NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Right. Good morning, everybody. Thank you for joining us at our Healthcare Conference. I'm Josh Buchalter, semiconductor analyst at TD Cowen. I'm really pleased to be joined by my colleague, Brendan Smith, who covers life science, tools and diagnostics and biotechnology to host a discussion with Kimberly Powell, VP of Healthcare at NVIDIA. Brendan, thanks for inviting me to your party. And Kimberly, thank you for joining us.

Thank you for having me.

I guess to start, maybe get the elephant out of the room. What's a semiconductor company doing at a healthcare conference. From a high level, how does NVIDIA support the health care market and how does it fit into the broader NVIDIA story?

Yes. So let me take 1 minute to as a reminder, this presentation contains forward-looking statements, and investors are advised to read our reports filed with the SEC for information related to risks and uncertainties facing the business. But I appreciate the question. Let's take a step back for about 2 decades now, NVIDIA has dedicated themselves to what's called accelerated computing. Accelerated computing was originally adopted by the medical community, some of the earliest applications, in fact, came from the medical community, medical imaging, molecular dynamics simulation, genomics analysis. So what have we done here? NVIDIA has -- we build computers that solve problems no other computers can. And these are -- led us to some of the most important breakthroughs that we're living through right now. One of those breakthroughs, of course, is artificial intelligence. But many other areas of science are also having incredible breakthroughs, whether that's climate science, now drug discovery, and you're, of course, hearing a lot more of the increase in what's going to transpire in physical AI and robotics. And so NVIDIA is not a semiconductor company any longer. We build what's called full stack data center solutions where there is a GPU of course, graphics processing unit, but it's accompanied with CPUs that we also design, networking, full data center scale technologies that really allow for us to make computing go through what is unimaginable in the previous generations of computing. For example, 1 million times of compute performance has been achieved in the last 10 years. This is orders and orders of magnitude, more computational capabilities than the previous Moore's Law of computing had brought us. And so here we are, we have invented our latest architecture, which is called Blackwell. Blackwell has been built for the modern AI of agents that we are now living in. There are new laws of scaling that you're all, I'm sure, starting to learn quickly about. The first law of scaling is now what's called pre-training, pretraining the AMLs. This is like sending your child to grade school all the way up through high school. It's the foundation at which they build the ability to learn, memorizing facts about history or other kind of fields of science. So pretraining requires a lot of data, a lot of infrastructure to create a model. There's a second law, which is called post training. This is a bit like getting a post doc. This is where you're now fine-tuning the skills with more domain-specific information or task-specific information so that you can have these AI models do jobs with utmost accuracy. And then there's a third scaling law, which we're living through, which is called test time compute. And this is actually where AI goes off and it thinks. You've all heard about what's called reasoning models. And so we've built the Blackwell architecture to advance these incredible 3 new scaling laws and really unlock the world that we're living in now, which is agenetic AI. And then I'll talk to you a little bit more about the very near future here that we're inventing, which is physical AI. And so all that we're building here in terms of the computing platform has huge applicability into health care. And I know we're going to talk more about those applications.

For sure. Thank you for that Kimberly. I definitely want to get back to the scaling laws. But maybe to start, I think you've described health care as potentially the largest opportunity within your AI ecosystem. And I think you described 3 main buckets, digital surgery, digital biology and digital health. Could you maybe speak to the maturity of those 3 verticals where they are today and the potential size of them long term within NVIDIA's model? And I'll turn it over to Brendan.

