NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

All right. Good afternoon, and welcome to JPMorgan's 44th Annual Healthcare Conference here in San Francisco. My name is Harlan Sur. I'm the U.S. semiconductor analyst for the firm. For the seventh year, we have the team from NVIDIA presenting. As all of you know, NVIDIA is the leader in accelerated computing in AI semiconductor, software, systems, enabling the development and deployment of the world's AI foundational models, like large language models, enabling next-generation reasoning and agentic-based frameworks and now moving the industry adoption curve to physical AI and driving compute innovation for cloud, hyperscalers as well as large vertical markets like health care and life sciences. Here with us today from NVIDIA is Kimberly Powell, Vice President and General Manager of Healthcare at NVIDIA. She's responsible for the company's worldwide health care business, including hardware and software platforms for accelerated compute and AI that power the ecosystems of imaging, genomics, life sciences, drug discovery and health care analytics. Kimberly, great to have you back again. Let me turn it over to you.

Thank you, Harlan. Thank you so much. Thank you. Good evening, everyone. This is the first time I've been between you and cocktails, but I'm going to make sure it's as entertaining as humanly possible. Just a few words that this is an absolute once-in-a-generation platform shift for the health care industry. And I am so honored to be invited back for the seventh year, we're in our 17th year of working on health care at NVIDIA. And I thank the conference very much for this opportunity, and all of the partners that we work with day in and day out, which I hope I'm going to be able to share what the future is going to look like. Before we get into it, please have a moment and look at our forward-looking statements. And action. Okay. 2025 was an absolute breakout year for agentic AI. Many things came together in just the last 12 months, you've heard the words reasoning, AI models that can reason. You've heard the words tool use, software that can actually use tools on behalf of the human user. You've heard the word retrieval being able to attach language models with trusted information and trusted knowledge. Agentic AI is here alive and being deployed faster in health care than any other industry. The ChatGPT moment, Jensen just described to us last week, has arrived in physical AI, the amount of progress we are now able to make in robotics because we've closed the loop in a very important domain called simulation, where robots and embodied AI will learn in a computer first before they're ever deployed to the real world is here, and it's having profound implications across the entire health care and life sciences industry. And then thirdly, and we've been working on this for some time, and many companies are here sharing with you that AI is starting to understand and learn the laws of nature, biology in particular. You might call us at the beginning of the transformer moment of biology. Let's start, though, with thinking about what's happening in AI. Let's take a minute to understand how AI is really making such rapid progress. One of the most important things we need in the world is open models and open software. Just like you could think about Linux back in the day as an operating system that created brand-new markets. We're exactly here at this time. And what's amazing to think about is open models are now reaching the frontier, which is giving an opportunity to every startup company and actually every enterprise to participate as fully as very well-stocked AI labs around the world. Open models and reasoning models that really came into light at the beginning of 2025 are absolutely the backbone of innovation. These are models that can think and they're much more relatable to humans, and they can create the essence of transparency and explainability. They can break down very complex tasks that otherwise were just untouchable or had to be hand coded in the past software. So 80% of startups today are built on open models and it's in a very, very important strategy to NVIDIA. Over the past several years, we have been amassing a huge body of work in the open source. In 2025, actually NVIDIA became the world's largest contributor of open-source AI on Hugging Face. We have over 650 language models that have been contributed, 250 data sets. And those are not only in language. They're also in biology, in chemistry, robotics and vision. And so key to developing an ecosystem is to not only provide open models. When we say open models, we actually made 3 things. One is the model itself. Two are the open data sets. For any company or any industry that is regulated, you need to understand how these models came to fruition. You might even need them for auditing purpose down the road. So open models, open data set. And the third is open tools. Just like all intelligence, learning is never finished. Every single user interaction you have with your software, with your application is training data to enhance the system going forward. So you need to create a whole tool chain for the end-to-end AI life cycle which is essentially a never-ending life cycle. So when we say open models, yes, the capability of open models are extremely important, but it comes with open dataset and the open tools. And so we've been pioneering in some very important places. We just announced the third generation of our Nemotron language models that are absolutely at the frontier for Agentic AI, we've recently announced our physics AI models, Earth-2, things that we can do in climate and weather simulation. Our own Clara models in health care for biomedical AI that spans everything from target discovery to molecular design and medical AI reasoning. And then we're at the dawn of this ChatGPT moment for physical AI. And a lot of that physical AI has come from the foundation models we have pioneered at NVIDIA, last year's CES was actually the breakout moment for Cosmos, which is the best innovation of the show for understanding the world, world foundation model, understands the laws of physics, understand spatial awareness, can create digital worlds of all kinds with millions, billions of permutations in which robots and physical things can learn in these environments. Group is for robotics, so that you can train robots to operate in these physical worlds, all of the different training policies, all the different tasks that needs to take us from very specialized robots to more generalized and they can complete really amazing tasks. And just last week, Alpamayo, which is for self-driving cars. The first time we've ever opened sourced. This is a model that is essentially thinking autonomous vehicle model. It has large language models at the root of it, and it's an end-to-end driving system. Incredible work that is going to lay the foundation for much of the physical AI to come. What I also love about 2025 and us moving forward is agentic AI has become hireable. I would hire this man. I think all of you would as well. And what do I mean by that? All of those breakthroughs that I just described, the ability to reason, the ability to call the tools necessary, the ability to interact with antiquated systems, whether they're scheduling systems or otherwise, all of that has largely now become solved in the age of Agentic-AI. And so health care systems all over the globe are recognizing that they can start hiring these agentic systems and platforms essentially as digital coworkers to close this extreme