Earnings Call: Q4 2018

Feb 8, 2018

Good afternoon. My name is Victoria, and I will be your conference operator for today. Welcome to NVIDIA's Financial Results Conference Call. The line phone lines have been placed on mute to prevent any background noise. After the speakers' remarks, there will be a question and answer period. Thank you. I'll now turn the call over to Simona Jankowski, Vice President of Investor Relations to begin your conference. Thank you. Good afternoon, everyone, and welcome to NVIDIA's conference call for the Q4 of fiscal 2018. With me on the call today from NVIDIA are Jensen Huang, President and Chief Executive Officer and Colette Chris, Executive Vice President and Chief Financial Officer. I'd like to remind you that our call is being webcast live on NVIDIA's Investor Relations website. It's also being recorded. You can hear a replay by telephone until February 15, 2018. The webcast will be available for replay up until next quarter's conference call to discuss our fiscal Q1 financial results. The content of today's call is NVIDIA's property. It can be reproduced or transcribed without our prior written consent. During this call, 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 the disclosure in today's earnings release, our most recent Forms 10 ks and 10 Q and the reports that we may file on Form 8 ks with the Securities and Exchange Commission. All our statements are made as of today, February 8, 2018, based on information currently available to us. Except as required by law, we assume no obligation to update any such statements. During this call, we will discuss non GAAP financial measures. You can find a reconciliation of those non GAAP financial measures to GAAP financial measures in our CFO commentary, which is posted on our website. With that, I will turn the call over to Colette. Thanks, Simona. We had an outstanding quarter and fiscal 2018, led by strong growth in our gaming and data center businesses. Q4 revenue reached $2,910,000,000 up 34% year on year, up 10% sequentially and above our outlook of $2,650,000,000 All measures of profitability set records. They also hit important milestones. For the first time, gross margins strongly exceeded 60%, non GAAP operating margins exceeded 40% and net income exceeded $1,000,000,000 Fiscal 2018 revenue was 9 $710,000,000 up 41 percent or $2,800,000,000 above the previous year. Each of our platforms posted record full year revenue with data center growing triple digits. From a reporting segment perspective, Q4 GPU revenue grew 33% from last year to 2,460,000,000 dollars Tiger Processor revenue rose 75 percent to $450,000,000 Let's start with our gaming business. Q4 revenue was $1,740,000,000 up 29% year on year and up 11% sequentially with growth across all regions. Driving GPU demand were a number of great titles during the holiday season, including Players Battle Ground, PUBG, Destiny 2, Call of World War 2, Star Wars Battlefront 2. PUBG continued its remarkable run reaching almost 30,000,000 players and recording more than 3,000,000 concurrent players. These games deliver stunning visual effects that require strong graphics performance, which has driven a shift towards the higher end of our gaming portfolio and adoption of our Costco architecture. Esports continues to grow, expanding the overall industry and our business. And one sign of their popularity, Activision's Overwatch League launched in January and reached 10,000,000 viewers globally in its 1st week. We had a busy start to the year with a number of announcements at the Annual Consumer Electronics Show in Las Vegas. We introduced NVIDIA, BFGB, big format gaming displays in a partnership with Acer, ASUS and HP. These high end 65 inches 4 ks displays enable ultra low latency gaming and integrate our shield streaming device, offering popular apps such as Netflix, gaming video, YouTube and hula. The BFGD won 9 Best of Show awards from various publications. We expanded the free beta of GeForce NOW beyond Macs to Windows based PCs and we enhanced GeForce experience with new features, including NVIDIA Freestyle for customizing gameplay with various filters and updated NVIDIA Ansel photo mode and support for new titles with ShadowPlay highlights for capturing gaming achievements. Additionally, the Nintendo Switch gaming console contributed to our growth as it became the fastest selling console of all time in the U. S. Strong demand in the cryptocurrency market exceeded our expectations. We met some of this demand with a dedicated board in our OEM business and some was met with our gaming GPUs. This contributed to lower than historical channel inventory levels of our gaming GPUs throughout the quarter. While the overall contribution of cryptocurrency to our business remains difficult to quantify, we believe it was a higher percentage of revenue than the prior quarter. That said, our main focus remains on our core gaming market as cryptocurrency trends will likely remain volatile. Moving to data center. Revenue of $606,000,000 was up 105% year on year and up 20% sequentially. This excellent performance reflected strong adoption of Tesla V100 GPUs based on our Volta architecture, which began shipping in Q2 and continued to ramp in Q3 and Q4. V100s are available through every major computer maker and have been chosen by every major cloud provider to deliver AI and high performance computing. Hyperscale and cloud customers adopting the V100 include Alibaba, Amazon Web Services, Baidu, Google, IBM, Microsoft Azure, Oracle and Tencent. We continued our leadership in AI training markets where our GPUs remain the platform of choice for training deep learning networks. During the quarter, Japan's preferred networks trains the ResNet-fifty neural network for image classification in a record of 15 minutes by using 10 24 Tesla P100 GPUs. Our newer generation V100s deliver even higher performance with the Volta architecture offering 10 times the deep learning performance of Pascal. We also saw growing traction in the AI inference market, where NVIDIA's platform can improve performance and