Yes, absolutely. So we all know that health care is about a $10 trillion industry globally. And in fact, of that $10 trillion, upwards of 30% of that is in the form of labor, labs and infrastructure, physical infrastructure in order to deliver health care. I think we can all recognize that even here where we live in a very well-endowed country, we're still at a drastic supply and demand shortage, demand of health care services and the supply of health care services. We're at an extreme shortage, tens of millions of human short which is equal to trillions of dollars, each 10 million people in the health care services is approximately $1 trillion. And so AI has to help. It's a bit existential in that way. AI has to help. And so we think about it in really 3 major buckets, as you described. One area, and this is the area that we've been in the longest is really in the area of digital devices, right? If you think about the patient journey, oftentimes the very first thing you're doing is you're receiving some sort of imaging to understand what's going on in the body, so they saw your diagnostic imaging, you're measuring the biology, whether that's lab test or genomic sequencing test. And then you're going into some form of therapy that could be radiotherapy or could be robotic-assisted. All of these sensors that these companies, great leaders in the industry, Siemens, GE, Intuitive Surgical, all of these are fantastic platforms and sensors that are going to become robotic in their own right. They are not only going to sense what's going on, but they're going to reason over it and they're going to facilitate the ability for a deeper understanding of our physicians. And so digital devices is a huge area where it's going to continue to take advantage of these amazing agentic AI systems and future robotic systems. In digital biology, this is in the area of drug discovery, right? We just actually added another $50 billion. Albert was here, $50 billion to a total of $300 billion is spent in R&D for drug discovery every single year. And we know we have extreme challenges here. We still have 90% attrition rate when we get into clinical. We are generating more data than ever before. This is where we can take advantage of the AI breakthroughs that happened in language and apply it to the languages of drugs. So just like our language, our spoken language is a sequence actually biology is a sequence. It's just in a very different shape. It's 3 billion base pairs long is your DNA, amino acid sequence for proteins and even smile strings for chemistry, but we can now start to use that to our advantage. All of the data we can now generate approximately 24 hours a day in these labs, we can start injecting those into these large language models, biology foundation models to represent biology in profound new ways. And we're seeing the R&D productivity massively increase. There's many examples. Schrodinger disrupting themselves in a way, using their own platform in fantastic ways to take what is otherwise years of work into 10 months. And instead of synthesizing hundreds, if not thousands of chemicals, they're doing it on 1 or 2 orders of magnitude less and so that R&D productivity is really profound. And then it doesn't just stop there in early discovery, it continues on into clinical development. We have friends here in the room who are acutely trying to tackle this problem of how can we do a better job using the language models to now strategize our clinical trial processes much better. Find the patients, find the sites, design the trials, keep the patient engagement high, all of that now can be facilitated by today's agentic AI. So there's going to be amazing progress that is going to be felt there. And then the third one being digital health. Think of these as digital humans that you can hire on behalf of this otherwise completely stretched and burning out health care force that we have. We need to augment them with a digital health force. And we now have the ability, Hippocratic AI, one of our fantastic partners just launched at the January Health Conference a whole agent AI marketplace where instead of the software systems in the past, the electronic health records, which were huge software deployments, billions of dollars, huge change cost, huge training and change management necessary in the health care system. We're now in the phase of just like you and I as consumers, we can order groceries or use a service like Spotify, on-demand services, you can now augment the health care system and start to navigate it through on-demand digital human digital health services. And so this is really, really fantastic. And this is happening right now as we speak. And this is not a -- we're going to see this in 3 or 5 years. This is what we're going to see happen in the next 3 months. So super excited about digital health. So those are really the 3 verticals that I think truly address a gigantic opportunity. Think about that $3 trillion of operating expenses and labor and how we're going to augment them. That can easily be a couple of hundred billion dollars in what we call AI factory, AI infrastructure to support all of those agents, supporting the robotic surgeries of the future, supporting the digital biology labs of the future and supporting all of our health care workers. So it's -- this a gigantic opportunity.

Yes, I think that's a really super helpful lay of the different use cases for AI within health care that you've just outlined. So I want to double-click maybe on the health and biology aspects. So maybe which types of research, diagnostic or clinical applications at this point, do you see as particularly ripe for AI-driven innovation? And I guess, what is it about those applications that might lead you to say that?