gap we have in terms of health care services and the number of health care professionals we have. As you know, the World Health Organization predicts we're going to be tens of millions of health care providers short by 2030. We can offload our amazing health care professionals who dedicate their lives to their profession and offload a lot of the clerical work that isn't necessary clinical work. I love this report from Menlo Ventures that describes that I would say, for the first time in my history and my being in this industry that health care is leading the pace at a technology, enterprise deployment and adoption. It's actually at 3x the pace of the U.S. economy. And that is because it's solving such acute challenges. And so it absolutely is here a USD 4.9 trillion market. And we are deploying AI at this incredible scale. These are paid and enterprise-grade software systems that are now being hired in the psychology of CIOs and health systems all around the globe are recognizing this as an opportunity. It is just not possible to go out and find another 500 doctors that you can hire into your system. But by offloading your amazing doctors with systems that I'm going to share with you in just a moment, we have an incredible opportunity to do so. I've talked about this -- we've talked about this for quite some time that the way software is being built has fundamentally changed. This is, in essence, what all agentic systems and what all Software as a Service platforms will look like when they are agentic. They are prompted with some input and this amazing reasoning system kind of understands the user intent. You will always use frontier models, and you will also augment these frontier models with specialized models because the work that gets done in industry is exquisite work, it's specialized work. There's subject matter expertise involved. And so you have to call upon a lot of different tools in order to connect these agents and connect these otherwise generalists to become specialists and deliver the value that's necessary in the industry. Let me share a couple of amazing examples. I think many of us here have heard of Abridge and actually, they're on the conference scheduled this year. Abridge is a clinical conversation AI platform. Their platform again, looks like those systems by connecting these systems in such exquisite ways, understanding workflows in a building block sense to transform workflows. They're giving 30% or more of doctors' time back at the end of the day, helping them generic reports and prior authorizations, deployed in over 200 health systems already, and that number is growing dramatically faster even as we seek in the last 6 months. Corti is a health care agent platform that is helping Europe and the NHS deploy agents of all kinds. And similarly, we have Speechmatics and Sully, who are creating agents to triage, creating agents to check you in, agents that can be deployed all over the hospital for, again, not necessary clinical work but amazing workflow that creates a win-win situation. It's a win situation for the health care systems because more patients can come through. And it's a win for the patient because the experience is far improved. Now these agents are going into another high stakes, very high-cost area of the industry, and that is in clinical development. This is a part of the drug discovery and medical device process that is absolutely necessary. But it's a very challenging part of the system. It's very labor-intensive. It's very manual. It's frankly error prone. We work with amazing companies like ConcertAI, which is helping to stratify these clinical trials and even simulate their outcomes so that you can do much, much better planning, the amount of money, time and resources that can be saved and the precision to get to where these clinical trials need to go faster. CytoReason is essentially building the capability to do drug development by building disease models, using knowledge graphs and otherwise to really understand and help build better modeling with all of that real-world data. And we've been working with IQVIA now for well over a year, and they have agentic systems being deployed from the commercial deployment of commercial teams who can give you an essence of in what region, what physicians you should call on with very relevant data so that they can have much more productive commercial teams as well as into all of the clinical trial, finding the right start-up studies and building those at a much, much faster pace than before. So this agentic digital health ecosystem is being built on NVIDIA, our open models, our tooling and our ability to help them connect and build out these agent systems to do incredibly wonderful things. Now agents are at also a very exciting inflection point where they are accelerating science. This is the loop of science that is emerging, not only in life sciences, but particularly in life sciences, it's having a very, very accelerated effect on how we're thinking about doing science. Science -- AI scientists are agentic systems, who you could imagine, could go off and read the literature, you can go back and forth and reason with them, they can help you design the experiment. They can call upon tools that could be foundation models for protein structure prediction or do virtual screening in the digital lab or they could actually go off and kick off and experiment in a physical lab. And you can also think about the computational dry lab as this connective glue to close this loop as we described. As you have the agentic system, you have the physical space, but you need to constantly take every experiment and build that into digital intelligence of the R&D work that's being done all over the world. There is a new emerging ecosystem of a category of AI science companies who are building on, again, NVIDIA's Nemotron, a huge additional breakthrough of last year that's come into vogue in order to create these agents that go beyond generalist understanding and into science and technology, and it's called reinforcement learning, using experimental data to reinforce these models and tune them into a very particular science task. Edison is the new commercial company that came out of FutureHouse. This is a stunning AI scientist. This scientist can go off and read 1,500 papers, write 40,000 lines of code and synthesize a research report in about 16 hours, work 16 hours straight and essentially do the amount of work that would otherwise be 4 to 6 months by a researcher, pretty mind boggling. LiLA is building a super intelligence, a completely integrated autonomous lab and agentic system. So literally, all experiments come back and feed the super -- to feed the super intelligence and creating a complete closed loop system. And Owkin is combining biological language models with deep patient data to really help biopharma teams have a higher confidence in their decisions. So the science agents are really here and they're making a profound impact on that -- there's really a new paradigm in science that is emerging. And as we know, life sciences is one of the largest science domains and pharma R&D is the largest of that. And so a $300 billion industry of R&D is going to be reinvented with this paradigm. I want to share with you how agents are entering the lab. They're not only going to be co scientists along with you that you're kind of talking with, working with, hypothesizing, they're creating reports for you, but they're actually going to do work for you on behalf of the scientists or with the scientists in the lab. Let's take a look. [Presentation]