efficiency by orders of magnitude over CPUs. We continue to view AI inference as a significant new opportunity for our data center GPUs. Hyperscale inference applications that run on GPUs include speech recognition, image and video analytics, recommender systems, translation, search and natural language processing. The data center business also benefited from strong growth in high performance computing. The HPC community has increasingly moved to accelerated computing in recent years as Moore's Law has begun to level off. Indeed, more than 500 HPC applications are now GPU accelerated, including all of the top 15. NVIDIA added a record 34 new GPU accelerated system to the latest top 500 supercomputer list, bringing our total to 87 systems. We increased our total petaflops of the list by 28% and we captured 14 of the top 20 spots on the Green 500 list of the world's most energy efficient supercomputers. During the quarter, we continued to support the build out of major next generation supercomputers. Among them is the U. S. Department of Energy's Summit System expected to be the world's most powerful supercomputer when it comes online later this year. We also announced new wins such as Japan's fastest AI supercomputer, the ABCI system, which leverages more than 4,000 Tesla V100 GPUs. Importantly, we are starting to see the convergence of HPC and AI as scientists embrace AI to solve problems faster. Modern supercomputers will need to support multi persistent computation for applying deep learning together with simulation and testing. By combining AI with HPC, supercomputers can deliver increased performance that is orders of magnitudes greater in computations ranging from particle physics to drug discovery to astrophysics. We are also seeing traction for AI in a growing number of vertical industries such as transportation, energy, manufacturing, smart cities and healthcare. We announced engagements with GE Health and Nuance in Medical Imaging, Baker and Hughes, a GE company in Oil and Gas and Japan Komatsu in construction and mining. Moving to professional visualization. 4th quarter revenue grew to a record $254,000,000 up 13% from a year ago, up 6% sequentially driven by demand for real time rendering as well as emerging applications like AI and VR. These emerging applications now represent approximately 30% of pro visualization sales. We saw strength across several key industries, including defense, manufacturing, energy, healthcare and Internet service providers. Among key customers, high end Quadro products are being used by Glaco, Smith, Climb for AI and by Pemex Oil and Gas for seismic processing and visualization. Turning to Automotive. In Automotive for the 4th quarter, revenue grew 3% on year to $132,000,000 and was down 8% sequentially. The sequential decline reflects our transition from infotainment, which is becoming commoditized, to next generation AI cockpit systems and complete top to bottom self driving vehicle platforms built on NVIDIA hardware and software. At CES, demonstrated our leadership position in autonomous vehicles with several key milestones and new partnerships that point to AI self driving cars moving from deployment to production. In a standing room only keynote that drew nearly a 1,000 attendees, Jensen announced that Drive Xavier, the world's first autonomous machine processor, will be available to customers this quarter. With more than 9,000,000,000 transistors, DRIVE Xavier is the most complex system on a chip ever created. We also announced that NVIDIA DRIVE is the world's first functionally safe AI self driving platform, enabling automakers to create autonomous vehicles that can operate safely, a necessary ingredient for going to market. Additionally, we announced a number of collaborations at CES, including with Uber, which is in using NVIDIA technology for the AI computing system in its fleets of self driving cars and freight trucks. We announced that ZF and Baidu are using NVIDIA DRIVE self driving technologies to create a production ready AI autonomous vehicle platform for China, the world's largest automotive market. Production vehicles utilizing this technology, including those from Chery are expected on the road by 2020. We also announced a partnership with Aurora, which is working to create a modular scalable level 4 and level 5 self driving hardware platform incorporating the NVIDIA DRIVE Xavier Processor. Jensen was joined on stage by Volkswagen CEO, Herbert Diess. They announced the new generation of intelligent VW vehicles will use the NVIDIA DRIVE Intelligent Experience or DRIVE IX platform to create the new AI infused cockpit experiences and improve safety. Later at CES, Mercedes Benz announced that Mbox, its new AI based smart cockpit uses NVIDIA's graphics and AI technologies. The MBUX user experience, which includes beautiful touch screen displays and a new voice activated assistant, debuted last week at Mercedes Benz A Class compact car and will ship this spring. And earlier this week, we announced a partnership with Continental to build AI self drive wing vehicle systems from enhanced Level 2 to Level 5 for production in 2021. There are now more than 320 companies and research institutions using the NVIDIA DRIVE platform that's up 50% from a year ago and encompasses virtually every car maker, truck maker, robotaxi company, mapping company, sensor manufacturer and software startup in the autonomous vehicle ecosystem. With its growing momentum, we remain excited about the intermediate to long term opportunities for autonomous driving. Now turning to the rest of the P and L. Q4 GAAP gross margins was 61.9% and non GAAP was 62.1 percent, records that reflect continued growth in our value added platforms. GAAP operating expenses were $728,000,000 and non GAAP operating expenses were $607,000,000 up 28% 22% year on year respectively. We continue to invest in the key platforms driving our long term growth including gaming, AI and Automotive. GAAP EPS was $1.78 up 80% from a year earlier. Some of the upside was driven by lower than expected tax rate as a result of U. S. Tax