provides a variety of acceleration engines for different functionality to do with IO, storage, security. It's got 7 billion transistors on it. To give you 1 example of the power of this technology, if you consider the different activities performed by the DPU and you were to perform them on the host CPU, you would require upwards of 125 x86 cores to perform the same functionality which is not really practical. And here, again, is the reason why we believe that as computing moves forward, every server will be equipped with one of these DPUs to make all of this processing feasible. Now as we move forward with our journey on GPUs at NVIDIA, we introduced the CUDA platform, which was the software abstraction that allowed our ecosystem of 2.3 million developers to proceed forward with a single API even as we advance the technology inside our GPUs. We are doing exactly the same thing with the DPU. And so Jen-Hsun today announced DOCA, Data Center Infrastructure and a Chip Architecture, which is our abstraction and interface for developers to access the capabilities of our DPU. It's based on open APIs, P4 for packet processing, DPDK for the network, SPDK for the storage. These are open APIs on which the developer ecosystem can build a variety of software-defined infrastructure that can now run on the DPU as opposed to running on the host CPU. This is just the beginning of our journey. As you know very well, NVIDIA believes very strongly in the power of AI. The power of AI can be brought to bear on data center infrastructure as well. And so we are embarking on a journey, as Jen-Hsun announced, to take the silicon of the DPU and add to it Tensor Cores from the silicon of GPUs as well, so that we can infuse AI into data center infrastructure. And this leads to a complete road map now for NVIDIA, from which the BlueField-2 is only the starting point. Here's a chart that shows our road map. At the bottom left, you will see the BlueField 2, which has certain capabilities in terms of network traffic as well as its computational power. All the DPUs are based on one architecture, which is DOCA. You can see the time line for our advancement. And on the y axis, we refer to the computational capability. A year from after BlueField-2, we will have to BlueField-3 and then followed by the BlueField-4, 2 years after the BlueField-2. As shown on the y axis here, the BlueField-4, which is a combination of the silicon of the DPU as well as our state-of-the-art GPUs, will have 600x the computational capability of the BlueField-2, it is a significant advancement. But we are not waiting 2 years to bring this capability. Before we get to the point where we integrate the silicon, we are producing new form factors where we combine the DPU and the GPU onto a single form factor or to a single card. This is referred to as the BlueField-2x which you can see vertically above the BlueField-2. It will follow the BlueField-2 to by only a few months, but it has 85x the computational power of the BlueField-2. It is a significant advancement, and it will put AI-infused data center infrastructure on a chip in the hands of the customer base much before BlueField-4 is available. I mentioned earlier that these infrastructure functions that are performed in software today typically live within a layer like a hypervisor. And as NVIDIA and VMware have announced over the last few days, we have a major new partnership with VMware to bring this to the enterprise customer base. As you know, the majority of enterprise customers today use the VMware platform for their virtualized data centers. What we are announcing and the work we're doing together is to take that VMware platform and to move its functionality onto this NVIDIA BlueField DPU, so that the whole CPU can be freed up to perform and run more of the applications. At the same time, we have also announced a partnership with VMware to bring accelerated AI applications to the VMware platform, so that we can truly democratize AI and make AI available in a seamless manner to every enterprise customer. This is a picture of our full stack of collaboration. As you can see on the top here, we are talking about traditional enterprise applications that are already running on the VMware platform as well as the workload that Paresh talked about that are accelerated by NVIDIA GPUs. All of these workloads can now be run on one infrastructure using the VMware platform, which then in turn runs on GPUs as well as DPUs present on every server to provide now both of these forms of acceleration, both the application acceleration for the domains of workloads accelerated by GPUs as well as the DPU-based acceleration that applies to every server and every workload. At the same time, Paresh talked about data analytics as the next frontier of workloads that can be accelerated. We have also announced a partnership now with Cloudera. Cloudera is the predominant platform used by all enterprise customers for their Spark deployments. Cloudera's platform will now be powered by NVIDIA and also accelerated by GPUs as well as in this model by DPUs. So with that, I'll hand it over to Justin to talk about the edge.

Great. Thank you, Manuvir. And good morning to everybody. While the big bang of AI originally happened in the cloud, AI is really about to transform every industry with the wave of new AI infrastructure announced today using the NVIDIA EGX edge AI platform. So I want to talk to you about some of those announcements and put this in context. Today, the Internet connects billions of people to giant cloud data centers. But in the future, there's going to be trillions of devices connected to millions of edge data centers. And this is going to create an Internet of Things that's 1,000x bigger than today's Internet of people. From smart retail to manufacturing and service robots, self-driving cars, smart streets and cities, computing is going to extend from the cloud data centers to the -- to every corner of the world. AI will sense, infer and act accordingly at the edge. The amount of data generated by high-resolution sensors is just simply too much to have the data moved back to the cloud. Some things just can't be done from the cloud as actions need to be immediate. The software powering this new Internet will not be written by humans, but by computers learning from the data. Now, this new way of computing is called AI. And the edge of AI is about driving tremendous acceleration in demand for computing, precisely at the time the Moore's Law has slowed down. And this requires a new approach in computing as legacy architectures just can't keep up. NVIDIA's accelerated computing platform is really the platform that will power the future of every industry. Today, we made several new announcements with the NVIDIA EGX edge AI platform. The EGX -- NVIDIA EGX is designed to make it easy for the world's enterprises to quickly stand up state-of-the-art edge AI servers. NVIDIA EGX can control factories of robots, perform automatic checkout at retail, orchestrate a fleet of inventory moving robots and help nurses monitor patients. EGX is a full stack solution, consisting of the AI computer, system software, AI frameworks, in fleet orchestration and management software. Security is really a top feature of EGX. Every aspect from secure and measured boot of the operating system, protecting data in motion and at rest, securing the applications in AI models with signing and encryption to tamperproof the infrastructure that will automate the future of industries. Today, we announced early access of NVIDIA Fleet Command. It's software-as-a-service for deploying and managing AI services at the edge. Fleet Command simplifies setup in management by ensuring a simple one-touch authentication to connect a new node, so you don't need Linux admins floating around in edge locations. This allows store associates or warehouse managers without IT experience to quickly set up new systems. One of the most important features of EGX though is the rich ecosystem of partners bringing AI to every industry. EGX servers are certified by all leading OEMs and now include NVIDIA BlueField-2 and Ampere GPUs. This ensures that customers can quickly find and easily buy hardware that's optimized for AI performance and can be securely managed and updated through leading enterprise platforms. To fully accelerate the run time of EGX servers, we're working with leading enterprise platform companies, including VMware, Red Hat, Canonical and Suse and others to test and validate that the performance of CUDA X can be delivered for AI training and inference through all of their platforms. We're building industry-focused AI optimized frameworks to make it easier for developers to bring new innovation to every industry. Every industry can benefit from EGX. With our network of partners, EGX can help companies in manufacturing, in health care, retail, logistics and transportation. OEM partners, software partners, industry-focused solution makers can all participate as it's an open platform. This is truly the iPhone moment for the world industries. NVIDIA EGX will make it easy to create and deploy new edge AI services. Now let me share with you the type of customers we've been working with to build and refine the EGX edge AI platform. Previously, we spoke about how we're working with leading companies like Walmart, Procter & Gamble, BMW and Siemens to bring a range of new -- or to bring AI to a range of industries to drive higher business efficiency. And -- but let me highlight a few more that we've announced today. KION Group is the largest logistics and automation supply chain solutions provider globally and operates over 6,000 automated warehouses worldwide. With Dematic, they're developing smart cabinets and adaptive speed conveyors to increase distribution center efficiency and throughput. But still, they're developing automated forklifts. KION is looking to simplify the management and deployment of AI applications to fleets of GPU-accelerated systems used for optimizing warehouse efficiency using NVIDIA Fleet Command. If we look at the retail industry, they lose 1.5% of sales or over 600 -- or $60 billion per year to shrinkage. AI can instantly detect the missed scans at checkout. Kroger is one of the largest super chain market -- or supermarket chains in the U.S. and can operate and operates close to 2,800 stores, with Everseen, a Vision AI (sic) [ Visual AI ] application, they're reducing customer errors and providing faster customer checkout to improve the shopping experience. And every health care provider in the world wants to reduce operating costs while improving patient care. At Northwestern Medicine, Whiteboard Coordinator is being used to operate a network of thousands of sensors, cameras and microphones with perception and conversational AI to help new nurses monitor patients reducing the load on nurses while improving care. So that's our EGX platform. It's a full stack, open platform, state-of-the-art computing, designed for security from the ground up. We have a broad ecosystem of partners to help enterprises around the world create AI services. This ranges from software vendors like VMware, Red Hat, Canonical, Suse, to industry-focused SIs including IBM and Accenture, as well as every major OEM and ODM, such as Dell, HPE, Cisco, Lenovo, Fujitsu and many more. We're working with hundreds of ISVs who are leveraging the power of our AI frameworks to build industry-focused AI for every industry. Every enterprise company understands the power of AI. They no longer need to be convinced why they need it, but they need a broad partner ecosystem to show them how to become an AI-driven enterprise, powered by the world's most widely adopted accelerators in a broad range of ecosystem partners. We're working to bring AI to every industry with NVIDIA EGX edge AI platform. Now let me hand it over to Kimberly Powell who runs our health care business.

