NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

So what we have right now is, given the ramp of overall A100 and as we discussed, starting first with the overall hyperscales and creating those cloud instances that many of our enterprises work, that's a good bridge to begin a quick adoption of the overall enterprises as well as we move forward. Right now and it changes quarter-to-quarter, what we're seeing is the hyperscales are just over 50% of our overall data center revenue. And what we're seeing is the vertical industries are just over 40%, and supercomputing can be less than overall 5%. So those are important to understand what the mix is and the importance of each of these. And some of our vertical industries are doing quite well and growing quite well, and some of them will probably take some time for them to move to the new architecture given a lot of the challenges that we have with the overall economy.

Okay. And on the Mellanox side, Mellanox exceeding expectations so far, what keeps you excited?

So we're thrilled that Mellanox is now with us here at NVIDIA. The Mellanox team, I would look at as a perfect match within NVIDIA to join our strategy and our focus in terms of the future of the data center. Our years of partnership, I think, is really key in terms of how that allowed our first quarter at NVIDIA to work so smoothly. We are now in the process of beginning our work on the future of the modern data center rather than just getting to know each other. So we're off to a great start. We couldn't be more pleased in terms of the work that they have here in front of us. That modern data center is really about accelerated using the GPU-enabled capabilities, disaggregated using Mellanox-enabled capabilities as well as overall software-defined to really take all the pieces in terms of the data center and accelerate each one of them piece by piece and not just a focus on the time that is overall computing or the GPU is enabling that accelerated computing. So we're excited to see all the great things working together with Mellanox, but you're correct. It's off to a great start, and we're excited to have them on board.

Okay. Colette, this is a tough one. But again, investors kind of always struggle with how to model the growth of the data center business. The team had talked about a $50 billion TAM by 2023 at its Investor Day last year, split into a 40, 40, 20, hyperscale, enterprise and scientific computing segments. Jensen has talked about cloud being $100 billion market growing at 40% compounded growth rate. How do you see the long-term data center market growth for NVIDIA?

If we step back and look at where we are in the overall ramp of overall AI across the world, AI is still nascent, and it's still advancing exponentially each and every month or essentially each and every year. The model sizes we've talked about, even over 3 years, have increased 3,000x on the training side before -- and since we've launched overall V100. It has allowed us to open up to new markets as we continue to fuel giant new models that can be trained, focused on natural language processing, probably one of the most difficult pieces of work to understand voice and translate that voice and return back in normal conversation. Additionally, we have proven the ability to use a programmable GPU, its overall software capabilities to take on a very important market such as inferencing. Inferencing is a huge market that will take place, not only in the data center, but also on the edge and also in terms of autonomous types of edge devices as we move forward. So we're still in those early stages of that. With the inclusion of Mellanox, we now have that ability to do more than just accelerating overall compute, but we have an ability to accelerate almost all of the data center with now a focus of the overall networking. We've likely added tens of billions to our TAM with the inclusion of acceleration and opportunities that we now have with overall Mellanox. So in the early stages, off to a great start. But yes, there is a large TAM and a large opportunity for our data center business as a whole going forward.

Okay. And this is a tricky one, too. Jensen mentioned on the last earnings call that NVIDIA works closely with ARM, and he respects the energy efficiency of the ARM architecture. Just conceptually, are there advantages or disadvantages, if any, on combining a CPU with a GPU in future compute architectures?

So first, I'll start off. Yes, we work with all CPUs. We are agnostic to each, any of the overall CPUs that are available. For example, within our DGXs and our new SuperPODs include the AMD CPU as well. We also work closely with ARM. With each of the workloads, it is a choice of the best overall CPU architecture and is considered when we put some of those workloads together. Some geographies don't want to be dependent on the x86 overall processors. And so in this case, ARM is the most popular instruction set in the world and essentially fuels billions of devices that are out there. ARM is also very power efficient and is considered for something in the data center to improve that overall data center TCO. ARM is the only CPU right now without accelerated compute support. So NVIDIA had started that work in terms of working with ARM to bring CUDA to their instruction set. We work with a variety of software -- excuse me, a variety of hardware, software companies in the ARM ecosystem. And we believe that CUDA and GPU acceleration for ARM-based servers, such as hyperscales, edge computing and HPC applications, can be with us in the overall future. So that's how we think about all the different future compute architectures, but our focus is making sure that we are agnostic to all the CPUs that are out there or essentially any of the players that are in the data center in terms of assuring that NVIDIA's overall GPU workloads can work effectively and well to accelerate everything in the data center.