reform and excess tax benefits related to stock based compensation. Our 4th quarter GAAP effective tax rate was a benefit of 3.7% compared with our expectation of a tax rate of 17.5%. Non GAAP EPS was $1.72 up 52% from a year ago reflecting a quarterly tax rate of 10.5% compared with our expectation of 17.5%. We returned $1,250,000,000 to shareholders in the fiscal year through a combination of quarterly dividends and share repurchases. Our quarterly cash flow from operations reached record levels at $1,360,000,000 bringing our fiscal year total to a record $3,500,000,000 Capital expenditures were $416,000,000 for the 4th quarter, inclusive of $335,000,000 associated with the purchase of our previously financed Santa Clara campus building. Let me take a moment to provide a bit more detail on the impact of U. S. Corporate tax reform on the quarter and our go forward financials. In Q4, we recorded a GAAP only one time net tax benefit of $133,000,000 or $0.21 per diluted share. This is primarily related to provisional tax amounts for the transition tax on accumulated foreign earnings and remeasurement of certain deferred tax assets and liabilities associated with the Tax Cuts and job act. We previously accrued for taxes on a portion of forward earnings in excess of the provisional tax amount recorded for the transition tax, hence the one time benefit. For fiscal 2019, we expect our GAAP and non GAAP tax rates to be around 12%, which is down from approximately 17% previously. This does not take into effect the excess tax benefit from stock based compensation, which depending on stock price investing schedule could increase or decrease our tax rate in GAAP in a given quarter. In terms of our capital allocation priorities, we continue to focus 1st and foremost on investing in our business as we see significant opportunities ahead. Our lower tax rate strengthens our ability to invest in both OpEx such as adding engineering talent, as well as CapEx, such as investing in supercomputers for internal AI development. In addition, we remain committed to returning cash to shareholders with our plan remaining at $1,250,000,000 for fiscal 2019. With that, let me turn to the outlook for the Q1 of fiscal 2019. We expect revenue to be $2,900,000,000 plus or minus 2%. GAAP and non GAAP gross margins are expected to be 62.7% and 63%, respectively, plus or minus 50 basis points. GAAP and non GAAP operating expenses are expected to be approximately $770,000,000 $645,000,000 respectively. GAAP and non GAAP OI and E are both and E are both expected to be nominal. GAAP and non GAAP tax rates are both expected to be 12% plus or minus 1% excluding discrete items. For the full fiscal year 2019, we expect our operating expenses to grow at a similar place as in Q1. Further financial details are included in the CFO commentary and other information available on our IR website. In closing, I'd like to highlight a few upcoming events for the financial community. We'll be presenting at the Goldman Sachs Technology and Internet Conference on February 13 and at the Morgan Stanley Technology Media and Telecom Conference on February 26. We will also be hosting our annual Investor Day on March 27 in San Jose on the sidelines of our annual GPU Technology Conference, which we are very excited about. We will now open the call for questions. Operator, will you hold for questions, please? Your first question comes from the line of C. J. Muse from Evercore. J. Muse:] Yes, good afternoon. Thank you for taking my question. I guess first question, when I think about normal seasonality for gaming that would imply data center potentially north of $700,000,000 plus into the coming quarter. And so curious if I'm thinking about that right or whether crypto is being modeled more conservatively by you guys. And so we'd love to hear your thoughts there. Which way is more conservatively? TJ? Yes, Chuck. When you say conservatively, which direction were you saying? Are you implying up or down? Well, just curious to hear your thoughts there. We model crypto approximately flat. Okay. And then I guess as part of the larger question, how are you thinking about seasonality for gaming into the initial quarter? Well, there's a lot of dynamics going on in gaming. One dynamic, of course, is that there's a fairly sizable pent up demand going into this quarter. But I think the larger dynamics that are happening relate to just the really amazing games that are out right now. PUBG is just is doing incredibly well as you might have known and it's become a global phenomenon. And whether it's here in the United States or in Europe or in China, in Asia, PUBG is just doing incredibly well. And we expect other developers to come up with similar genre like PUBG that are going to be coming out in the near future. And I'm super excited about these games. And then of course, Call of Duty, there's Star Wars, there's just so many great games that are out in the marketplace today. Overwatch and League of Legends still doing well. There's just a countless number of great franchises that are out in the marketplace and the marketplace and the gaming market is growing and production value is going up and that's driving increased unit sales of GPUs as well as ASPs of GPUs. And so I think those are that's probably the larger dynamic of gaming. Your next question comes from the line of Mark Lipacis with Jefferies. Hi, thanks for taking my question. First question, the checks we've done indicate that the tensor cores you put into Volta give it a huge advantage in neural network applications in the data center. And I'm wondering whether the tensor cores might also have a similar kind of utility in the gaming market? Yes. First of all, I appreciate you asking a tensor core question. It is probably the single biggest innovation we had last year in data centers. Our GPUs, the equivalent performance to one of our GPUs, one of our multi GPUs would take something along the lines of 20 plus CPUs or 10 plus nodes. And so one GPU alone would do deep learning so fast that it would take 10 plus CPU powered server