Great. Thank you, Justin, and I hope everyone is well this morning. The health care industry is on an extraordinary moment in time. The global pandemic has created the biggest threat to humanity in our lifetime. The race to discover new therapies has never been more critical and today, the health care industry is producing more biomedical data in a couple of months than the last several hundred years. This is a perfect storm to catalyze AI and drug discovery. This is why we're so excited to announce NVIDIA Clara Discovery, a suite of state-of-the-art tools to tackle the most pressing and future challenges in drug discovery. The end-to-end drug discovery process is incredibly complex, starting with biology to understand human disease and why we get sick in the first place; then to chemistry to combine molecules that can inhibit or enhance biological behavior; next to patients and uncovering biomarkers that are the medical science associated with our disease or our response. The pharma industry is huge at $1.5 trillion large, yet it's still very much an unsolved problem, taking well over 10 years, costing $2 billion and still has such a high failure rate of 90%. Drug discovery draws on every computer science domain from accelerated computing, data science and machine learning, deep learning and natural language processing. the Clara Discovery brings together more than a decade worth of work and working with the industry's most popular GPU-accelerated tools like Schrödinger for computational chemistry and where there weren't any tools, we built our own with NVIDIA Clara Parabricks for genomics, Clara Imaging for pathology and radiology, Bio-Megatron and Bio-BERT for natural language processing and NVIDIA Rapids for GPU-accelerated machine learning. This suite of tools is powering the next generation of computational drug discovery to accelerate discovery from months to minutes and using AI in this new trove of data to improve success rates of discovering new life-saving drugs. The U.K. is an epicenter for health care research. Cambridge is home where -- to where Francis Crick and James Watson discovered the structure of DNA, and it's home to the world leaders in the pharmaceutical industry with a rich university and start-up ecosystem focused on health care. Researchers and scientists in the U.K. needed a state-of-the-art computing infrastructure. Because there's no more important time, NVIDIA is building the U.K.'s fastest supercomputer, we're calling Cambridge-1, it's a 400 Petaflop AI Performance supercomputer based on of NVIDIA's DGX Superpod. It will become the fastest supercomputer in the U.K., and it will be in the top 30 on the top 500 and also the top 3 in the [indiscernible]. Cambridge-1 will host our collaborations already underway with the U.K. AI and healthcare researchers in academia, industry and start-ups. Our first partners are GSK, AstraZeneca, King's College London, guys in St. Thomas' NHS Foundation Trust, and Oxford Nanopore Technologies, all of which are already using NVIDIA GPU computing. Cambridge-1 lets them do experiments too large for their infrastructure or a resource while they're building up their own. Cambridge-1 will accelerate the use of AI in the vast and wide health care ecosystem. Also exciting to announce today is GlaxoSmithKline is leading the way in the pharmaceutical industry in adopting artificial intelligence and data-driven drug discovery. We're partnering with GSK in one of the world's first AI drug discovery labs. GSK has been pushing the frontiers in drug discovery and data-driven drug discoveries for years using genomics to prove target selection early in the drug discovery process and a recently established GSK's London-based AI hub. GSK AND NVIDIA together will expand the use of biomedical data in the field of digital pathology, radiology, genomics, natural language processing using Clara Discovery to optimize computational discovery applications. In addition to GSK's investment in DGX A100 system, GSK will also be able to access NVIDIA's new Cambridge-1 supercomputer. And just to conclude, we are in a perfect storm for AI and health care. With the race against time in the global pandemic, the explosion of biomedical data and the utility of AI, we can accelerate health care research and discovery from months to minutes, 16x across many domains used in drug discovery and harness the biggest AI breakthroughs in natural language processing to tap into invaluable biomedical literature and clinical data. Health care vocabulary is domain-specific, complex diseases, proteins and drug names, the case in point is in COVID-19. So this is a tipping point for AI and health care, and we're delighted to be building the industry's computational platforms and partnering with the world leaders in health care. Thank you.