Yes. I think that actually, just recently, Roche announced a new platform. When I was here last year, we had lots of new sequencing and genomics platforms. There's still massive innovation in the ability to be able to read biology at incredibly new resolutions and in same throughput. And genomics is just a very obvious one because humans are just not built to read it or understand it. But what has just happened, in fact, I think it's only a week old. We did some really groundbreaking work with the Arc Institute in California. We helped build the world's largest biology foundation model for the first time. This biology foundation model is incredible. It was trained on some 9 trillion nucleotides. It allows for a context length of 1 million which is, in some cases, a whole genome of certain kind of species bacteria. These are the kind of domain specific things that we need to work on in order to take the breakthroughs of large language models and transition them into the areas of biology. And so what is really happening here when you can study genomes from all these different species, you create a representation of biology that now allows for a study of the biology and understanding of the biology in incredible new ways. And in fact, can predict variance at extreme accuracy like they proved with the BRCA 1 genomic variant, but this model can also do some other incredibly exciting things like right biology. So you went from, okay, we're still creating new platforms that can read biology at new scales. We now have the ability to write biology, and we also need these models to help us really do the analysis and understand the inner workings of biology. And so there's a lot of people in this room I know are thinking about it. This idea is how can we eventually get to what's called the virtual cell, the AI virtual cell where you're able to now maybe represent biology at multi-levels and then be able to have these AI models be able to reason across these multi scales and connect what is a very complicated. The world's probably most complicated study of biology to not only understand all of the molecules that's going inside cell but also cell to cell, cellular systems all the way up to organ. And so we have a lot of work to do there. So it's absolutely right. And it's one of our fastest-growing areas. In fact, is helping the pharmaceutical industry, the -- this burgeoning sort of tech bio industry, which is coming at it from what data can I generate, what can I measure and then how can I represent models to then drive what kind of medicines we can go after. And so there's model-driven drug discovery and molecular discovery and then we also need to say, how can we do model-driven biology discovery because we all know that we need to also expand our understanding of targets. And so those 2 things are absolutely right. So right in the areas of being able to measure it, read it, write it. And so that is really just starting to come into view here is super exciting.

So I guess specifically within the drug discovery piece of what you covered there, I guess, what do you feel people fundamentally either misunderstand or underappreciate about the applications for AI that are already happening today?

Yes. Thank you for asking that question because it's something that were thinking a lot about actually during COVID, as Albert was saying, like we saw what was essentially what feels like a missed opportunity in that, that $300 billion of R&D spend a substantial amount of it is in what's very deep domain expertise, experimental processes, lab work for a lack of better words. This is the IP of most biotech, tech biopharmaceutical companies and we could -- we were witnessing that they just were not taking enough advantage of it. And so with our collaborators in the industry, you could imagine that what is happening now is how can we create the conditions that AI is in the loop or lab is in the loop with AI. So every single experiment that you do, every idea your scientist has can now be codified to build upon an institutional knowledge that otherwise is maybe sitting in an electronic lab notebooks somewhere. But how could you represent even those decades of electronic lab notebooks in a model? And then every single experiment you make put it back in. And so we've invented a platform called BioNeMo to serve just that. And you see the reason I talked about these 3 scaling laws is because this lab in the loop is essentially going to require all 3 of those. It's going to require the very large-scale pretraining of these very large foundational models, tens of billions, hundreds of billions of parameters. We just reached $40 billion with this latest foundation model we built. You're then going to take that model and you're going to use it to do all sorts of tasks generally ideas, the scientists otherwise wouldn't. And then you have to prune those down. You have to use prediction models to prune those down. And then you might want to even optimize them before they go into the lab. So you're doing all of this AI work upstream of the lab after you go into the lab, you take all that very, very high value, essentially gold standard data and you get it back into the model and you go through this loop again and again. And that is what is really starting to shrink what is otherwise a 5- or 3-year process. We're talking months now. And the readouts from the industry are happening very often. And so I think it was very underappreciated. They know their data is valuable. They didn't realize that they could put it in a model where it would be completely reproducible, reusable, build upon this institutional knowledge. And let's face it, this industry is a challenging one because it's very transient. Even the scientists are jumping from one company to another, and you have the potential to lose that deep level of understanding, when you can represent it in models, it's now there for your company, your enterprise, to extract the most value out of it. And so I'm super excited about that, and you see the work that we're doing with Genentech and our friends here in relation like lab in the loop is absolutely how R&D should be conducted. Capture every ounce of data that is so high value to teach us how to make the best next scientific decision. And we're right at the cusp of having AI scientists, fantastic papers have just recently been published, where you're going to be a biologist, you're just going to talk to the computer. You're going to banter with it. You're going to hypothesize with it. It's going to automatically go off and run some experiments for you, come back, synthesize what it witnessed in those experiments. And you have dynamic conversations with the AIs. And so that's coming, and it's really exciting.