nodes to keep up with. And then Tensor Core comes along last year and we increased the throughput of deep learning, increased the power of GPUs. It's very unlike a CPU where the power of GPUs. It's very unlike a CPU where the instruction set remains locked for a long time. And it's hard, it's difficult to advance. In the case of our GPUs and with CUDA, that's one of its fundamental advantages. We can continue to year in and year out continue to add new capabilities to it. And so Tensor Core's boost of the original great performance of our GPU has really raised the bar last year. And as Claude said earlier, our Volta GPU has now been adopted all over the world, whether it's in China with Alibaba, Tencent and Baidu, iFlytek to here in the United States, Amazon and Facebook and Google and Microsoft and IBM and Oracle and in Europe and Japan. The number of cloud service providers that have adopted Volt has been terrific. And I think everybody really appreciates the work that we did with Tensor Core and all of the updates that are now coming out from the frameworks. Tensor Core is a new instruction set, it's a new architecture. And the deep learning developers have really jumped on it. And almost every deep learning framework is being optimized to take advantage of Tensor Core. And on the inference side, on the inference side, and that's where it would play a role in video games, You could use deep learning now to synthesize and to generate new art. And we've been demonstrating some of that at GTC if you've seen some of that, whether it's improved the quality of textures, generating artificial characters, animating characters, whether it's facial animation with for speech or body animation, the type of work that you can do with deep learning for video games is growing. And that's where Core could take a could be a real advantage. If you take a look at the computational capability that we have in Tensor Core, compare that to a non optimized GPU or even a CPU, it's now 2 plus orders of magnitude greater computational throughput. And that allows us to do things like synthesize images in real time, synthesize virtual worlds, animate characters, animate faces, bring a new level of virtual reality and artificial intelligence to these video games. Your next question comes from the line of Vivek Arya with Bank of America. Thank you for taking my question and congratulations on the strong growth and the consistent execution. Jensen, just a near longer term question on the data center. Near term, you had had a number of strong quarters in data center. How is the utilization of these GPUs? And how do you measure whether you're over or under from a supply perspective? And then longer there seems to be a lot of money going into startups developing silicon for deep learning. Is there any advantage they have in taking a clean sheet approach? Or is GPU the most optimal answer? Like if you were starting a new company looking at AI today, would you make another GPU or would you make another ASIC or some other format? Just any color would be helpful. Sure. In the near term, the best way to measure customers that are already using our GPUs for deep learning is rekey customers. When they come back another quarter, another quarter and they continue to buy GPUs, that would suggest that their workload is continuing to increase. With existing customers that already have a very deep penetration, another opportunity for us would be using our GPUs for inference and that's an untapped growth opportunity for our company that's really, really exciting and we're seeing traction there. For companies that are not at the forefront, the absolute forefront of deep learning, which with the exception of 1 or 2 or 3 hyperscalers, almost everybody else I would put in this category and their deployment, their adoption of deep learning, applying deep learning to all of their applications is still ongoing. And so I think the second wave of customers is just showing up. And then there's the 3rd wave of customers, which is, they're not hyperscalers. They, their Internet service applications, Internet applications for consumers, they have enormous customer bases and that they could apply artificial intelligence to, but they run their application in hyperscale clouds. That third phase of growth is it is now really spiking and I'm excited about that. And so that's kind of the way to think about it. There's the pioneers, the first phase, are they returning customers, Then there's the 2nd phase that's now ramping, the 3rd phase that's now ramping. And then for everybody, we have an opportunity to apply our GPUs for inference. If I had all the money in the world and I had, for example, 1,000,000,000 and 1,000,000,000 of dollars of R and D, I would give it to NVIDIA's GPU team, which is exactly what I do. And the reason for that is because the GPU was already inherently the world's best high throughput computational processor. A high throughput processor is a lot more complicated than linear algebra done that you instantiate from a Synopsys tool. It's not quite that easy. The computation throughput, keeping everything moving through your chip with supreme levels of energy efficiency, with all of the software that's needed to keep the data flowing with all of the optimizations that you do with each and every one of the frameworks, the amount of complexity there is just really enormous. The networks are changing all the time. It started out with just basically CNNs and then all kinds of versions of CNNs now. It started out with RNNs and simple RNNs and now there's all kinds of LSTMs and gated RNNs and all kinds of interesting networks that are growing. It started out with just 8 layers and now it's 152 layers going to 1,000 layers. It started with mostly recognition and now it's moving synthesis with GaN. And there's so many versions of GaNs. And so all of these different types of networks are really, really hard to nail down and we're still at the beginning of AI. So the ability for our GPUs to be programmable to all of these different architectures and networks, it's just