Okay. This is Colette Kress, and I'm just going to do a quick overview for you about what GTC Fall 2020 is all about. First, you've seen us talk so much about NVIDIA's AI and NVIDIA's overall momentum in terms of gaining across so many different pieces. One thing that bounds us all together is really about CUDA and the overall development platform. We have more than 20 million CUDA downloads a year, more than $6 million in loss last year. What we are seeing also is an expansion of NVIDIA and being able to power so much of the overall enterprise. You've seen Manuvir discuss in terms of what we're seeing with our collaboration with VMware and Cloudera as well as the introduction of the NVIDIA DPU data center on a chip architecture software as well. You also got to hear about NVIDIA's edge AI for the Internet and the trillion things out there on the edge. And our fleet command reoccurring revenue service that we will be adding later. And then here with Kimberly, we've heard about our focus in terms of NVIDIA health care as a very key part in both drug discovery and key partnerships with important areas in the U.K., such as GSK. This is allowing us to broaden our overall ecosystem, our customer adoption, every server, every storage OEM, hundreds of ISVs, thousands of enterprise. And just keep in mind here at GTC, we are also exposed to more than 2 million, 2.3 million developers focused on our overall computing platform. I'm going to take this next opportunity to discuss an important piece of the high-level view that sizes up our market opportunities for NVIDIA compute. First, NVIDIA RTX targets the large and growing markets of gaming and professional visualization. Our trailing 12-month revenue in these 2 markets is close to $8 billion and representing an 18% CAGR over the last 5 years. Computer graphics is the first and holistically, the largest application of NVIDIA GPUs. Our graphics growth going forward will be fueled by the expanding universe of gamers, creators and professionals, which already number over 1 billion around the world. One day, we expect every human will be a gamer or connected to others in virtual worlds. Our RTX platform and the Ampere architecture launch this year was a giant step to making that future a reality and a foundation for our growth in graphics over the next decades. While gaming was historically our largest revenue driver, last quarter, for the first time it was eclipsed by our data center platform, driving -- driven by AI. Last quarter was also the first to include our Mellanox acquisition. So let me take this moment to update you on our total addressable market opportunity for data center, which is significantly expanded with the inclusion of Mellanox. We see $100 billion TAM by 2024 across the 4 main markets within data center including high-performance computing, hyperscale and cloud, enterprise and edge. As you may recall, last year, we sized our data center TAM at $50 billion by 2023. So let me help you understand the drivers for this expansion. First, with Mellanox. We are now addressing the large data center networking market with a particular focus on high-performance hyperscale and software-defined environments. This adds over $20 billion to our TAM. Second, as you heard from Manuvir's presentation, we are introducing a new class of processor in the data center, called data processing unit or DPU. The DPU offloads a substantial amount of the processing currently done by the CPUs as well as processing done by other data center infrastructure today. In other words, data center on a chip. This new process adds more than $10 billion to our estimated TAM in this period. And third, you heard from Justin's presentation, we are enabling a merging edge AI market with our EGX computing platform, this adds approximately $10 billion to account. Each of these opportunities is uniquely enabled by the combination of NVIDIA compute and Mellanox networking, and we are delighted to have Mellanox team on board. So that concludes our prepared remarks for today's presentation. We will now turn it back over to the operator, and we will open up the line for Q&A.

[Operator Instructions] Our first question today comes from Aaron Rakers from Wells Fargo.

Yes. I just want to, Colette, if I can, just touch on briefly the TAM assumptions that you're making I guess, first of all, on just kind of the general kind of GPU, TAM, can you give us any kind of framework of how you're thinking about attach rates on GPUs for the data center? And kind of the similar question, as you think about the time horizon, the $10 billion opportunity on data processing units, what's the underlying assumptions of kind of the industry's move to having every server incorporating some form of a DPU in them?

Yes. Thanks, Aaron, for the question. First, let me start off with -- it's a great place to look at in terms of the overall server environment out there and the use of overall GPUs and acceleration as well as AI in many of those servers. Nothing has changed in terms of our view over the future, very similar to everybody would be a gamer. We do expect everything in the overall data center environment to be accelerated, and the movement of AI is getting us there. So it is early in these days to talk about that attachment, we have discussed already our continued growth in that overall attachment, but coming after a very, very, very small base of folks that have been focused on that AI. Our expansion that you've seen us do here in today's presentation in terms of GTC as of overall is really expanding to all of the different types of workloads that are out there as well as the different customers, whether they'd be enterprise, whether they'd be focused on the ad or in the core of all the workloads within that data center to really have a stronger attachment as we go forward. So in summary, nothing's changed. Our goal in terms of getting all of the servers to the accelerated with the use of AI is still there.

Our next question will come from C.J. Muse from Evercore.

I guess, if I could ask two. First for Jen-Hsun. You discussed a number of AI platforms in your keynote this morning. And just to help us prioritize where we should be focused, what do you think are the biggest revenue opportunities looking over the next 1 to 2 years? And then perhaps a question for Kimberly. You announced a partnership with GSK, working with AstraZeneca. You got Cambridge-1. Really curious how you think about your go-to-market strategy and the revenue model for your medical business?

Yes, C.J., the AI started in research. And the first 5 years of the work that we did and most of our conversations that we had was related to the groundbreaking work that was being done, super human image recognition, super human speech recognition and now near human natural speech synthesis, the ability to process data at a scale no humans can so that it could predict recommendations, conversational AI because the recent take through the natural language understanding. The first 5 years was really focused on groundbreaking work and the early works of self-driving cars, robotics, et cetera. The first wave of economic growth of AI, the economic impact of AI was in the cloud. And I would expect the next couple of years to still unquestionably still be in the cloud. The vast majority, not -- I would expect that the next couple of years not only will the cloud grow in very significant percentages, but off of a quite a large base now, and it's a multibillion-dollar business. And as we said it is now past the tipping point, where any service, any application can take advantage of NVIDIA GPU inference because it's in every single cloud and it's in such large abundance. The amount of computation, aggregate computation NVIDIA GPU for inference now exceeds and cross -- surpass CPU and it's growing at a factor of 10 per couple of years. In a couple of years, 90% of the world's total inference compute capacity would be GPU accelerated -- NVIDIA GPU accelerated. And so I think that the -- as you've seen in other type of platforms, once it reaches the tipping point, the acceleration of adoption actually goes up. And for obvious reasons, because people feel really, really safe now that they could take advantage of it because they can always count on the NVIDIA architecture in every cloud and abundance of it in every single cloud. And so I think the next couple of years, you'll see NVIDIA AI growing in the cloud and service providers into large numbers and ever larger numbers. The next wave is enterprise. And enterprise, we described in 2 ways. Enterprise is helping to automate company, which is a lot of the work that Manuvir was talking about. It requires us to re-architect the data center, the system software has to be re-architected. Once -- and the reason for that is because enterprise software and the enterprise data center infrastructure is very different than that of clouds. And so we have to work with partners, particularly the VMware, to do a lot of computer science around the current stacks. And then, of course, the data analytics applications. We're going to grow in enterprise even before we finish with VMware because many of the early adopters are perfectly comfortable building multiple infrastructures, but that's going to get turbocharged incredibly when the work that we do with VMware over the next several quarters come to market. Meanwhile we're preparing the ecosystem as we speak. The next wave is edge and autonomous. And so the -- now that the enterprises and the companies are comfortable and have mastered processing large amounts of data, they're also collecting a giant amount of data and that data is connected to sensors or web services or applications or products. And before you know it, these things that are out all over the world today. It could be anything, it's -- a lawnmower, a refrigerator, an elevator, air conditioner, you name it, they're all going to be connected to either 5G or WiFi. And that allows them to collect even more data and turn all of these products into essentially smartphone connected smart devices. And so the iPhone moment is coming. That's the wave after that. And so we're preparing for each one of these waves. And they'll -- we hope that they happen just bam, bam, bam and just keep on happening over the course of the next 5 years, surely. But you're looking at one of the largest computing opportunities ever. Kimberly's got the next question. Go ahead, Kimberly.