We're super excited to announce today that we're working with Thermo Fisher Scientific, the world leader in lab instrumentation and services to build what we're calling the fundamental AI infrastructure for the lab. You can see there that little gold box is essentially a bench top AI supercomputer called DGX Spark. This DGX Spark can run any workload of AI and accelerated computing. You can hold it in the palm of your hand. And so we pioneered this first agentic system where you essentially can make the instrument intelligence right there with that amazing gold box and some agents that we built together. And you can close the loop from the scientists there. Sometimes, you can close the loop right with the instrument. The instrument with an automatic quality control agent can understand, oh, I need to go clean something in the instrument and it can autonomously self-relieve it and come out with much better experimental data. And so it is very clear that this is going to drive -- this is going to scale the throughput of labs. It's going to increase the quality of experiments and no longer are humans going to be kind of the thing that's bottlenecking the amount of data that we can come in and do science. So this is just an amazing partnership and we're delighted to be partnering with Thermo Fisher. Now getting into physical AI, labs are one of the most chaotic and bespoke physical environments that are out there. And so we really want to think about scaling labs with robotic intelligence and into those real worlds of physical labs. And so you -- there's a journey here that we want to have not only specialized robots that really understand a given instrument, we want them to also be generalized and actually, we want the best of both worlds. And so to get to that best of both worlds, NVIDIA has created the physical AI 3 computer platform. I was describing this earlier, where you use simulation and our Cosmos World Foundation Model to create digital worlds to train these robots in, you can vary the lighting. You can move beakers. You can practice all sorts of different tasks and train your robot in simulation first. You can use Isaac in the training platform to train all sorts of different types of robots or for tons of different tests. Sometimes they need to be contamination tasks, sometimes they have to have different perception because they're looking for barcodes or they're looking for a glass like this size or their pipetting. And then we have also the edge computer for you to go off and deploy this. So AI and lab automation is reaching far into the physical world. And so again, a new class of companies is emerging into not only lab automation, but robotic lab automation. We're working with some fantastic companies in this space. Multiply Labs is using our Isaac platform to train their robots. They're doing amazing work in cell and gene therapy, biomanufacturing labs where they're using the Isaac system to train literally thousands of different tasks because as you're going through some of these more complex therapies, it's many, many steps involved, and these are very precise steps, the precision and you don't actually want humans involved because of the contamination aspects of it. And so they've made some tremendous breakthroughs where, take a cell therapy that costs $100,000 to manufacture, and they're reducing that down to $30,000, over 70% with their robotic systems. And they're essentially putting 100x the throughput in a given square space, square footage lab environment. These are the breakthroughs that are going to scale medicines to where we need to go. Similarly, HighRes build complete lab automation at a very large scale, exquisite robots that are learning, again, in our environment using Isaac and Cosmos to train them and learn all of these different task to take automation to yet another level. And then Opentrons very well known for their liquid handling, deployed in 10,000 labs around the world, again, using our platform to build the simulation environments and increase the velocity at which these robotic systems are able to tackle more and more complex tasks in the lab. All right. We are in a final chapter here of AI starting to learn the laws of nature. I took a few headlines. The NeurIPS Conference is the flagship AI conference of the year. And there was over 30 workshops, developers of all kind. There were 50 or so biology and life science company parties at this event. And it was written up as biologies transformer moment. AI revolution and drug making is well underway now. And we're really starting to see AI-enabled medicines reach the later stage of clinical development, which is extremely exciting and a lot of those companies are here. We are working really hard to help push the frontiers in this area of biology transformation models. We don't aspire to necessarily be a biology foundation model company. But all of the methodology, all of the challenges in which it takes to scale these models at a domain-specific level, language