Okay. Then switching to auto. Ampere architecture powers next-generation auto platform called Orin, delivering more than 6x the performance of Xavier and 4x better power efficiency. How does the improved power management profile on Ampere compare with competition? Will this significantly help to close the gap on Level 2 and up?

Yes. So we're very excited in the last 6 months what we have brought to market for many of the auto manufacturers, the Tier 1s and many of the overall start-ups. As you know, we have been focused on an end-to-end platform, enable the cars to take what they have in terms of production of AV from a Level 2, a Level 2+, all the way up to a fully autonomous, and also taking care of what we may want to do in terms of overall robotaxis. Now an important piece of that is making sure that they have that end-to-end system that enables the ability to advance to Level 3, Level 4, even with the overall same hardware. The energy efficiency is important. But generally, all of our overall GPUs and all of our overall platforms continue to be the best in terms of performance and energy efficiency together in terms of what we're bringing into market. This last quarter was an important quarter of us to announce one of the most transformational overall deals in the automotive industry. Our work in terms of with Mercedes is very key and a very important thing that a lot of eyes are on. What it did was take a company, such as Mercedes, and such a rock-solid company that's been available for more than 100 years, and working with us to develop the full software stack that will go into their AV solution. Using our hardware, that the hardware at the onset can continue to upgrade with software and enable many different levels of overall AV over the life of the ownership of the overall car. The great piece of that, which is so great for us as well, is the ability to share that revenue with Mellanox, share that software revenue back to overall NVIDIA. We're not only working inside of the car with our Orin pieces and many other parts of our overall system, but keep in mind that we are working with Mercedes back in their data center to both support leading up to overall production as well as the fleet that will be out there on the road and the data on continuing to work with that data over the time. So we're really thoughtful in terms of thinking through a supermodel that can be leveraged across the overall industry, but our work with Mercedes is definitely transformational.

Great. Then moving to the financial model. Gross margins in second half because of the gaming mix are coming down. Do you see it trending back up to 66% as data center mix normalizes, let's say, in first half of next year?

Yes. So in our Q2 results, our Q2 results last quarter, we hit record levels at 66% overall non-GAAP gross margin. Remember, the mix of our products still remains our largest driver for our overall gross margins. Last quarter, with the inclusion of Mellanox, our data center revenue reached record levels and helped drive our overall gross margin expansion. With our Q3 guidance for non-GAAP gross margins, they're just slightly below that record level in Q2. New products that we are introducing have slightly lower margins earlier in the ramp, but we expect those margins to tend to increase as both yields mature and as we optimize a lot of those overall component costs. So when you think about that overall Ampere architecture as a whole, keep in mind, we have a very strong mix expected of gaming as we move into Q3 from Q2. Within that gaming mix, we also expect all of our platforms to grow, including our consoles. Our consoles do not have as high a gross margin as the company average. So in total, we expect to continue to focus on our overall gross margins and focus in terms of launching this Ampere architecture across both data center and gaming. And we expect over the long term to have an ability to move them up over time.

Okay. And then on the OpEx side, hiring and retaining an AI developer, software talent in Silicon Valley is a tough job because you're competing against Internet and software companies. How do you incentivize your employees? And how should we think about the OpEx growth relative to the sales next year?