Yes. Thanks, Jen-Hsun. C.J., thanks for your question. So a bit about health care's go-to-market strategy, it's much like all of NVIDIA's enterprise go-to-market strategy and industries. If you think about what we're building with Clara, we are building a domain-specific computing platform for health care, for computational health care. And we deliver what we call a full stack, silicon systems and software. And that software, if you think about what I described in Clara discovery, there are aspects of drug discovery where there's still just an insatiable demand for compute. Doing that -- looking at exabytes of genomic data in the near future and trying to use artificial intelligence to extract associations out of these very large populations of genomic data or looking at the vast chemistry space of 10 to the 60 potential combinations and using a combination of search, to docking, to simulation, all done in silico today. We still can totally consume the world's fastest supercomputers in doing that work. And we're doing that with COVID today. The systems of COVID are more complex than we've ever simulated. There's still an insatiable demand there. And then as we move into these new budding areas of natural language processing, as Paresh talked to you about, the fact that they are growing in complexity and size to train these models every 10 months, it's no different. I mean, these models, biomedical, specific natural language models take tens of thousands of GPU hours to train. So we have just an incredible amount of opportunity ahead of us, and that's what we're really inspired to do with the Clara platform. We hope to catalyze and quickly disseminate the capabilities through building out of Cambridge-1 and enabling the industry leading researchers with it. And then longer term, over time, we have plenty of opportunities to monetize across all 3 systems, silicon and software.

Your next question will come from Vivek Arya from Bank of America Securities.

I actually had 2 as well. One for Colette first and then one for Jen-Hsun. Colette, I was hoping if you could give us a sense of the supply situation for Ampere on both the data center and the gaming side, good to have lots of demand, but just how is the supply situation working out? And there, if you could talk specifically to the gaming side as well? And then Jen-Hsun, my question is, where does Arm fit into your data center vision? Because from what we heard today, if more of the workload and value are going to the preprocessing or the DPU or the SmartNIC which I realize contains some embedded CPU, but more of the value is going to that DPU and various kind of accelerators and GPUs. Does it matter to you one way or another, whether it is Arm or x86 that's a CPU architecture of choice? In other words, is it really critical to owning Arm? Or do you think you can achieve similar levels of success by just optimizing DPU and accelerators because that's where most of the value in the data center is shifting anyway?

Vivek, so thanks for the first question to discuss in terms of supply. We're very comfortable with the supply and where we are in terms of that supply for our outlook that we have provided. When we turn to our overall data center and we look at the overall Ampere architecture, keep in mind, the A100 and that going forward is a very complex product. And it will probably take multiple quarters to really work through all the supply needs and get that to market in its full capacity. You're also focused in terms of Ampere architecture for overall -- for gaming. We're in the initial stages of ramping that. It will take some months for us to fully supply to the channel, but we are right on track in terms of providing that. Sure. We'd love to have more supply sooner when we're ramping, but we're also executing to our plan. So we feel very comfortable with the supply and what that means to our outlook.

We can achieve extraordinary success and all of the success we've talked to you guys about without Arm. However, with Arm, there are some really exciting things we can do. Let me highlight 2 of them. The first one is extending NVIDIA's architecture accelerated computing to the Arm ecosystem. You might notice that we've -- that accelerated computing is surely here. It is past the tipping point. And everybody acknowledges that this is the way to go forward, that Moore's Law has ended. It's ended last week, it didn't revive this week. And in order to extend computing, further, we have to bring accelerated computing to all device, all computing devices, including Arm. The benefit of owning Arm is that we could also offer NVIDIA accelerated computing in soft IP form, not just hardened IP form but soft IP form. NVIDIA is an IT company, we're not a chip company. NVIDIA is an IP company because I'm pretty sure that TSMC makes the chips, and I'm pretty sure that what we delivered to them was effectively an e-mail after completion of a multibillion-dollar project. And so NVIDIA is a soft IP company, whereas -- we're an IP company. And the benefit of having Arm is to have this team that has a vast network to the ecosystem of Arm and all the devices that they're in, the billions and billions of units that are sold every year. And we can extend Arm with accelerated computing and AI computing that NVIDIA is renowned for. Second, we are going to bring a lot of platform technology to Arm in a whole bunch of new data center environments. It could be high-performance computing, cloud data centers, enterprise data centers as we were talking about earlier with DPUs, edge data centers, a whole bunch of new technologies that we're rolling out. As we turn the CPU core of Arm, which is world class, I mean, this is absolutely the most energy efficient CPU core the world. As we turn the CPU core into computing platforms, we're going to bring -- we're going to deliver a lot of value to Arm. We're going to create a lot of value in Arm beyond the mobile device. These are all markets that I'm talking about that are really nascent. And as we create all of that value around the Arm platform, it would be great to own it first. And so I think we have 2 enormous opportunities to extend NVIDIA's accelerated computing to a large ecosystem around the world. And secondarily, we're going to create a lot of value around data centers and servers and computing platforms that are very nascent to Arm, and we would love to own it as we create the value around it.

Our next question comes from Timothy Arcuri from UBS.

I had 2 as well. I guess the first question is for Colette. How much of the $100 billion TAM is China? That would be my first question. And then the second question for Jen-Hsun. And really, it's expectations around your share of that $100 billion TAM. Is there a way to sort of handicap? What would be a reasonable share assumption within that nongaming TAM? I guess I asked because if I sort of take what your nongaming and your non-Mellanox revenues this year relative to the TAM that you set forth a few years ago, it seems like you're maybe mid-teens to like 20% of that TAM. So I guess the question is, would you be disappointed with say, 20% share of that TAM by 2024? And I guess, ask it a different way, is it really that you expect to gain share within that rising TAM? Or the story really is riding the growth in the TAM?

So let me first start with the question regarding our TAM and a regional breakout. We don't have the ability at this time to really look at that opportunity by a region. But keep in mind for each of the areas that we're focused on, we have the opportunity to take that to each and every single region. We can look at it by the additions that we have added with Mellanox, with the DPU and with the edge or we can look at it in terms of the types of customer markets that we will also be addressing. This is an opportunity for us to focus on both our hyperscales and the massive expansion in terms of the cloud. That will be a considerable portion of our overall TAM as a whole. Additionally, you can think about the overall enterprise opportunity. A lot of growth just announced today in terms of key areas of the enterprise that we can have a focus on. High-performance computing has been a big part of us for more than 10, 12 years. So again, expansion in terms of bringing further acceleration in AI to that is also an area. And now the edge, and you can think about those devices even in all of the regions has the opportunity to grow. So we've expanded each and every single one of those different areas today in terms of our increase to $100 billion with the Mellanox, the DPU and the overall edge. And yes, a region, such as China or the U.S., can definitely benefit from that.

