All right. Good morning, and welcome to JP Morgan's 41st annual Healthcare Conference here in San Francisco. My name is Harlan Sur. I'm the semiconductor analyst for the firm. For the fourth time in five years, we have the team from NVIDIA presenting. In fact, in the 41-year history of the healthcare conference, NVIDIA has been the only semiconductor company to present here. Very honored to have them again here this year. For those of you that don't know NVIDIA, leader in accelerated computing semiconductor systems, hardware and software platforms in areas like artificial intelligence and deep learning, powering some of the world's most powerful supercomputers. Driving compute innovation for cloud and hyperscalers, as well as large vertical markets like healthcare and life sciences. Here with us today is Kimberly Powell, Vice President of Healthcare at NVIDIA.
She's responsible for the company's worldwide healthcare business, including hardware and software platforms for accelerated computing, AI, visualization that power the ecosystems of imaging, genomics, life sciences, drug discovery, and healthcare analytics. Kimberly, thank you for joining us today, and let me go ahead and turn it over to you.
Thank you, Harlan. Thank you very much. Good morning, everybody. Nice and dry. An honor really, and I thank you, Harlan and JP Morgan for inviting us back. This is absolutely a different kind of feeling year, and I'm excited to really share it with you. Let's just do the first housekeeping. Let me start with a reminder that the presentation and QA contains forward-looking statements, and investors are encouraged to read our reports filed with the SEC and the information that relates to our risk and uncertainties facing our business. Super. You might know NVIDIA as an amazing chip company, and in fact, we are. We're much, much more than that. We are an accelerated computing platform company. Some scientists, some industries actually describe us as a time machine. We launched our GPUs into general purpose programmable processors with our programming model called CUDA.
We now have well over 33 million CUDA developers. For the last 15 years we've been developing a full stack of computing approach. The ability to program the chips, but also be able to introduce acceleration libraries, applications, and purpose-built computing platforms, whether that's large scale data centers for artificial intelligence or embedded supercomputers for the medical devices industry. For each field of science and industry, and applications, we create a full stack. Everything from gaming to design, earth and life sciences, and self-driving cars and robotics. This allows us to really serve $100 trillion worth of industry. We're here to talk about healthcare. For healthcare, we created a platform called NVIDIA Clara. It's after Clara Barton. She was the inventor of the American Red Cross, not after Santa Clara, where our headquarters is. It's an AI computing platform for healthcare.
We recognize that healthcare is becoming the absolute largest data generating industry, and we have global challenges increasing the cost of healthcare delivery and access to healthcare. We build computing platforms to serve these grand challenges. Leveraging world-class chips and systems, we build platforms that span embedded edge all the way through to cloud applications and frameworks that span from healthcare delivery to disease and drug discovery. I hopefully show you some of that today. Our computing platform is widely adopted. We're relied on by all of the top medical imaging companies to turn sensor information into rich images. We're quickly becoming the de facto standard for surgical robotics platforms for real-time sensor processing. We're helping the medical and cancer centers be able to be really data-driven in their clinical treatment decision-making.
We're increasing the accuracy and throughput of sequencing so that we can reduce the overall cost of genomics and hopefully bring it more into the standard of care. We're turbocharging the drug discovery process using artificial intelligence to really illuminate biological meaning and explore literally the infinite possibilities of small molecules and antibodies. In healthcare delivery, imaging is the essential tool. It's used at every stage of the patient journey, which is why it was one of the application areas we chose to focus when I started at NVIDIA 15 years ago. As sensors evolve, so do the computing platforms and the applications in healthcare delivery need to evolve. From 2D sensors that allow us to do annual screening into advanced 3D and 4D to do quantitative image analysis and to capture things like function in the new pho t on CT scanners that have recently been released.
Now we're entering into what I call the fifth dimension, where in real time, with real-time sensing, these devices can really actually take action. They can do things like self-navigate. They can do things like adapt while you're in the middle of a radiation treatment. Very, very exciting times. Much like the self-driving car industry, I love to use that parallel because we can all sort of understand it, the healthcare industry is becoming software-defined, and it's able to deliver great value through software. In order to do so, you need two computing platforms. You need an AI development platform and AI deployment platform. You need to connect these two so that the data that you are experiencing on the edge is very dynamic, and it's that data that you can collect at the edge to bring back in to improve these applications and redeploy them.