I wanted to follow up on BioNeMo actually. So in the tech community, we've become used to sort of an arms race to build foundational LLMs. But with BioNeMo, it seems like NVIDIA is doing more that work on behalf of and in partnership with. Is that the right way to think about the go-to-market and your relationship specifically in digital biology. And then I guess on that note, given how critical data sovereignty and ownership in creation and capture is, how does that work? How much of the frontier model -- or call it a frontier model development you're doing with BioNeMo?

Yes. I appreciate that question also very much. We are 100% an ecosystem and platform go-to-market. We trained some biology foundation models to understand the process. The end-to-end life cycle of building these models is extremely -- we are in that business to codify that process into a platform like BioNeMo and give it to the industry for exactly that reason, their data is their IP and should remain that forever. So how can we build enough of a foundation model that we took some of the hard work out. We codify, okay, how do you have to process the data beforehand. What model architectures are needed. How do you share that model up across a data center of thousands of GPUs to train it in the shortest amount of time and cost footprint, total cost of ownership of training those models. How do you then package those models in with, again, the lowest cost to run it? That whole process is what we essentially build into BioNeMo and we give it to our friends like relation or Genentech, and they train their own models. And so that is our go-to-market is in the BioNeMo platform being a platform that can be deployed anywhere, it would be like on-prem in a public cloud, in a supercomputing center somewhere right next to an instrument, it's really completely versatile, but we turn it into tools that are much, much easier to use, and we take a lot of the hard computer science and we codify the data science know how. So our go-to-market, as I said, we partnered with Arc Institute because they had the data, they had the scientific expertise. We have the computer science expertise, and then we turned it into reusable recipes that is essentially codified in what is called BioNeMo. So every other researcher can take and benefit from it.

Can I actually just ask now where you stand today? Are there any specific breakthroughs across the AI health care space that you think will be potential watershed moments for even the biggest skeptics that we have today? Is it first AI generated drug to be approved by FDA? Like what does that look like to get the last of people on board with this?

I absolutely do think that it will be a watershed moment when one of those drugs gets through. It's not that there isn't a huge amount of activity going on using AI for drug discovery. But it is the case that it's hard for investors to stay in it for a bit of this long haul and see that AI is actually the R&D productivity is not this AI found in new drug. It's all about the process of doing R&D discovery in a much more efficient way, but that's not as sexy to invest in. And so I do believe that there are several late-stage AI discover drugs now that when that happens, I think it will be a huge wake-up call. I kind of akin it to the Tesla Motors moment in the automotive industry, where it's going to be then a foregone conclusion. We have to go electric. It's a foregone conclusion. We've got to have AI in the loop in everything that we're doing. So I would say a hefty part of the industry is definitely -- there's the top that is doing AI in the loop, there is the next generation that is thinking about doing it. If that were to happen, it would just be everybody is going to see it as the only way forward because it truly is the only way forward. So I think that is going to be a fantastic watershed moment.

I guess just really quickly before I kick it back to Josh. Are there any important considerations from a regulatory perspective that investors should bear in mind as we try to understand kind of the market potential for AI within health care, obviously, it's a moving target daily at this point, but where things are today, anything that we should really be aware of?

I don't think that there's anything that has changed substantially. I do think the regulatory market industry, excuse me, they're really trying hard to keep up in it. I hope we all feel a bit of onus in trying to educate them because it's really crazy. Even in medical imaging, Aidoc is one of our great partners. They've created now an imaging foundation model, and they were able to get an FDA approved algorithm that is on the basis of that foundation model approved within months. This is a good sign. But of course, we don't want to do things that are unsafe either. Just like in the car industry, for example, the RAND Association declared you've got to drive 1 billion miles before you could ever be considered safe to drive autonomously, right? And if you did the math on that, the current fleet of instrumented cars, it would take a decade or more to be able to drive that amount of miles. So how could technology potentially help you solve that? One of the things that we're leaning into, and this is going to be huge in health care robotics space is simulation environments. You can train the car to drive in a virtual world because we now have the ability to do such high fidelity and create the corner cases that you would probably never still see on those 1 billion miles traveled, and we can build that understanding into the model and make them much more robust. And so we're going to have to help the FDA understand that there are technological advancements we can apply to show how much safer in fact, by using computational methods. We're going to be making a lot of these AI algorithms or these medical systems going forward. So I think they're working really hard at it. I really respect the work that the FDA is doing. I think we can play a pivotal role in creating technologies. And then I think the conversation has to just be incredibly real. Nobody has seen this pace of innovation in 100 years, right? I mean we're all breathing from a firehose. We're all waking up every single morning being like what's new in AI today. So try doing that in a regulated industry, and it's something that requires the entire industry to work on.