is short words, assembled into a sequence of a sentence that looks very different than a 3-billion-character long DNA. And so we need to think about context length when we're talking about biology. We need to have different model architectures. These models have to get grounded in physics. So there's a lot of interesting challenges. And so we've been adding to our Clara open models to help the entire research industry really accelerate the ability to train larger models as we go here and multimodal models. And we're really proud of some of the work we're doing in La-Proteina, which allows you to add atomic scale, essentially design proteins, our very new one that we're announcing is RNAPro for RNA design, the first that we've had. And Merck was just up here. We did some exciting work with them on that KERMT model, which is all around predicting toxicity. And so we're really trying to work across the drug discovery process with these open models. We have a reasoning model for molecular synthesis in our version 2. Really excited about these models. And so we also announced today a pretty massive extension to the NVIDIA BioNeMo platform. So not only are we investing heavily in these open models. But additionally, and as I said, it's not just about the model, it's about the data sets. And we have a road map to continue to invest in data. This industry, like other industries where it's self-driving cars will benefit from synthetically generated data. And so we've generated some synthetic proteins. We also care a lot about doing data processing, things like cheminformatics workflows like RDKit, we now have a GPU accelerated nvMolKit that is 100x faster in chemistry processing. And so this platform expansion is really, really foundational. And as I said, it's part of that glue. It's that digital dry lab that is going to take all of the intelligence from experiments and continuously enhance these models going forward, which can then be called by the agentic systems as their tools. So what's exciting is we're seeing enterprise adoption of BioNeMo in some pretty exciting platforms. Basecamp Research is an AI-native company who announced their EDEN platform today here at this conference. This is a GPT 4 size biology model that was trained on 10 trillion biology tokens. And it's able to now do things what they're calling gene insertion and their lab result -- their validation results of what it's able to do in antimicrobials and in cancer is really pretty groundbreaking, amazing EDEN platform. We're working with Natera, who is for cell-free DNA, they're training their own models and then also building into their platform, agentic systems to advise in clinical development and advise in the clinical decision support. And then TetraScience is a scientific data platform, again, to try to connect all of these. As we know, science is done oftentimes in very narrow. But when you can try to now start to learn across many different data sets, to ask all of these scientific questions, we're working with them. They're deploying BioNeMo models inside. They're deploying Nemotron, and it's an amazing platform for scientist and so this is the vision. This is the new paradigm in science. And as I said, there's an amazing group of new markets, new companies that are being built all around this vision, of AI scientists that can call tools that are constantly getting smarter by the experiments. And these experiments have automation that will come from the agents setting it up, but you're still going to need robotics and otherwise to help you execute that in the physical world. To bring this whole vision together, we are so excited to announce an extension to our partnership with Lilly. Today, we announced a first of its kind co-innovation AI lab with Lilly. This is the first time we are going to be joining together world-leading scientists with world-leading AI researchers in South San Francisco here in the Bay Area, co-locating with the amazing science and lab understanding that comes with doing drug discovery, we'll be investing $1 billion over the next 5 years to really push the frontier in this new paradigm of science and new paradigm in acceleration of drug discovery. This is building upon their deep belief that they see this transition of 90% wet lab to 10% compute, imagining the paradigm with that's deeply, deeply flipped in the next coming years, and it's going to accelerate the breakthroughs. We're going to work on clinical development, and we're going to work on manufacturing and lab automation, just like we described that they're doing. Lilly is world-class in manufacturing and accelerating the ability to deploy physical AI throughout labs and manufacturing is also going to continue to be transformative and help them meet the amazing demand that they've created for medicines in the world. So this has been a phenomenal kickoff to the year. I'm going to leave you with one last video, and then I'm going to join Harlan over there for some Q&A. [Presentation]