This is the as-the-service architecture, and NVIDIA is building this end-to-end computing platform to serve it. Here's an example. NVIDIA Clara provides the imaging industry, radiology, pathology, all the surgical data. The AI development platform is called MONAI. MONAI is an open source AI framework that we co-develop with the industry, the academic medical centers and experts. We open sourced it, we have an AI deployment platform called Holoscan. Using MONAI as the AI framework for imaging, we've had over 800 million downloads of this framework. It's absolutely accelerated, and it's a driving force in the exponential growth in the AI research. You can see that through the PubMed publications here. To help close the gap between research and clinical deployment, we announced last year NVIDIA Holoscan.
It's meant to be a commercial, off-the-shelf, more general-purpose computing platform for these applications to live so that not every medical device needs to reinvent their computing platform every time a new sensor technology comes to market. A scalable real-time AI sensing platform, and it can scale, as I said, from embedded to cloud. We have developer kits already out there in market, and we announced design wins with several robotic surgery companies already. The innovators in the industry are using NVIDIA platform to deliver new applications across radiology into therapy. As you can see, just as we described, Carestream uses artificial intelligence to really have a complete smart AI X-ray room. It's at the room level all the way down to the device level to streamline workflows. You have Varian and Elekta, where they can now use AI to do adaptive therapy, real-time adaptive therapy for their patients.
I love the Intuitive Ion because it allows you to navigate into the lung and biopsy areas otherwise no humans could travel. These devices are absolutely opening up tremendous new capabilities, and we're excited to see the medical device market continue to go through this revolution. I wanna go quickly next to a modality that is just super exciting, and I think we all saw a 2022 that changed the landscape of genomics tremendously. Genomics is a modality that is delivering great value to healthcare as well as research and drug discovery. It is the largest data generator in healthcare and growing rapidly. We're witnessing the continued decline in cost, enabling large-scale genomics programs to transpire.
However, we need to be sensitive to the fact that a lot of times when you advertise the cost of sequencing, it's just the cost of sequencing and not the downstream analysis which is ultimately the insights that we need to care for our patients and to deliver insights into discovery. T his is why we are partnering across the genomics industry from new sequencers to bioinformatics platforms to cloud services and large pharma. 2022 was an absolute breakout year for NVIDIA accelerated genomics. We partnered with Oxford Nanopore and Stanford and many others to achieve a Guinness record in clinical sequencing. We made NVIDIA Clara Parabricks free for research and partnered with the Broad Institute and put it in the Terra platform. And it's also available in every public cloud.
Just this week, our partners at Bionano announced our work in accelerating 96 optical genome mapping workflows for high throughput on-prem and cloud deployments, where they're having tremendous outcomes helping patients with this new structural variant capabilities. AI is becoming vital to all areas of genomics analysis, in fact. Primary analysis, which takes the signal or image and turns it into the genomic base calls, is all done now through artificial intelligence algorithms because it helped realize a step function in accuracy as well as performance. The higher the throughput of the instrument, the lower the cost of sequencing. AI is also bringing that speed. Just a few months ago, Seattle Children's and the University of Washington were able to deliver a genetic risk assessment within three hours of a newborn's life to rule out a disease that their sibling had.
By reducing the cost, increasing the speed, and partnering with the clinical community, we can bring, you know, the condition to move sequencing more into the standard of care . We're really excited to do that across the board with all of our sequencing partners. With more sequencing platforms and modalities entering the market, we're going to be pushing, you know, these 40 exabytes of genomic data out there into the world, and we already have 500,000 genomes into 1 million genome databases associated with patient data, and these are becoming readily available. We must harness the latest breakthroughs in artificial intelligence to drive our biological understanding in therapeutic discovery. So we had some incredibly exciting breakthroughs at the end of last year at the Supercomputing Conference. It's a super neat example of how you can harness large genomic data to drive genomic understanding.
We worked with NVIDIA- Argonne National Labs, University of Chicago, and achieved this late last year and won the Gordon Bell Prize. You know, the latest technology that is taking the world by storm is generative AI and large language models, most notably ChatGPT. These models are trained on extremely large, unlabeled datasets, and they can learn context and meaning by tracking relationships in sequential data. Sounds very much like genomics, if you ask me. The same can be used on genomic sequencing data. We did just that. Co-developed a model called GenSLM, and it's the first genomic scale and the largest biological language model to date. It used 110 million bacteria sequences from the PATRIC database, and then we fine-tuned it with 1.5 million SARS-CoV-2 genomes.