On the topic of hurdles, do you feel like the health care industry is ready for all the innovation you're driving and I ask for 2 reasons. One, at least from the outside looking in, is not necessarily the first one to adopt new technology usually in my dad machine, he is a physician. And then also from a cost standpoint. Do you feel like we're at the point where cost is still longer prohibitive. Obviously, it's going to continue to move forward. But how do your customers think about where they are in the cost curve when they're adopting AI?

Yes. This is a great question, too. I mean, you're right, the cost is getting completely manageable now. And in fact, we've improved inference costs by 200x in just 2 years. That means the cost curve of technology is going like this. the cost curve of human is on the other extreme. And so I think that is starting to be understood. What I don't think people quite understand yet is the total cost of ownership and actually the return on investment can be huge. I have this fantastic example of building an avatar, a digital human with Deloitte for the Ottawa Hospital. And every time you go in for a surgery, there's always what's called a preoperative appointment. These preoperative appointments can range from an hour or 2 generally with a trained clinical health care staff member. And a good 50% of that conversation is nonclinical. Where do I park? When do I need to stop taking my meds? When do I not -- I need to stop eating? Can I exercise before all of those kind of things where -- imagine you had a digital agent that you could call up any time of day and ask as many questions as you wanted. They recognize that they do 80,000 surgeries a year. Think about that 160,000 hours of time spent. Imagine being able to deploy a digital system to take on 80,000 of those hours. If you did the math on that, that's essentially like being able to hire another 40 nurses. So the economics are fantastic actually in that situation. And not only are the economics and return on investment, fantastic for the hospital, which as we know here in North America, our economics are extremely poor and declining. So we need all of that capture back but the patient experience, they studied the patient experience, it was 100%. Think about it as a patient, you have a lot of angst. You want to ask a lot of questions. If you're doing it with a human who you know is short on time, you're not going to ask as many questions. And so the patient experience went way up as well. So there was a win-win situation here on deploying digital technology. So that's an example where I don't think we're doing a good job describing that return on investment in that TCO. That was very simple math and pretty mind-boggling math. When it comes to -- let me use the second example you gave of dad and fax. And unfortunately, our health care system was architected for billing and things that are not really in service of the patient from the get-go. So we have a system that's been really, really tricky to innovate with. I believe that is not any longer. I believe that because of Agentic AI and the way that it is architected, it can overlay on this system, unlike introducing technology in the past. For example, what can agents do? They can connect. So they have the foundation model, but that wasn't enough. They can connect into an API, we did build that into the health systems so they can API in, get your personalized data. They can API in into a tool that is maybe an antiquated scheduling system. And so they can kind of navigate autonomously over what is otherwise an antiquated and very complex and disaggregated health care system. And so we have a lot of friends in the room who are building agents on top of this data to create new patient experiences, a bridge I don't know if your dad gets to use a bridge, but this is a clinical documentation platform that is a phone that the clinician sets in your room with them says, do you mind if I record and now has all the opportunity to focus on you and not the computer. It summarizes, it does all of the clinical summarization and has a way after the clinician checks it for whatever it needs to do for the minor correction, and it goes right into electronic health record. This system is giving clinicians back 2 to 3 hours a day, a day, right? So this is, I think, what we are right at the precipice of these. And as I said, these systems are for hire. They're on demand. They don't require you to hire a system integrator and get a $3 billion budget to hire a digital preoperative appointment agent. You can essentially do it, do it on demand. And so I think that's -- that transition is going to be right here in front of us. And you can already feel it as a patient because a bridge is in my doctor's office.

Well, unfortunately, the red light says we're out of time. But Kimberly, what you and your team are working on is important. It's inspiring and it's exciting. So thank you and good luck and thank you for doing this.

Thank you. Thank you for having us.

Thank you, everyone.