Yes. So first, talking about our employees and our engineering workforce, both from a software side and a hardware side. We do have more overall software developers than we actually have hardware developers. The great thing is our software development is done in many places around the world and now is actually done at home as we still have almost all of our sites closed as we are all working from home. There's no place on the planet that you can go and work like an NVIDIA place. NVIDIA has the ability to produce software for so many different overall workloads and some of the most important workloads of the future. When we think about overall ray tracing for gaming, when you think about everything from virtual reality or augmented reality, enterprise types of graphics, even before we think about the workloads that we are doing in terms of in the data center and the focus on AI and autonomous vehicles. There are some great pieces that we can entice a lot of that talent, but I think what is more important is the talent enjoys NVIDIA. We have probably one of the best retention that we do in terms of our talent here. People come to work in NVIDIA and do their life's work, their life's work in terms of focus on that GPU overall architecture and thinking about the use of acceleration in so many of these different platforms going forward. So right now, we feel very good in terms of both our ability to retain and attract talent to overall NVIDIA. As we turn the corner into next year, as we've just talked about, we've got many very important workloads, very big markets in front of us that will continue to need our focus and continue to see our investment. So we don't necessarily look at OpEx as a percentage of sales. We look at it that says, where is the best use of overall OpEx to assure that we can meet these large opportunities in front of us. So we'll provide more once we turn the corner at the end of this year to focus on what we see as OpEx going forward into fiscal year '22.

Okay. Colette, TSMC is building a fab in U.S. next year. We hear U.S. semiconductor industry lobbying for $37 billion, with $17 billion in research and a chip factory. There are geopolitical tensions with China. What does it all mean for NVIDIA?

So the idea of bringing a high-end fab here into the U.S., we couldn't be more excited about it. We work with 2 of the best overall fab providers on the planet today. Having one of those closer into the U.S. would be great for us. We deal right now with a worldwide workforce working with those fabs, but having something closer to home here in Santa Clara would also be great for us as well to think about it. So we couldn't be more supportive of something in that space. We are working also with the semiconductor industry because I think we need to work together as a group. We are just one piece of this, and we do support the needs of that overall semiconductor industry.

Okay. We have about 3 to 4 minutes left. I'm going to switch to audience questions. Colette, could you comment further on opportunities for data center for new verticals? And in particular, for enterprises, could you make some examples? What is missing for penetration to increase further adoption in enterprises?

Yes. So we've had a great rollout in the last 2 years focused on building out many of the workloads for the enterprises. What has enabled us to be quite successful in this piece is the focus in terms of the software that will need to be available and the overall production of development resources to enable these workloads as we move forward. Now we have focused on very key areas. One, we focused on manufacturing. We focused on metropolises and building out the overall smart cities. You've seen us focus in terms of on Clara. So when you think about our data center platforms, it's not just focused on that hardware. We have built so much of the software libraries, components, compilers and working with many of the different frameworks that enable it. We have enabled in terms with our Tensor cores for depending on the workload and depending on that precision in order to address the overall workloads that are out there. Our focus in terms of e-tail and retailers have been very key as well. You've seen us focus with large lighthouse retail, such as Walmart, focused on very large distribution, such as the U.S. Postal Service. So each and every one of these workloads continues to drive overall acceleration into the enterprises. We also know the enterprises are focused on cloud computing as well. So we hit in both areas, both in terms of edge computing that we can provide to them, but also what we can provide in those cloud instances. Let's step back and look at what A100 provides. A100 provides a solution for you to not have to choose or determine at the onset whether or not you need to do training or whether or not you need a virtual VM, okay? You can do both. You can do inferencing and training on the same platform. This is great not only for the hyperscales as they built out their data centers, but it is also a very useful technique for the enterprises as they think about the reuse and redistributing that overall compute for the latest types of workloads that we're doing. So our work is essentially workload by workload or industry-specific and really stems with the application and the software. One more piece that is important is our work in terms of the data scientists that are out there or the data analytics. This is a very important piece of almost every enterprise, and it's not necessarily industry-specific. The enterprises are out there spending time working with their data scientists to understand how to leverage their data to make overall decisions. It's a relatively inefficient process. They spend a lot of time loading data. They spend a lot of time reanalyzing data, rerunning overall queries to get the information that they were looking for. We have now been able to stitch ourselves together with Spark, Spark 3.0, which is a very key component in terms of the work that they do. Accelerating that entire process, from end to end in terms of the data load, the working with the data and feeding back out that data, can improve something that is integral to all of the overall enterprises that are out there. So more to come, but this is what all of our software engineers continue to do every day is more and more workloads that address the industries and the enterprises around the world.

Great. We're almost at the end of time. Colette, thank you for being part of the Citi conference, really appreciate it.

Okay. Thank you so much for hosting us, and we'll see you next year as well.