The model was able to not only predict the evolution of the virus, so potentially be useful for an early warning system but it was also able to accurately identify variants of concern. This is all published in the paper is in the source there. Really super amazing breakthrough here. I want to kind of give you a sense to what are these models learning and how are we going to be able to really understand, you know, how they learn and represent biologically relevant information. Let me just kind of play this video the team made this to sort of give you a description of what's going on. You can see through clustering that the model's finding semantic meaning in the latent space. First, you can see it cluster genomes by their sequence length. Then we're going to move into here.
You can see it with its GC content, indicating secondary structure stability. Then we can even zoom in on particular enzymes, and we can see the structural differences. This is something actually the model was never trained for. These models are actually helping us with interpretability. No longer are the days of deep learning black boxes. These foundation models give us a new ability and become the bedrock for us to really be able to read and start to understand biological meaning. Very, very exciting opportunities here. This is, you know, language models applied to genomics. We are just getting started. You know, like generative AI has set off broad use across all industries. AlphaFold set off the ImageNet moment for AI and biology.
Standing on the shoulders of giants, and the breakthroughs coming from DeepMind and OpenAI, biology labs at Meta, Rose Lab, Baker Lab, Barzilai Lab and thousands more published papers in 2022. Generative AI and biology is witnessing that same broad applicability across life sciences and drug discovery. Like I said, from pandemic early warning systems to target discovery, to protein structure prediction, to virtual screening, you know, drug target interactions, and protein engineering, literally touching every phase of the drug discovery process. Biology is going from an empirical science, experimentation and exploring the physical and natural world to computer science and AI. Where it's quickly moving from a science to an engineering. Just this morning, we're announcing exciting results of a collaboration between NVIDIA, InstaDeep, who congratulations, who just announced they are being acquired by BioNTech, and Technical University of Munich. It's resulted in a state-of-the-art genomics language model.
Genomic language models are still in early investigation, DNABERT being one of the early notable ones. You can imagine that much of this technology has been applied to natural language processing but the genome has, you know, 4 letters in 3 billion, a 3 billion long sequence, so it presents new challenges. We use the NVIDIA Cambridge-1 supercomputer. We trained a collection of large language genomic models and the highest performing model called Nucleotide Transformer. It achieves state-of-the-art on not only 1 of the benchmarks, but on 15 out of 18. That means that this large foundation model is able to generalize across many, many tasks which otherwise were built model at a time, quite narrowly. The paper is really, really informative, and it shows you that multi-species data was super important to be good at generalizing across these tasks.
It was also very important in showing that you need larger and larger language model sizes. That's why NVIDIA is here, so we can enable that to happen. The highest performing model it ranged from 500 million parameters to 2.5 billion. The 2.5 billion parameter model absolutely won out. This will be published in a paper in the coming hours, you know, as arXiv chunks through it. We will also make a fraction of these models readily available in the coming weeks. This is exactly why we created NVIDIA BioNeMo Service. Announced in September of last year at our GPU Technology Conference. To enable the over $200 billion of R&D market and drug discovery, how can we give them access to the tools, the frameworks, the applications at data center scale to accelerate drug discovery?
BioNeMo is making it easier and more efficient to build and use generative AI and large language models across every stage of that drug discovery process I just enumerated. BioNeMo is in early access, and we're working across the entire ecosystem with the leaders in biology and drug discovery field from research to the tech bios all the way through to the large pharmaceutical companies. Today, we are also announcing with our partner Evozyne that we've built a generative AI model for protein engineering called ProT-VAE. It's a protein transformer variational autoencoder. As you know, proteins are the building blocks of life. Every cell contains proteins. They're present in our everyday life, from clothes we wear, food we eat, air we breathe.