We address the entire TAM, except for x86 CPUs. That's really the simplest way to think about that. And in almost all computing platforms that we serve, accelerated computing is the most important part of it. And going forward, if we believe in the thesis, 2 important thesis, that the first one, all applications in the future will be infused by AI. For example, something like Microsoft Office will be infused by AI. Then I would suggest that AI would be in every computer. And I would suggest that every computer will be accelerated. I am certain every computer will be accelerated, in fact, just as I was convinced 30 years ago that every PC would be accelerated with GPUs. And so I'm certain of that. I am certain of the fact that CPUs alone will not do the job. That is complete certainty. And so I believe in the thesis of AI, I believe that AI will be in every application, and AI will open new markets and create new applications that weren't writable before that no humans know how to write. And so that, I believe, and therefore, accelerated computing is going to be everywhere. And in all the platforms that we are in, accelerated computing, the GPU and the networking, frankly, dominates the vast majority of electronics inside those competing platforms. Second, I believe in the world of zero trust. I believe that protecting data centers at the perimeter is historic. It's like building a big wall. It makes no sense. That the future of security is about zero trust, and it's about securing every single transaction, every single node, every single application. And therefore, you need to take what it used to be security appliances at the perimeter of the data center and put it into the servers, every single one of them, every single network. That's the reason why the networking chip is going to be the most important security chip in the future because that's where all the input and output comes from. That's exactly where you want to put it, and that's exactly why the GPU is invented. Security is going to force every single computer on the planet to have something like a DPU. And so I believe that every single server node will be accelerated by a DPU. And every single server application will be accelerated by GPU. And therefore, the vast majority of the world's TAM, minus the x86 CPUs would be our opportunity.

Your next question comes from Stacy Rasgon from Bernstein Research.

I have 2 as well. First, to go back to your chart showing GPU inference workloads in the cloud versus exceeding all CPUs. How does -- I guess, given the trend, how does your cloud revenue today split between inference and training? And how are those each -- like what were the trajectories of growth there relative for those 2 relative to each other? The second question, can you tell us a little more about how the revenue model for the VMware partnership works? Is it just working to incentivize broader GPU use in enterprise? Or is there a more direct revenue share something about that?

So let me see if I can...

Can I answer both of them?

Yes. Why don't you give it a shot.

Okay. The reason why our inference -- well, first of all, do you know that our inference. Hello?

Hello, yes.

Yes. We've got some reverb. That's okay. The first thing that's, of course, Stacy, you know that our inference acceleration business has been growing very, very quickly and -- for all the reasons that we've already talked about. We turbocharged it even more this time with Ampere because Ampere is our first universal GPU. We used to have essentially 3 GPUs in data center. And one of them would be very heavy-duty training systems to build these large models, AI models. The second is a cloud training platform, which is based on PCI Express, like the new GP that we announced today, the NVIDIA A40. It's going to go into all the clouds, and it's easy to deploy. It's easy for them to put into all their PCI Express servers because it's based on PCI Express. The A100 is based on SXM2, a different type of networking and a different type of system architecture. In fact, very different. And so the second, which is cloud-based training was our second GPU. And our third was our inference GPU. Well, what we did with Ampere as we combined it into one architecture. And so you could literally now use the Ampere the A100 SXM based for both training as well as cloud training as well as inference, one single architecture. And then -- and if you like, if you would like to have PCI Express versions because your cloud data center is able to facilitate a lot more PCI Express servers. And we have the A40 GPU, which now allows you the -- A100 and the A40 GPU, depending on sizes, that allows you to do both training and inference. So we now have an architecture that's universal. It does training, it does inference, it does computer graphics, it does all the things that you know that we do very well, all into one architecture. And so that's the reason why our inference performance is going to not only continue to grow at the historic rate, which was really, really high, both in number of units and the fact that we're increasing throughput by a factor of 20 generationally. We now have the ability to put a lot more unified aggregated GPUs in the cloud that are inference based. And so that strategy was a really good strategy. And Stacy, what people want is they want to make sure that if they were to use a cloud, develop software for a type of technology, they develop software for capability, it could be video decoding, it could be whatever it is, x86 or Arm or in our case, NVIDIA's accelerated AI. Whatever capability they use in the cloud is available in every cloud, so they have the flexibility, which that is established. And number two, abundance of capacity. And that's why the tipping point of our aggregated AI influence throughput is such a big deal. And I think at this point, because of the rate that we're growing, it's a foregone conclusion. It's going to be the vast majority computing and cloud. Gosh, what was the second question?

The second question was on VMware and...

Monetization?

Yes, VMware. There are 3 ways that we benefit. And then I'll leave the most important one last. The first way we benefit, of course, is VMware running on -- VMware, as you know, is the data center operating system. They represent 70% of the world's data centers. This operating system is really computationally complex these days. And the reason for that is because of software-defined data centers. The networking stack, the storage stack, the security stack, the virtualization stack, it's all running in VMware. And so the first thing I'm going to do is going to offload and accelerate and isolate the data playing from the application plane. And that offload alone creates really fantastic opportunities for our DPU. It is cheaper, it is purely faster and unquestionably more secure to have VMware running on x86 plus a DPU instead of x86 without a DPU, okay? So the first is creating an opportunity for our DPU. Second is every one of the VMware stack goes with a virtualization stack of our GPUs. And that hypervisor, there's the VMware hypervisor. And then there's simply the NVIDIA hypervisor for virtualizing all of our GPUs. The virtualization of GPUs is really complex. And it opens up CUDA, opens up our graphic stack, RTX, it opens up Cuda-NN, opens up all of the capabilities that is buried underneath the hypervisor otherwise. And so the second thing is it opens up the virtualization stack, we call it the GPU for all of our data centers, okay? And so the third and this is really the biggest one, which is the ability for enterprises to be able to accommodate all 3 domains of computing from scale out, virtualized to micro services. All 3 of these harmoniously and basically transparently in their data center using VMware. And so the ability for enterprise IT to easily adopt and video accelerated AI is now -- you don't even have to think about it. It was just like before. VMware has never fully transparently integrated in video GPUs or any accelerators aside from the CPU until now. This transition is a big, big deal. And it opens up great opportunities for VMware into all of the worlds of AI, opens up great opportunities for us to be able to transparently, seamlessly, easily integrate NVIDIA AI into our world data centers. So 3 ways.

Your next question will come from John Pitzer from Crédit Suise.

Just 2 questions here. Jen-Hsun, both in your presentation and I believe Paresh's presentation, you talked about software writing software and sort of this iPhone moment. I'm wondering if you could just help elaborate on that a little bit. Is the analogy here that you're sort of the iOS and you'll be collecting a recurring revenue stream on some of these sort of AI apps the same way Apple does? And if so, should we expect to hear about other Mercedes likes deals coming down the pipeline? And as you think about monetizing software, is that part of your $100 billion incremental TAM that Colette talked about this morning? Or is that above and beyond? And then secondly, for Colette, you talked about kind of Ampere being extremely strong. And you guys clearly guided to that in October. You also guided gross margins to be under a little bit of pressure as you ramp the new product. I'm just curious, is the strength that you're seeing just that demand is outstripping supply or you're having some supply issues as well? And how should we think about some of your gross margin comments?