an enormous advantage. You don't ever have to guess whether NVIDIA GPUs could be used for one particular network or another. And so you could buy our GPUs at will and know that every single GPU that you buy gives you an opportunity to reduce the number of servers in your data center by 22 nodes, by 10 nodes, 22 CPUs. And so the more GPUs you buy, the more value you save. And so I think that capability is really quite unique. And if I could just give you one example from last year or from previous year, we introduced 16 bit mixed precision, we introduced 8 bit integer, we introduced NV Link the year before this last year. This year, this last year, we introduced Tensor Core, which increased it by another factor of nearly 10. Meanwhile, our GPUs get more complex, energy efficiency gets better and better every single year and the software richness gets more amazing. And so it's a much harder problem than just a multiply accumulator. Artificial intelligence is the single most complex mode of software that the world has ever known. That's the reason why it's taken us so long to get here. And these high performance supercomputers is an essential ingredient, an essential instrument in advancing AI. And so I don't think it's nearly as simple as linear algebra. But if I had all the money in the world, I would invest it in the team that we have. Your next question comes from the line of Stacy Rasgon with Bernstein Research. Hi, guys. Thanks for taking my questions. I have a question for Colette. So if I correct for the Switch revenue growth in the quarter, it means the gaming business ex Switch was maybe $140,000,000 $150,000,000 In your Q3 commentary, you did not call out crypto as a driver. You are calling it out in Q4. Is it fair to say that like that incremental growth is all crypto? And I guess going forward, you mentioned pent up demand. Normally your seasonality for gaming would be down probably double digits. Do you think that pent up demand is enough to reverse that normal seasonal pattern, or normally down in Frankfurt? Do you think gamers can even find GPUs at retail at this point to buy in order to satisfy that ton of demand? So let me comment on the first one. We did talk about our overall crypto business last quarter as well. We indicated how much we had in OEM boards and we also indicated that there was definitely some also in our GTX business. Keep in mind that's very difficult for us to quantify down to the end customer it is. But yes, there is also some in our Q3 and we did comment on it. So here we are commenting in terms of what we saw in terms of Q4. It's up a bit from what we saw in Q3. And we do again expect probably going forward. I'll let Jensen answer regarding the demand for gamers as we move forward. Yes. So one way to think about the pent up demand is we typically have somewhere between 6 to 8 weeks of inventory in the channel. And I think you would ascertain that globally right now the channel is relatively lean. We're working really hard to get GPUs out into the marketplace for the gamers and we're doing everything we can to advise e tailers and system builders to serve the gamers. And so we're doing everything we can, but I think the most important thing is we just got to catch up with supply. Your next question comes from the line of Mitch Steves with RBC. Hey guys, thanks for taking my question. I actually want to circle back on the auto since I was at CES. So it's still kind of on track for calendar towards calendar year 2019 at the end of that where we see the autonomous kind of ASP uplift. And just to clarify, the expected ASP uplift is somewhere around $1,000 Is that about right? Yes, it just depends on mix. I think for autonomous vehicles that still have drivers, passenger cars, branded cars, ASPs anywhere from $500,000 to $100,000 makes sense. For robot taxis, where they're driverless, they're not autonomous vehicles, they're actually driverless vehicles, the ASP will be several $1,000 And in terms of timing, I think that you're going to see larger and larger deployments starting this year and going through next year for sure, especially with robot taxis. And then with autonomous vehicles, cars that have autonomous driving capability, automatic driving capability starts late 2019, you could see a lot more in 2020. And just almost every premium car by 2022 will have autonomous automatic driving capabilities. Your next question comes from the line of Toshiya Hari with Goldman Sachs. Great. Thanks very much for taking the question. Jensen, I was hoping to ask a little bit about inferencing. How big was inferencing within data center in Q4 or fiscal 2018? And more importantly, how do you expect that to trend over the next 12 months to 18 months? Thank you. Yes. Thanks a lot, Tush. First of all, just a comment about inference. The way that it works is you take the output of these frameworks and the output of these frameworks is a really complex, large computational graph. When you think about these neural networks and they have millions of parameters, millions of parameters, millions of anything is very complex. And these parameters are weights and activation layers and activation functions and there are millions of them. And it's millions of them that composes consists of this computational graph. And this computational graph has all kinds of interesting and complicated layers. And so you take this computational graph that comes out of each one of these frameworks and they're all different. They're in different formats, they're in different styles, they're in different architectures, they're all different. And you take these computational graphs and you have to find a way to compile it, to optimize this graph, to rationalize all those things that you could combine and fold, reduce the amount of conflict across all of the resources that are in your GPUs or your processor. And these conflicts could be on chip memory and register files and data paths and it could be the fabric, it could be the frame buffer interface, it could be the amount of memory. I mean, you got this computer is really