The field of protein engineering looks to discover new proteins that can help design more effective drugs and remove carbon from the air or make more environmentally friendly clothes. However, the number of potential proteins, you know, far exceeds the number of particles in the universe. Evozyne was able to use in a few short weeks, be able to use NVIDIA BioNeMo, use our pre-trained models, inject their special sauce with the V-VAE encoder, and train a model and be able to, from sequence, generate proteins that they were able to experimentally synthesize and validate in the lab. The beauty about this is they use BioNeMo to fine-tune that model for families of proteins. You can fine-tune it for a set of proteins that has the given properties, function, and characteristics that you want, and then it can generate a library of those.
I wanna, again, give you a little flavor for what that looks like. Here's an example of a protein called human PAH. It's a protein responsible for, you know, a precursor to making pigments, hormones, also neurotransmitters, and it can even cause some pretty rare disease disorders. The Prot-VAE model, it was trained on a PAH protein family. Here we sampled two of the proteins generated and validated. The first protein generated had 51 mutations, which is about 85% similarity to the original, and it has a 2.5 times, a huge 2.5 times enhancement in function. This is what exactly you want when you're developing new therapies. ProT-VAE also generated another protein with 167 mutations, so only 50% similar. This is where no human would, you know, intellectually go, but was able to still achieve enhanced function.
The protein language model, ProT-VAE, it generated proteins nature's never seen before, but it was able to maintain the function for the desired properties that they were measuring for. This is the promise of these large language models, the ability to explore way outside the space. We're going to extend what is today, you know, the common use of directed evolution and extend it into machine-guided directed evolution. It's really an accelerator to be able to discover new proteins. I'll conclude with NVIDIA's building the compute platforms to address the breadth of health, healthcare, to benefit from extraordinary capabilities enabled by generative AI and accelerated computing. MONAI for imaging, Parabricks for genomics, BioNeMo for drug discovery, are helping the industry, you know, harness the computation and massive accelerate R&D and the workflows that drive healthcare and life sciences. The time for generative AI in biology is now.
An explosion of seminal work happened in just the last three months of last year. I've exemplified them here with these models, and they are elucidating biological meaning. This is gonna help us understand disease, and it will accelerate our ability to discover new therapies. We can build representations of large and complex data sets now and make meaningful predictions. With that, I hope I leave you with a very bright and sunny future of what 2023 holds to really harness this technology, and I look forward to answering your questions. Thank you.
Great presentation. Are there any questions out there? I would just ask if you wait for the mic. Got one back there.
NVIDIA's revenues seem to have peaked in April of last year and have declined quarterly since. Can you explain that and why you think it's temporary?
I cannot. I am not in our investor relations team, so I will politely decline that answer.
Thanks.
Questions? I've got one right over here in the middle.
Yeah. Can you just talk about the competitive landscape, your customers, what other alternatives they have? How much is done internally, how much they need to go outside for AI solutions?
Yeah, I think the competitive landscape in artificial intelligence is, you know, we are, we're in a unique position at NVIDIA. We're the only AI company that can also work with every other AI company. Our platform is well-adopted in every single public cloud. It's adopted by every single computer maker on the planet. It ranges in sizes from embedded computers all the way through to cloud platforms. If you look at that sort of feature description of what we can offer the market, it's a ubiquitous computing platform. It allows us to give the application developers, the industries, the ability to transform business models and run these exciting new application workloads completely at scale. As you know, we dedicated the company to artificial intelligence, going back some 10 years ago. Yeah, at least 10 years ago.
The description language I described called CUDA, has been around even longer, 15 years. That's why I really describe NVIDIA not as a chip company, but we are an accelerated data center company. To compete at level is very, very difficult. The software investment that we've made and our ecosystem partners have made, is really what differentiates us and allows us to innovate continuously at the speed of light because we're a full stack company.
Questions? I've got a question. Maybe this will answer the gentleman's first question. Obviously, the consumer part of your business was weaker last year because of the weaker macroeconomic trends, some of the lockdowns in China. The data center business, which is where your franchise belongs, that's, I think that grew 40%, 50% last year, and it's still going at a very strong sort of double-digits year-over-year clip. In fact, the data center business for NVIDIA has grown at a 70% CAGR over the past 3 years. It's 60% of your total revenues. Underneath that, accelerated compute spending within healthcare continues to go at an extremely rapid rate as you sort of reflected in your presentation.
Can you just give us a sense snapshot of the revenue scale, the growth rate for your healthcare franchise over the past few years, and how do you see the same opportunity for healthcare over the next three-five years?