Our $100 billion TAM does not include many of the things -- well, almost all of the software things that we've talked to you guys before, all right? It didn't even include what I mentioned just now to Stacy about our virtual compute. It doesn't include GeForce now. It doesn't include DRIVE. It doesn't include Fleet Command. It doesn't include our software stacks for enterprise. And so we'll have plenty of opportunities to talk to you guys about that in the future. Today, we want to just stay a very, very focused, keep it nice and conservative, and we have plenty, plenty to talk about. NVIDIA is a full stack accelerated computing company, as you know. And our opportunities -- and we're an open computing platform. We work with our network of partners in any way they would like. And for some, they would like the whole stack. And the reason for that is because ours is just so good. We're so good at it. For some, they would like to develop some parts of it and use ours. And so we work with them to figure out which part of it they would like to use of their own and which part of that they would like to use ours. And some people would like to build all of their own and use a lot of our open-source tools or our libraries, like CUDA-X or RAPIDS open source stacks to build their data analytics services. And so we're an open computing platform that works across the multiple layers from the technology to the system, SDKs to the application frameworks. We want to be able to work with the entire world's industries and democratize AI and bring accelerated computing to as many places as we can. And -- but what we've only captured so far in our TAM is the hardware stuff, the hardware stuff, which is...

Just on the Mercedes deal, as we run numbers, it's not hard to envision kind of a software recurring revenue stream which is as large as the hardware revenue stream, is that how we should think about the potential for service and software recurring revenue relative to that $100 billion TAM?

Yes, that's a great way to think about it, if not more. And the reason for that is this. And I don't mean for you guys to go make any changes in your models, okay? But just listen to the strategy and think about the implications and think about what we're going is. I'm building it all in public and all these pieces are being talked about at GTC. And so you guys know where we're going. There is no question that in the case of autonomous vehicles, the business model that we've created with Mercedes is really quite extraordinary. It's potentially the best one. And the reason why is because, first of all, it's incredibly hard. To be able to create a computing platform that integrates into the most safety concern, the most safety conscious companies and industries with accommodation for all of the heritage and all of the existing technologies to be able to integrate into that, infuse into that harmoniously is a great challenge. And so that's the beginning. That took us 10 years to learn. Now that we're inside the car and we're the computing platform, then we can create the applications that sit on top of it. Because unlike a smartphone, you're not going to be able to create an application in the cloud and download and works in every single android phone. It's just not going to be like that. And the reason for that is because of safety. These applications are safety concerned first. And so each one of the application opportunities belong to the car companies. And so this is one of the reasons why such an extraordinary thing is if Daimler could figure out a way to make all of their -- all of the cars software defined and turn them into application platforms, they'll grow that application platform 2.5 million cars a year. They'll grow instantaneously. Over the course of a decade, you could just imagine the economic opportunity that they're going to create. Our business model with them is to share that. And so we're doing a lot of the hard work, of course, as well, and that becomes one of -- that's a significant opportunity for us that we would like to do this. We're an open platform company. And we would like to do this with other car companies. And there will be other opportunities. And the reason for that is very simple. The number of companies in the world that could create an end-to-end self-driving car stack that is world-class and that you can deliver on real streets is very, very few, that integrates into the existing car industries, very, very few. And so I think this is a great opportunity, and we're going to continue to scale it. But you're going to see other examples like that. You're going to see other examples like that. We'll always have free community versions. That's just one of the policies we have. The community versions will be free to develop, a version could be free. There will always be free versions. But for some companies, they want to make sure that we're on the hook on it, and they want to make sure that we have some kind of an enterprise agreement and an enterprise business model that allows them to get in front of open source or allows them to get in front of developers and others, so that they can get their software enhanced or software debugged.

And let me see if I can touch based on your second question that was regarding overall Ampere demand and the impacts in terms of our gross margins. So, so far, the gaming demand for our RTX 30 series has just been off the charts. We had expected a really great holiday season. We knew that our overall platform that we were bringing was the best generation to generation performance ever. We've got a great release of fall games that are coming out. And the work from home is even bringing more and more to the gaming and the entertainment and social arena. So we're racing to catch up to that demand. But the ramp is going well. The yields are very good. So all of that is intact. Now when we think about our gross margin, let's remind what we did in terms of our gross margin outlook for the Q3. Our outlook for Q3, and as usual, with most quarters, mix is our driver of overall gross margin. We expect a very strong sequential increase for our gaming. And with that gaming piece of our business being so strong, we took a slight sequential dip in terms of our guidance for gross margin. Everything seems to be in place, intact. So no change in terms of our gross margin in terms of what we're seeing.

Your next question will come from Mark Lipacis from Jefferies.

The question, I think this one is for Jen-Hsun. To provide an integrated data center scale architecture, I'm trying to make sure I understand how far across the data center value chain and how deep in the value chain at NVIDIA -- you feel NVIDIA has to go in order to deliver that. And I think it's pretty clear NVIDIA is not a chip company, but a platform company. But maybe if you could compare what you feel you need to deliver across that value chain today to today's data center value chain? Is this -- is NVIDIA effectively becoming the equivalent of companies that are selling processors in the data center today, companies on the servers, on the software, the networking companies, on the software, from the OS, application side, how far up the software stack you're doing? If you could provide maybe if you had an analog in today's data center versus the data center scale architecture you're delivering in the future, maybe that would be helpful to level set.

Sure. The first -- there are 3 questions I'll hit right away, and then I'll explain. One, we're not like a company that exists today because AI is a problem and an opportunity, a challenge and an opportunity unlike any software that's ever been written before. Otherwise, why could we do all the things that we're doing right now, number one. Number two, we are not doing integrated data center, we're not doing that. Number three, we innovate as much as we need to and as little as we can. We innovate as much as we need to as little as we can, which is our guiding principle of NVIDIA, to do as little as we can. We're not trying to do everything. We're only trying to do the things that we have to. We're only doing the things that the world relies on us to do, is that we're only doing the things that we do. If we don't do it, the world just doesn't have it. It is absolutely the case that if we don't do what we do today, the world doesn't have it. It is absolutely the case that when Gelsinger and I worked on the VMware partnership with us, if the 2 of us don't do it, it just doesn't get done -- it just won't get done ever. And so we have to go and do the things that we do because unless we do it, the world doesn't have it, okay? So I answered those 3 questions very quickly. Let me now get back off a little bit and explain. One, the way that AI is written, the way that AI develops software is a computer that's off learning from data, it's learning from data that we collect. And we ask, we coach it, we influence it on the type of neural network architecture and the type of data that we present it and the way that by which we want it to learn. We coach it. We coach it, we're like a coach, we're like a teacher. And then it goes off and it runs for days and weeks, and it does it over and over and over again on giant amounts of data, and it writes the software. When it's done writing that software, we can't read it, it's unreadable. It's like a brain dump of somebody's brain, and it's not readable. And it requires a new type of computer to run it. And so from the way that the software is written, the methodology of which it is written, the infrastructure pressure it creates. And during the past, there were some really -- first of all, the 4 speakers were fantastic today, I really appreciate their work. And you could hear in their phrases, the tip of the iceberg of the challenges in computing that we're solving. And so the software is different, the way it's written is different, the tools are different, the pressure on the infrastructure is different and the coding that is different. In no time in history, did we see that all of a sudden the data center now, the enterprise data center, as it wants to be hybrid cloud, the enterprise data center has to manage 3 computing environments. That's never happened before. One computing environment is bare metal, they allow, distributed competing like a supercomputer. Number two is virtualized multi-tenant, virtualized, easily manageable, easily scalable, easily secured multi-tenant and third containerized microservices, deploying far out at the edge. You'll never visit it again. You drop the server, you connect it to your network. And hopefully, you never go back to that warehouse, and you never go back to that store room ever again. And you manage it from one point glass very, very far away. These 3 types of computing domains has never happened in one company before. We've got to go make it happen. And it's never done it for AI, and it's never done it for GPUs. And so we have to go create the necessary pieces. When we are done creating it, when we're done creating it, and this is very, very important, when we're done creating it, we open it up as SDK. We open it up as SDKs, chip SDKs, system SDKs, APX, EGX, AGX, they're all -- they're hardware components that OEMs can integrate. Then we put on top of it, our entire software SDK. And then we put our libraries on top of that. And then for the application developers, you create tools like application frameworks, which are basically AI skills that we pre-train. All of this stuff is put into the cloud. All this stuff is put into the cloud and is all certified, it's all optimized, it runs in the data center, it runs in the cloud. We -- as a result, can connect up a network of partners of software developers, system makers, solution makers, cloud service providers and partners all over the world, and they can all run NVIDIA AI, they can all run NVIDIA accelerated competing, they can all run NVIDIA Clara, they can all run NVIDIA Rapids, they can all run NVIDIA Isaac, Jarvis, Merlin, all of it, SDKs that we created. So I hope that's helpful. I know we look very different. However, accelerated computing needs to look different because Moore's Law is finished. Number two, AI is different because it's not written by a human, it's written by a machine. So the world has changed. That's why a new type of company needs to be created.