complicated across all these different processors. And the interconnect between GPUs, the network that connects multiple nodes. And so you've got to figure out what all these different conflicts are, resources are and compile and optimize to take advantage of it to keep it moving all the time. And so TensorRT is basically a very sophisticated optimizing graph compilation graph compiler. And it targets each one of our processors. The way it targets Xavier is different than the way it targets Volta, the way it targets our inference, the way it targets for low energy, for different precisions, that all of that targeting is different. And so first of all, TensorRT, the software of inference, that's really where the magic is. Then the second thing that we do, we optimize our GPUs for extremely high throughput and to support different precisions because some networks could afford to have 8 bit integer or even less. Some really can barely get by with 16 bit floating point and some you really would like to keep it at 32 bit floating point so that you don't have to second guess about any precision that you lost along the way. And so we created an architecture that consists of this optimizing graph computational graph compiler to processors that are very high throughput that are mixed precision, Okay, so that's kind of the background. We've been sampling our Tesla P4, which is our data center inference processor. And we're seeing just really exciting response. And this quarter, we started shipping. Looking outwards, my sense is that the inference market is probably about as large in the data centers as training. And the wonderful thing is everything that you train on our processor will inference wonderfully on our processors as well. And the data centers are really awakening to the observation that the more GPUs they buy for offloading inference and training, the more money they save. And the amount of money they save is not 20% or 50%, it's factors of 10%. The money savings for all of these data centers that are becoming increasingly capital constrained is really quite dramatic. And then the other inference opportunity for us is autonomous machines, which is self driving cars. TensorRT also targets Xavier. TensorRT targets our Pegasus robot taxi computer, and they all have to inference incredibly efficiently so that we can sustain real time, keep the energy level low and keep the cost low for car companies. Okay. So I think inference is a very important work for us. It is very complicated work and we're making great progress. Your next question comes from the line of Blayne Curtis with Barclays. Hey, guys. Thanks for taking my question. Just kind of curious, as you look at the gaming business, I've kind of lost track of what seasonality is. You clearly have a big ramp ahead of you. I'm just kind of curious as you think about Pascal versus seasonality ahead of Volta, if you can just kind of extrapolate as you look out into April and maybe July? Well, we don't we haven't announced anything for the for April, July. And so the best way to think about that is Pascal is the best gaming platform on the planet. It is the most feature rich, the best software, the most energy efficient and from $99 to $1,000 you can buy the world's best GPUs, the most advanced GPUs. If you buy Pascal, you know you get the best. Seasonality is a good question. And increasingly, because gaming is a global market and because people play games every day, it's just part of their life. There's no I don't think there's much seasonality in TV or books or music. People just whenever new titles comes out, that's when a new season starts. And so in China, there's iCafes and there's Singles' Day, November 11. There's back to school in the United States. There's Christmas. There's Chinese New Year. Boy, there's so many seasons that it's kind of hard to imagine what the exact seasonality is anymore. And so hopefully over time, it becomes less of a matter. But the most important thing is that we expect Pascal to continue to be the world's best gaming platform for the foreseeable future. Your next question comes from the line of Harlan Sur with JPMorgan. Good afternoon and congratulations on the solid results and the execution. I know somebody asked a question about inferencing for the data center markets, but on inferencing for embedded and edge applications, on the software and firmware side, you talked about printer RT framework. On the hardware side, you've got the Jetson TX platform for embedded and edge inferencing applications, things like drones and factory automation and transportation. What else is the team doing in the embedded markets to capture more of the TAM opportunities there going forward? Yes. Thanks a lot, Martin. The NVIDIA TensorRT is really the only optimizing inference compiler in the world today. And it targets all of our platforms. And we do inference in the data center that I mentioned earlier. In the embedded world, the first embedded platform we're targeting is self driving cars. In order to drive the car, you're basically inference or trying to predict or perceive what's around you all the time. And that's a very complicated inference matter. It could be extremely easy like detecting the car in front of you and applying the brakes. Our work could be incredibly hard, which is trying to figure out whether you should stop at an intersection or not. If you look at most intersections, you can't just look at the lights to determine and where do you stop. There are very few lines. And so using scene understanding and using deep learning, we have the ability to recognize where to stop and whether to stop. And then for Jetson, we have a platform called Metropolis, and Metropolis is used for very large scale smart cities, where cameras are deployed all over to keep cities safe. We've been very successful with smart cities, just about every major smart city provider and what is called intelligent video analysis company, whether almost all over the world is using NVIDIA's platform to do inference at the edge, AI at the edge. And then we've announced