Great presentation. Thank you, Kimberly. I'm going to kick off the Q&A.

We've seen the deployment of massive what Jensen calls AI factories by the leaders in your segment of the market like Amgen, Genentech and recently, Lilly with their Blackwell Ultra base DGX SuperPOD, right? And this potentially signals the shift from pilot programs to industrial scale development and deployment, can you walk us through the economic conversations you're having with pharma CFOs today? Are we at the point where they view GPU compute investments, not just as an R&D expense, but as essential capital infrastructure that integrates AI agents, AI tools that directly determines their pipeline throughput and probability of success?

Yes. I do think, as I was just describing, this is a new paradigm in science completely, and the amazing scientists that -- and if you think of this as a scientist or an employee, you really want them to be as productive as humanly possible. And so if there's a new scientific methods sort of emerging with the ability to take all of -- think of Lilly in many pharmaceutical companies, hundreds of years of science is written down in electronic lab notebooks and it's kind of shoved all over the different parts of the company. You actually have the ability now to build all that back into the Cardinal knowledge of the company, every scientist that works there, it's a very transient industry, frankly. But why lose all of that deep understanding of a scientist when they leave the company, you can actually inject that back into the system. So from that respect, it's taking all of that amazing high-value data and doing something with it to empower the whole organization and then this transformer moment that we're in is becoming very clear. I mean we're in the fifth year now post AlphaFold and its true initial impact, and it was the inspiration that has driven a lot of work in the area of models. And now there's thousands and thousands of biology and molecular models being built every single day. And with the sort of democratization, if you will, of what we're doing with open models and the data sets and the tools, we're giving the capabilities for science teams who are not open AI like, they are scientists for the job they're hired for, but they can now become AI scientists and AI researchers just the same because we're making it much more accessible for them to develop these things. And so I think that the visual of you have agent scientists working along with and then you have a dry lab wet lab. I mean you said it is going to be exactly like the wet lab and you're going to have the dry lab be as intelligent and reasoning with you and calling upon all of the data that you've ever built and all the new data that you're building. And so I absolutely think that it's going to be thought of exactly like your wet lab, and I see this 90-10 flip starting to go the direction. And it's not going to be less lab expense we're just going to do much more science. That's what this is all about. This is not a paradigm in which, if you flip it to computation, you don't need more of that. Absolutely not. Just think about radiologists and the reports are coming out. Everybody thought because you can read -- you can do the task of reading an image and finding something in it, which is one of the tasks that a radiology does, that we weren't going to need radiologists soon. That was what one of our godfathers said, well, in fact, reports are just out. We've increased the number of radiologists that are being hired because there's more work to do. And so we should just think of this all as we're going to do fundamentally much more science, which essentially will also lead to many, many more breakthroughs.