Your next question will come from Matt Ramsey from Cowen.

Jen-Hsun, I had a couple of questions for you. The first one, I noticed in the BlueField DPU road map, eventually, you guys integrate sort of the full AI engine into the DPU. And so I wondered if you could talk a little bit about on the acceleration side, what AI opportunities and low-hanging fruit might be available in -- particularly in the security domain? And then the second question is, you now have AI acceleration, you have a DPU acceleration. For an integrated stack, I mean, whether you buy Arm or don't buy Arm, it seems like you could make a CPU. And maybe you could talk a little bit about the pros and cons of going after that one piece of the TAM that you're not addressing today.

Yes. Excellent question. Number one, giving you an example, intrusion detection. Intrusion detection will be distributed not at the perimeter. The data center will be protected at every single transaction, at every single node, at every single application, it will not be protected only at the doors. And the reason for that is, don't forget, all of the intruders are largely inside the building already. In the future, you also have public clouds. The entire data center is open to the world. You can't allow to have east, west intrusion. The moment an intruder goes inside a data center, go sideways, East West into the data center, imagine the damage they can do. Security is incredible. Every single node will become a super firewall. Firewall technology today, intrusion detection technology today based on AI. We've got to put AI processing right at the network, number one. Number two, network shaping, network traffic shaping. That's an AI problem. It's an optimization problem that cannot be done with a simple set of equations. And it's a heuristic problem, it's a dynamic problem. It's one of those computer science problem that goes, well, it depends. What's the solution? Well, it depends. And so the well it depends requires intelligence as we want to put intelligence right at the network. What computer science is and what we've described in the past as in network computing, these are 2 examples of in-network computing, okay? So a very, very big deal. We're super excited about doing that. And this is one of the reasons why and if you go back to the very early days when I talked about the acquisition of Mellanox, I talked about in-network compute. And I talked about how the network itself will become a fabric where we do a lot of computation, a lot of AI. This is the first step. There are several things that we can do -- there are many things we can do with Arm. There are many things we can do with Arm, can do with Arm. And we can build a CPU. However, there will never be another CPU built like Arm, ever, and it's not a computer science problem anymore. It started out as a computer science problem 30 years ago, an energy-efficient architecture that's designed like no other. That was visionary. And the reason for that is because if you are energy efficient and the world hits the wall because of the end of Moore's law, you've got more runway. There's more runway in Arm than there is in x86. There's just more. I mean that's just the bottom line. The architecture is more energy efficient, therefore, they got more runway. However, Moore's Law has ended for them as well. And so we need to bring accelerated computing to Arm. Arm, number one, is just the genius of an architecture, but it's also a gene has of a business model. And the reason for that is because they wanted Arm to be the most popular CPU in the world. They wanted to be used in all kinds of things from -- well, from everything to everything, cars, the phones, the televisions to you name it. And so that required a business model that allowed them to license their IP in soft form, in soft flexible form that fits into other people's chips because many computers in the future are full -- it's just the whole data center is on one chip, the whole computer is on one chip, the phone is on one chip, TV is on one chip. There's no such thing as 2 chips, just on one chip. And so the second thing that they have because of the business model created the third thing, which is ultimately the most valuable thing today, which is their vast ecosystem. The execution machine of Arm that knows how to build soft IP, their IP for smartphones, for embedded systems, micro controllers to data centers and now increasingly PCs, there are a number of CPU cores, the engine they have behind it that creates the soft IP, productizing it and delivering it to customers, helping them integrate it into their chips, that phenomenal. That's phenomenal. And as a result, they created this ecosystem of thousands of chip companies. They shipped 22 billion chips last year. NVIDIA shipped 0.1 billion. So the difference between NVIDIA and Arm is 22 billion. And so that just kind of puts it in perspective the reach of their ecosystem. That's the value to us. You can't do that by building another CPU. It doesn't matter what another CPU is. I don't care what it is. I just don't think the ecosystem will ever be as rich as this one. It took 30 years to build. It took enormous character, enormous vision to built it. The team that built it is phenomenal. They love Cambridge. They work in Cambridge. It's a great computer science team. I love the work that they've done. That ultimately is the asset that we're buying, that combination of all that and the ecosystem that they've built up that just simply won't get replicated again.

Your next question will come from Harlan Sur from JPMorgan.

One of the powerful dynamics that the team is creating for itself is leveraging the entire portfolio to target vertical markets. And so a question for Kimberly for the vertical market focus, like health care, the drug discovery opportunity that you talked about is primarily focused on high-performance computing platforms like DGX, but moving across the portfolio, how is the team leveraging the edge, an influencing portfolio like EGX and Jetson product families within your health care franchise? And how involved is your team in helping to define next-generation hardware and system platforms?