recently success with Fanuc, the largest manufacturing robotics company in the world Komatsu, one of the largest constructions equipment company in the world to apply AI at the edge for autonomous machines. Drones, we have several industrial drones that are inspecting pipelines, inspecting power lines, flying over large spans of farms to figure out where to spray insecticides more accurately. There's all kinds of applications. So you're absolutely right that inference at the edge or AI at the edge is a very large market opportunity for us and that's exactly why TensorRT was created. Your next question comes from the line of Joe Moore with Morgan Stanley. Great. Thank you. You had mentioned how lean the channel is in terms of gaming cards. There's been an observable increase in prices at retail. And I'm just curious, is that a broad based phenomenon? And is there any economic ramifications to you? Or is that just sort of retailers bringing prices up in a shortage environment? Thank you. We don't set prices at the end of the market. And the best way for us to solve this problem is work on demand excuse me, work on supply. The demand is great, and it's very likely that demand will remain great as we look throughout this through this quarter. And so we just have to keep working on increasing supply. We have our suppliers are the world's best and the largest semiconductor manufacturers in the world And they're responding incredibly, and I'm really grateful for everything they're doing. We just got to catch up to that demand, which is just really great. Your next question comes from the line of Chris Rolland with Susquehanna. Hey, guys. Thanks for the question and great quarter. So just to clarify, Jensen, on pent up demand, one of your GPU competitors basically said that the constraint was memory. I just want to make sure that that was correct. And then in the CFO commentary, you mentioned opportunities for professional biz, like AI and deep learning. Can you talk about that and what kind of applications you would use, Quadro versus Volta or GeForce? Thanks. Sure. We are just constrained. Obviously, we're 10x larger of a GPU supplier than the competition. And so we have a lot more suppliers supporting us and a lot more distributors taking our price to market and a lot more partners distributing our products all over the world. And so we I don't know how to explain it aside from the demand is just really great. And so we just got to keep our nose to it and catch up to the demand. With respect to Quadro, Quadro is a workstation processor. The entire software stack is designed for all of the applications that the workstation industry uses. And it's used it's the quality of the rendering is, of course, world class because of NVIDIA. And but the entire software stack has been designed so that mission critical applications or long life industrial applications and companies that are enormous and gigantic manufacturing and industrial companies in the world could rely on an entire platform, which consists of processors and system and software and middleware and all the integrations into all of the CAD tools in the world To know that the supplier is going to be here and can be trusted for the entire life of the use of that product, which could be several years, but the data that is generated from it has to be accountable for a couple of decades. You need to be able to pull up an entire design of a plane or a train or a car a couple of decades after it was sent in production to make sure that it's still in compliant and if there are any questions about it that it can be pulled up. NVIDIA's entire platform was designed to be professional class, professional grade, low life. Now the thing that's really exciting about artificial intelligence is we now can use AI to improve images. Like for example, you could fix a photograph using AI. You could fill in damaged parts of a photograph or parts of the image that hasn't been rendered yet, you want to use AI to fill in the dots, predict the future rendering results, which we announced and which we demonstrated at GTC recently. You can use it to generate designs. You sketch out a few strokes of what you want a car to look like. And based on the inventory, safety, physics, it could it has learned how to fill in the rest of it, okay, design the rest of the chassis on your behalf. It's called generative design. We're going to see generative design in product design and building design and just buy everything. The last, if you will, 90% of the work is after the initial inspiration or the conceptual design is done. That part of it can be highly automated through AI. And so Quadro could be used as a platform that designs as well as generatively designs. And then lastly, a lot of people are using our workstations to also train their neural networks for these generative designs. And so you could train and develop neural networks and then apply it in the applications. Okay. So AI think of AI really as in the final analysis the future way of developing software. And so you And so you could teach you could use data to teach a software to figure out how to write the software by itself. And then when you're done developing that software, you can use it to do all kinds of stuff, including design products. And so for workstations, that's how it's used. Your next question comes from the line of Craig Ellis with B. Riley. Thank you for taking the question and congratulations on the very good quarterly execution. A lot of near term items here on gaming, so I'll switch it to longer term. Jensen at CES, I think you said that there are now 200,000,000 GeForce users globally. And if my math is correct, then that would be up about 2x over the last 3 to 4 years. So the question is, is there anything that you can see that would preclude that kind of growth over a similar period? And given the recent demand dynamics, I think we've seen that NVIDIA's direct channels have been very good sources for GPUs at the prices that you intend. So as we look ahead, should we expect any change in channel management from the