Yeah. I mean, healthcare is an extremely important industry, and it's quite broad. I hope I was able to let you know that we're focused on some very important large markets within it, whether that be genomics, medical devices, the biopharma industry. So we believe that we could be the next $1 billion industry for NVIDIA. You know, you can a s I described, there's these two computing platforms that are needed for the future of how we're gonna innovate, develop software and as a service. One proxy you could use, Harlan, is look at the growth in AI research papers. It's all being done on that development platform. You know, in the last two years, 120% growth. We're addressing the ability to deploy these applications.
They're maturing now, they're running in a data center, but now they also wanna run at the edge because there's a lot of incredible value you can provide the industry when you're developing real-time insights at the edge, and then also have a flexible computing platform to continue that analysis and development back in the data center in the cloud. That's kind of the proxy that we use, and I think 120% growth sounds great.
Yep, absolutely. Questions? Right there.
Hi. As a follow-up sorry, I've lost my voice. As a follow-on from that perspective around this potentially becoming a billion-dollar industry for the data center, sort of value proposition that NVIDIA can underpin, which of the traditional healthcare stakeholders need to enable or buy into the value chain of your end customer? Do you need reimbursement or regulatory or sort of government buy-in to the innovative products that get built on the stack for you to realize the $1 billion at your end?
Yeah. I think I don't know if that's gonna be our rate limiter at the moment. I think as I sort of alluded to, these large language models that we just described that were built in just the last three months of last year, they take an entire supercomputer to train them. Just to be able to explore this new capability in generative AI and large language models is a data center in itself for said customers. The pharmaceutical industry, you can hear it all over the floor here at JP Morgan. They're building centers of excellence in AI. They're using all of the massive datasets that they've acquired over the last decades of drug development.
There was $40 billion of investment that was pumped into what I, what the term I love, which is tech bios, technology first biology companies, because we're able now to generate so much biological data to really drive this model development. I think purely from that realm, we can reach that potential $1 billion opportunity amongst many of the other opportunities. You know, the analysis of genomics, Generative AI and large language models is just a clear, easy one to understand.
How much of your healthcare business today comes from genomics versus radiology and robotic surgery and more conventional healthcare segments?
You know, I don't think I don't know if I would be that granular about it. It used to be, call it, you know, sort of like an 80/20 split where we would be so in the medical devices. That kind of group what you described, we would be very medical device-centric. That's where we started. That's where our heritage lies. Going inside of CT scanners, inside the genomic sequencing instruments, you know, powering the robotic surgery platforms. All of that was kind of, call it, 80%, and 20% was kind of coming from, you know, stimulation that's going on in the pharmaceutical industry, the genomic analysis. We're definitely now coming into more of a 50/50 because of some of the dynamics that I just described of massive amount of computing that is being that is happening in pharmaceuticals.
The balance is there, but they're both growing at, you know, the sort of indicated pace of what you're seeing in the AI research.
Just a quick follow-up. Did you say what your healthcare revenue is today? You said during this.
I did not.
Okay. You don't release that?
Questions?
Thank you for a great presentation. A huge fan of NVIDIA. We buy a lot of your stuff as a, as an ADD startup. I see that recently the tech bio market has kind of declined. Some companies have, of course, grown dramatically, hopefully like us, but some have declined. How do you see the split now between big pharma and tech bio going forward in terms of the GPU sales and also cloud sales? Because many pharmaceutical companies are still kind of in a little bit of an Intel world, but now, of course, everybody's switching to NVIDIA. How do you see that split?
Yeah. I don't, I don't wanna predict the market, Alex. As a platform company, we love to raise all boats and support the ecosystem the best way that we can. Your fantastic partner at Insilico paved the way in showing that generative AI can be used across the complete drug discovery process, coming to GTC for the last 10 years. You know, one, you have to be a believer. I think pharma is still trying to figure out, are we a believer in this technology? They're doing a lot of dipping into it, and they're also kind of waiting and watching and partnering with the tech bio company. I don't, I don't wanna make a prediction. I think the industry and the number of therapies and the personalization of medicine, there's room for everybody if it were up to me.
I don't, you know, I think that there's a tremendous amount of promise. I think there is a need to think differently in this day and age with the tools that we have, with the automation and digital biology, that, you know, we need to see what the future holds.