Last week, we held the consumer electronics show, right, Jensen had a big NVIDIA live event where he announced his new -- the team's new GPU computing platform called Vera Rubin, right? And one of the key highlights of that was Vera Rubin is going to continue to drive cost per token or cost per inference lower by up to 10x, right? And every generation of GPUs that the team has brought to market cost of inferencing, cost per token is going down somewhere between 3x to 10x. This is per year, right? So with that in mind, in hospitals, you've highlighted the labor shortage crisis and introduce agentic AI as a solution for everything from patient triage to administrative coating. For a hospital CEO operating on razor thin margins, what is the immediate ROI of deploying NVIDIA-powered agents compared to the traditional staffing? In other words, is the cost of inferencing finally now low enough to make this viable for mass market sort of health care adoption?

Yes. And you're right. In the last 4 years, we've had Hopper to Blackwell, to Rubin. And in those 4 years, we've reduced inference by well over 100x. So if you're paying $1 to run an agent, you're now paying $0.01. And you need this, you need this for rapid adoption. And so there -- these companies that I just described, a bridge has hundreds now of millions of users, right? OpenEvidence has hundreds of millions of users, and they're using it constantly. And so we have to continue to drive the cost down. Now the return on investment is very clear. If a doctor has 30% of their time back, that's either 30% of life that they can return to having with their family and keeping them employed and safe at work or you can also see 30% more patients. I mean it comes with all sorts of benefits. It's a win for the patient. It's a win for the health system. And so you can measure a lot of those companies that we talk to, they're literally measuring how many clinical minutes they're giving back to the organization. And I think it was the Sully and Speechmatics, there's something like 57 years they've already measured since the platform has been in deployment that they've given back to the health system. So that's clearly measurable in ROI because the more you can give back of free minutes is essentially the more patient throughput you can have.

When we think about accelerated compute and AI, we typically think about use cases, customers as being the large cloud, hyperscalers, your corporate and enterprise partners, but Jensen always reminds us, right, that there's a sovereign AI opportunity. It's a $20 billion per year market opportunity. We've seen Japan's Tokyo One. We've seen Denmark's Gefion supercomputers launched with a very heavy focus on health care and genomics driven by the need for data sovereignty, national competitiveness. Do you view sovereign health care cloud as a stand-alone growth factor for the team separate from enterprise and do you expect every major economy to build a similar type of infrastructure build-out?

Yes. So to answer your last question first, yes, I expect every country to be able to take advantage of this incredible again, once-in-a-generation opportunity. Some countries will go from 0 health care services to complete AI-native health care services, and that's a fantastic opportunity. And now that we've made it so accessible to do so, we've done several things. NVIDIA's platform is inside of every public cloud, NVIDIA has also pioneered essentially a generation of what they're calling neoclouds. So clouds that are residing within the walls of certain countries, giving every country the opportunity, I mean if you think about what AI infrastructure is, it's just as important as roads, electricity, water, it is a necessary infrastructure for any country to prosper in the future. And so they can get it in the public cloud, if that's good for them, they can start building their own, a lot of telecom companies are transitioning themselves into cloud companies that can be hosted. You can build it inside your own enterprises, if you like. And so to answer your first question, last, no, it's not a separate. It is all part of our enterprise business. We've just now created the conditions that everybody can, should and will build their own infrastructure to serve their own country to prosper.

Perfect. Well, we are just about out of time. Kimberly, thank you for your participation. Looking for another strong growth year for the team.

Thank you so much.

Thank you.