NVIDIA & Lilly: The AI Revolution in Drug Discovery

Concepts and Ideas

Core Framework

Accelerated Computing as a Force Multiplier

NVIDIA's central thesis is that by co-designing the problem, the algorithms, and the hardware together from first principles, you can compress years of progress into years of engineering rather than decades of incremental improvement. The result is a millionfold improvement in AI throughput over a decade -- 10,000x beyond what Moore's Law would predict. This level of compounding is what makes problems previously considered intractable, like simulating protein folding or drug-target interactions, suddenly within range.

Computer-Aided Drug Design (CADD) -- the Biology CAD Moment

Jensen Huang draws a direct analogy to the 1980s arrival of computer-aided chip design, which transformed semiconductor engineering from an empirical craft into a deterministic discipline where 15,000 engineers can build a 220-trillion-transistor system expected to work on its first tape-out. Huang argues that biology is now at an equivalent inflection point -- where sufficient compute and model quality finally allow in silico simulation of molecular and cellular behavior with enough fidelity to guide real-world synthesis decisions.

The Synthetic Data Flywheel

The most powerful concept in the partnership is the closed-loop data generation system: AI generates candidate proteins or molecules, those are synthesized and tested in robotic wet labs, the experimental results are fed back into the model, and the cycle repeats -- continuously. Eventually the lab itself becomes a model, and the primary model can generate synthetic data to fill in the experimental space. The flywheel only works when the models are grounded periodically by physical-world truth, but the ratio of real experiments to AI-generated candidates can shift dramatically in favor of the machine.

Drug Discovery Principles

Inventing Inside the Patent Cycle

Pharma companies that cannot invent fast enough to replace revenues before patent cliffs hit are forced to rescale every 10 to 15 years -- shedding talent and institutional knowledge in the process. Lilly's core strategic response was to compress the R&D timeline by systematically eliminating friction at every step of the candidate-to-approval process, making it possible to bring new drugs to market while existing ones still generate revenue. This is the structural condition required for compound growth -- not just better ideas, but faster execution on ideas of any quality.

Target Discovery vs. Drug Engineering

Drug development has two distinct bottlenecks. The first is drug engineering -- taking a known target and optimizing the molecular "key" to fit it better, with higher efficacy and lower toxicity. The second is target discovery -- identifying the biological locks worth designing keys for in the first place. AI is already accelerating the engineering side. The harder and more valuable problem is using AI and robotics to discover and fully profile new targets, which requires massive experimental data generation rather than pattern recognition alone.

GLP-1 Incretin Biology

GLP-1 (glucagon-like peptide-1) and GIP are hormones produced by the gut that signal the body it has been fed, triggering insulin release and appetite suppression. The native human GLP-1 has a half-life of about 7 minutes, making it clinically useless until Lilly and others engineered extended-release versions -- first twice-daily, then once-weekly, with monthly dosing now in development. Tirzepatide (Zepbound) fuses both GIP and GLP-1, producing synergistic effects on weight loss and metabolic health that exceed either peptide alone.

Obesity as a Systemic Disease Driver

Obesity is not framed here as a cosmetic condition but as a chronic inflammation amplifier that drives or worsens more than 200 downstream diseases including cardiovascular disease, type 2 diabetes, joint degeneration, and potentially dementia. The evolutionary basis is maladaptation -- human physiology evolved for caloric scarcity and has no built-in off switch for abundance. GLP-1 drugs create that switch synthetically, and the downstream cascade of chronic disease reduction is the true market opportunity.

Infrastructure and Platform Thinking

Build the Instrument Before You Have the Scientists

Jensen Huang's account of building NVIDIA's internal supercomputer before having enough AI researchers to use it is a case study in conditions-first investment. The logic is that world-class scientists need world-class instruments, and if you wait until the researchers are hired to build the infrastructure, you will never attract them. This principle applies directly to the NVIDIA-Lilly co-innovation lab and to the biology supercomputer in Indianapolis.

Federated Learning and the Tune Lab Model

Lilly's Tune Lab platform allows multiple biotech companies to jointly train AI models without aggregating or commingling their proprietary datasets -- using federated learning infrastructure originally built on NVIDIA's NVFlare framework. Each participant's data improves the shared model, and each participant benefits from the improved model, without revealing their underlying data. The outputs from the NVIDIA-Lilly co-innovation lab will be contributed to Tune Lab, making the research benefit extend beyond the two companies.