Yes. Thanks, Harlan. Thanks for the question. So in drug discovery, you're right, the DGX systems as it's a full stack architecture, it does literally cross every computer science domain. Yes, it's very heavy and upcoming in artificial intelligence, but of course, it takes advantage of accelerated computing. It's also going to be very instrumental in data science and machine learning and data analytics. As we move into these gigantic data sets or genomic data sets or even doing the compound screening, what we do is we generate huge outputs of that analysis that needs analytics to really pull out the necessary information. So we literally leverage every corner of NVIDIA's extreme and powerful and world-class computing architectures, whether it be accelerated computing, data science, machine learning, deep learning and natural language processing. And the other areas, and Justin touched on it briefly. We have, for decades, been working in edge devices frankly, revolutionizing the medical instruments that care for us and see inside our bodies and extract our DNA and build out the 3 billion letters that make us up. And so these instruments are one of the edge platforms that in the future, just like our phones and our cars, they want to be software-defined. You -- this sensor technology that has created in these edge medical instruments is incredibly powerful, and we continue to get amazing insights and new information out of this sensor technology by applying artificial intelligence. But it can't remain the old way of deploying these instruments where you would sell a couple of million dollar CT scanner, and it would refresh every 10, sometimes 15 years. We can't carry on that way anymore. So edge computing is absolutely going to be vital to what is going to be the software-defined future of medical instruments. And then you can imagine, not only the new instruments will have -- will leverage everything we've built in our Jetson platform, but it will also leverage what we're building in our EGX platform. Being able to remotely manage provision and securely operate edge nodes that need to be updated with new AI applications over time is extremely vital. The example Justin gave was at Northwestern Hospital, where actually, there are plenty of sensors that already exist in the health care environment, microphones, cameras, that can now be coupled with artificial intelligence and conversational AI, so that brand-new services can enter the health care and hospital environment, just like we would expect, just like we have at our homes, we can talk to smart speakers. We're unlocking and enabling that environment in the health care system today using what is DeepStream for -- that's used in smart cities, what is Jarvis that is being used in a lot of our conversational AI platforms. We can leverage all that technology and create a domain-specific application framework to over -- literally overnight allow application developers to develop new applications that can be deployed and then leverage the Fleet Command system of EGX to deploy them in the tens of thousands of installations and environments that they're going to want to live at the edge. So whether it's new instruments, augmenting existing instruments with compute, coupling all of the amazing sensors that already live in our health care environments with AI and then being able to manage securely and deploy applications with EGX. The future is incredibly bright. And we see smart hospitals now in these applications popping up literally out of the wood work, of course, as you can imagine, to respond to the great demand that the pandemic is putting on the health care system.

Your next question will come from Raji Gill from Needham & Company.

Just when we're thinking about this new cloud class of data center products, the DPU, how do we think about kind of the pricing dynamic in terms of the revenue opportunity relative to other class of chips that you're selling. Just trying to understand how this will be kind of integrated and what the go-to-market strategy will be for that?

Yes, there are 2 pillars that we could look towards. One, of course, is the baseline, which is what a SmartNIC goes for, it's a few hundred dollars. And then the other is the amount of CPU offload that you provide such that the application performance of the server gets the boost. You're effectively going to add a DPU to a server. And my expectation over time is that you'll double the performance of the server. And that kind of puts the value up. So some of it goes to is pricing, and we'll work it out as we build.

Your next question will come from Ambrish Srivastava from BMO.

Jen-Hsun, I had a question on inferencing. And then I had one on gaming as well, maybe Colette, you could answer that. So on the inferencing, you shared a pretty revealing piece of data in terms of crossing over of CPU. And then you gave a projection for a 90% market share. So Intel is the incumbent. And so could you please just help us understand kind of in terms of how much of the training you could translate into NVIDIA inferencing, which gives you the confidence that you get to a 90% share? And then also, what are you assuming the competition does in that market? And I know Jen-Hsun, you don't take competition lightly. And then on the gaming side, Colette, I know in the past, such events, you guys have been very kind in giving us the components of the CAGR in terms of units and ASPs. So I was wondering if you could help us out there as well. And then you also said that gaming -- sorry, I'm asking multipart question, but you said this is a fastest ramp ever. So if you could just give a little bit more details around that will be helpful.

Yes. Yes. So it wasn't actually about share, it was just about compute, aggregate compute in the cloud. For example, in my PC right now, in my PC, I have a decent view. And the aggregate -- my -- the computational throughput of my PC is 99% my GPU. In fact, the aggregate CPU versus GPU compute inside the data center, inside a supercomputer or high-performance computer is about 99% GPU. And the reason for that is because that's its job. Its primary job is this to do the acceleration, primary job is to with compute. And it's not to say that the CPU is not useful, that's just not a job. Its job is to manage the application, manage the operating system, manage -- orchestrate the processing of the applications, figuring out who gets priority. Those things are really important, moving things around, launching things. Those are really important. And if those single-threaded performance, single-threaded application code is not properly processed, then the single-threaded part of your code comes the critical path, otherwise known as Amdahl's Law. And so that's all is that the vast majority of the cloud going forward would be accelerated. In fact, that is important on conclusion at this point. And it's a corollary to what Moore's Law has ended. And therefore, you have to look for another approach to accelerate applications because Moore's Law has ended. And yet, on the other hand, the emergence of the new type of application called AI that requires so much computation at exactly the time when CPU performance is not going to double every couple of years, and you can't just wait for it. And so the world has to lift their code and refactor it and take advantage of acceleration. It happened at a perfectly good time. And the reason for that is because of cloud compute. Because of cloud computing, the world had to refactor its application, and it disaggregated it. And when it disaggregated it, whenever you disaggregate you containerize certain modules, certain micro services, you might as well have accelerated. And you might -- because you're infusing it with AI anyway. And so I think that the confluence of both the end of Moore's Law and the beginning of AI and the emergence of this new type of data center, data center scale computing, we call it, are all working in its favor. Colette?

Yes. So let me see if I can answer your question regarding our split in gaming between our ASP growth and our unit growth. Both of these are very important to our overall growth. And over this 5-year period of time, both of them have contributed. One of the key things to note in terms of what is both influencing our units and our ASP growth is the onset of laptops, notebooks, gaming, high-end gaming notebooks for this market have really grown quite well, and they have great ASPs for us as well. So we continue to uplift overall ASPs as our new gamers coming on board tend to take on the RTX, tend to take on our higher-performing overall GPUs just to start off with. We still provide a slew of different overall price points to attract every single gamer, but you can see our ASP is probably over this period reaching double-digit growth. And there is still a great opportunity as we go forward. We also announced that right now, the overall Ampere architecture for gaming is growing quite well. The launch is probably the best launch in history. We are in an opportunity to do it a little bit different than what we did with overall Turing. Turing was an opportunity for us to address ray tracing for the very first time and really start that market in essentially a chicken and the egg type of market. The hardware availability began the beginning of the overall software that was there. We also had to take a little bit of a pause in terms of some of that launches and really address the whole scope of GPUs over a longer period of time. So we're excited both in terms of the performance improvement that you have with the overall Ampere as well as the great price points. And so far, the launch, the ramp is growing quite well, and we're just really pleased in terms of how things are going.

This will bring us to the end of today's question-and-answer session. I turn the call back over to Jen-Hsun for closing remarks.

This is an amazing time for the computer industry and the world. The age of AI has begun, and NVIDIA is in full throttle to bring this capability to the world. The breakthroughs of AI can now bring automation to the world's largest industries. This new type of software requires a new type of computer to write the software, validate the software, deploy the software. AI is understandably complex, from the chips, systems, software, algorithms to applications. AI requires reinventing every layer of the computing stack. NVIDIA is in full throttle, building the full stacks for each computing domain and each computing environment and from cloud, PC, enterprise, autonomous solution to edge. NVIDIA is in full throttle building a computing company for the edge of AI. Thanks for joining us at GTC.

Thank you, everyone. This will bring us to the conclusion of today's conference call. You may now disconnect.