company? Thank you. Yes. Thanks a lot, Craig. In the last several years, several dynamics happened at the same time and all of it were the favorable contributions to today. First of all, gaming became a global market and China became one of the largest gaming markets in the world. The second, because the market became so big, developers could invest extraordinary amounts into the production value with the new game. They could invest a few $100,000,000 and know that they're going to get the return on. Back when the video game industry was quite small or when PC industry PC gaming was small, it was too risky for a developer to invest that much. And so now an investor, a developer could invest 100 of 1,000,000 of dollars and create something that is just completely photorealistic and immersive and just beautiful. And so the production when a production value goes up, the GPU technology that's needed to run it well goes up. It's very different than music. It's very different than watching movies. Everything in video games is synthesized in real time. So when the production value goes up, the ASP or the technology has to go up. And then lastly, the size of the market, people have wondered how big the video game market is going to be. And I've always believed that the video game market is going to be literally everyone. In 10 years' time, 15 years' time, there's going to be another 1,000,000,000 people on earth and those people are going to be gamers. We're going to see more and more gamers. And not to mention that, almost every single sport could be a virtual reality sport. So video games is every sport. So e sport can be any sport in every sport and every type of sport. And so I think when you consider this and put that in your mind, I think the opportunity for video games is going to be quite large. And that's essentially what we're seeing. Your next question comes from the line of William Stein with SunTrust. Great. Thanks for taking my question and congrats on the great results and even better outlook. I'm hoping we can touch on automotive a little bit more. In particular, I think in the past you've talked about expecting sort of a lull in revenue growth in this market until roughly the 2020 timeframe when autonomous driving kicks in, in a more meaningful way. But of course, you have the AI co pilot that seems to be potentially ramping sooner and you have at least one marquee customer that is ramping now, I guess, but volumes aren't quite that large on the autonomous driving side. So any guidance as to when we might see these two factors start to accelerate revenue in that end market? Thanks. Yes. Thanks a lot, Will. I wish I had more precision for you, but here are some of the dynamics that I believe in. I believe that autonomous capabilities, autonomous driving is the single greatest dynamic next to EVs in the automotive industry. And transportation is a $10,000,000,000,000 industry. Between cars and shuttles and buses, delivery vehicles. I mean, it's just an extraordinary, extraordinary market. And everything that's going to move in the future will be autonomous. That's for sure. And it will be autonomous fully or it will be autonomous partly. The size of this marketplace is quite large. In the near term, our path to that future, which I believe starts in 2020, 2019, 2020, but starts very strongly in 2022. I believe the path to that in our case has several elements. The first element is that in order for all these companies, whether they're Tier 1s or startups or OEMs or taxi companies or ride hailing companies or tractor companies or shuttle companies or pizza delivery shuttles. In order to deliver in order to create their autonomous driving capability, the first thing you have to do is train a neural network. And we created a platform we call the NVIDIA DGX that allows everybody to train their neural networks as quickly as possible. So that's first, it's the development of the AI requires GPUs and we benefit first from that. The second is, which will start this year and next year, is development platforms for the cars themselves, for the vehicles themselves. And finally, Xavier is here. We have a first silicon of Xavier, it's the most complex SoC the world's ever made. And we're super excited about the state of Xavier and we're going to be sampling it in Q1. And so now we'll be able to help everybody create development systems and it will be 1,000 and tens of thousands of quite expensive development systems based on Xavier and based on Pegasus that the world is going to need. And so that's the second element. The third element in the near term will be development agreements. Each one of these projects are engineering intensive and there's a development agreement that goes along with it. And so these three elements, these three components are in the near term and then hopefully starting from 2019 going forward and very strongly going from 2022 and beyond the actual car revenues and economics will show up. I appreciate that question. And I think this is our last question. Yes? Well, we had a record quarter, wrapping up a record year. We had a strong momentum in our gaming, AI, data center and self driving car businesses. It's great to see adoption of NVIDIA's GPU computing platform increasing in so many industries. We accomplished a great deal this last year and we have big plans for this coming year. Next month, the brightest minds in AI and scientific world will come together at our GPU Technology Conference in San Jose. GTC has grown tenfold in the last 5 years. This year, we expect more than 8,000 attendees. GTC is the place to be if you are an AI researcher or doing any field of science where computing is your essential instrument. There will be over 500 hours of talks of recent breakthroughs and discoveries by leaders in the field such as Google, Amazon, Facebook, Microsoft and many others. Developers from industries ranging from healthcare to transportation to manufacturing and entertainment will come together and share state of the art in AI. This is going to be a great GTC. I hope to see all of you there.