Thank you. I actually think that $1 billion is probably lowballing it. I think that we are already at $1 billion.
Yeah, within the next five years.
Thank you, Kimberly for great speak. Aspiration for the future. Yes, I came from VinBrain in Vietnam, and it's very of we bet on NVIDIA as a platform for training system as well as for production. Everything we move from CPU on to GPU now. I have two questions for you. One is, what do you think about ChatGPT that will be influenced into healthcare from your point of view? Second is the generative AI, like, you know, sequential maturity of that, what do you foresee about? Thank you.
Yeah. Thank you. Yeah. ChatGPT is just it's an amazing technology, and it's You got to ask yourself, it's just about what do you wanna use it for? I kind of think about it right now, and just because in my own little world, back when we started work on AI and radiology, everybody jumped to the conclusion that AI and radiology has to diagnose somebody. No, it doesn't. It has a massive amount of utility throughout the entire workflow and process of radiology. It's really about what do you want to do with the AI? Do you want it to write your essay? Did I want it to write my presentation? No, I did not. You know, I might wanna do it to learn about how an enzyme react. You know, you can ask it some interesting questions like that. I think it will.
I think the type of technology will have and should have utility in healthcare. I'm not saying that it does today, but I think it will. Its utility will be different than maybe where some people wanna, you know, leap to. It doesn't have to be that far of writing up a medical report and discharging a patient. Then, you know, for generative AI, I think there's been, you know, Alex, who's just here, already has proof points how you're able to now discover novel targets and therapies that are in human trials. That is well underway and being proven. If that isn't promise enough, you know, where you have this large body of biologists and scientists who may not come from the computer science world, what I love about generative AI and these language models is we can now interrogate them in an interpretable way.
We can now learn together that the computer science community and the science community can learn this together and really start to unlock biological meaning. We know we don't know enough about the genome. These models, my hope is that they're gonna help elucidate a lot of that. Generative AI like DALL-E, where you can go from one domain to another, from text to an image, how could that not have massive utility in going through the different types of datasets in healthcare, from health records to genetic marker to what they saw in your image to what your pathology report is looking at? You know, with the likes of Genomics England and UK Biobank, you know, we're looking at can we harness these new big datasets to see if there's a DALL-E in healthcare that has utility. It's not a panacea. It's a utility. It's a tool, right?
The whole point is can we accelerate? Can we increase our understanding, and really, move this field along and find more therapies for patients who need them. Thanks for the question.
Questions? I'm going to put my semiconductor cap back on because your chips and your hardware systems are the foundational building blocks, right? You guys unlock all the innovation via your vertical markets with the platforms. The team just rolled out its next generation compute acceleration platform, the H100, that's based on the Hopper architecture, 4-nanometer leading-edge manufacturing technology, 80 billion transistors on a single piece of silicon, one of the largest chip designs in the world. Typically, the team focuses a lot of its efforts on cloud and hyperscalers at their early stages and followed by enterprise and vertical segments like healthcare. Talk to us about the adoption curve for the team's prior generation A100 platform and how do you see the uptake and momentum for the next generation H100 that the team is rolling out now?
Yeah. I think it's, I think it's gonna be incredible. We've already started early access and early work in the area of genomics. As I said, right? There's an insatiable demand for compute to deal with the amount of data that's coming off those instruments, and the more throughput, the lower the cost. We already have a tremendous pickup in that. What I'll say is, you know, in the past, you had to wait for the technology to become available. We're moving at speed of light, and throughout the first quarter of Q1, you're gonna see H100 popping up in all of the public clouds. Tens of thousands of them. We're very excited that it's gonna be readily available, so these workloads can automatically take advantage of it.
With some of the incredible features that it presents from a perspective of these large language models, it has a transformer core in it for these transformer large language models. It has INT8 to really squeeze down the efficiency and take these large language models, which would other be very expensive to inference and be able to do that very cost effectively. I think it's gonna be a rapid migration to H100, the market is prepared and ready for it.
Perfect. We're just about out of time. Kimberly, thank you for your participation.
Look forward to having the team back next year.
Thank you so much for having us. Yeah.
Look forward to monitoring the progress and execution of the team this year.
Thank you all for coming. Appreciate it.
Thank you.