BioNeMo as the Biological AI Stack

BioNeMo is NVIDIA's platform for training, fine-tuning, and deploying foundation models that understand geometric biological structures -- proteins, molecules, RNA, and DNA. It includes pre-trained models such as Llama Protina for protein design, Evo Two as a DNA foundation model, and Codon FM for RNA design. The platform sits above NVIDIA's accelerated computing infrastructure and below domain-specific applications, functioning as the biological equivalent of an operating system for AI-driven life sciences.

Frontier and Emerging Areas

Addiction and Self-Learned Behavior Loops

Appetite, it turns out, is driven less by actual caloric need than by self-reinforcing learned behavior loops -- a structure that closely resembles addictive behavior. GLP-1 drugs appear to down-regulate these loops more broadly, with early evidence suggesting efficacy against alcohol consumption, smoking, gambling, and opioid addiction. Lilly is running dedicated studies on a GLP-1/GIP formulation specifically optimized for brain conditions, which could represent an entirely new class of addiction treatment.

Dementia as the Next Disease Frontier

As cardiovascular disease, diabetes, and infectious disease mortality have been progressively reduced, dementia and other diseases of the aging brain have emerged as the primary limiting factor in quality of life for people living into their 80s and 90s. Lilly is investing heavily in programs targeting the intersection of chronic inflammation and protein misfolding in the aging brain -- and sees AI-assisted simulation of brain tissue and target interaction as essential to finding interventions in a system too complex to drug empirically.

Implementation

Implementation steps are auto-generated from the transcript content and are provided for informational purposes only. They do not constitute professional advice of any kind. Always consult a qualified professional before acting on any information presented here.

Map Your R&D Process End to End

Lilly's 40% speed advantage came from taking apart every step between candidate selection and first approval and compressing each one systematically -- not by finding better ideas, but by eliminating friction in the process of developing any idea. For any organization running an innovation cycle, the first move is to make the full process visible: name every stage, measure cycle time, and identify which steps are waiting time versus value-adding time. The ones you can automate or parallelize with AI tools are the highest-leverage interventions.

Build the Instrument Before You Have the Team

Jensen Huang's most counterintuitive insight is that world-class talent follows world-class infrastructure, not the other way around. If you are trying to attract researchers, analysts, or engineers who operate at a high level, the constraint is usually the tool environment -- not the salary. Before hiring, invest in the platform, the data access, the compute environment, or the lab capability that makes elite work possible. Build it first; the people who need it will come.

Close the Loop Between Generation and Ground Truth

The synthetic data flywheel only works if it is periodically grounded in physical-world reality. Any AI-assisted discovery or prediction system -- whether in biology, in business, or in operational forecasting -- needs a regular feedback mechanism that imports actual outcomes back into the model. Design your workflow so that AI-generated hypotheses are regularly tested against real data, and that real data is systematically imported back into the training pipeline. The flywheel degrades if you treat the model as an oracle rather than as a closed-loop instrument.

Distinguish Engineering Problems from Discovery Problems

Ricks draws a sharp line between drug engineering (optimizing the key for a known lock) and target discovery (finding the right locks to design keys for). Most AI applications in life sciences are accelerating the engineering side, where the combinatorial space is large but the fitness function is well-defined. The discovery side is harder and more valuable. For any domain, it is worth asking which of your open questions are engineering problems -- where AI can systematically search -- and which are genuine discovery problems that require new experimental data and new hypotheses.

Think in Flywheels, Not Projects

The NVIDIA-Lilly model is not a one-time research initiative -- it is a self-reinforcing system where each cycle of generation, testing, and model refinement makes the next cycle faster and cheaper. Project thinking produces a deliverable; flywheel thinking produces compounding capability. When scoping an AI initiative, ask whether the outputs of the work improve the inputs for the next round of work. If not, redesign the system so that generated data, labeled outcomes, or validated hypotheses are fed back into the model at each iteration.

Co-Design Across Domain Layers

NVIDIA's accelerated computing philosophy is built on co-designing the problem, the algorithm, and the hardware together -- solving from the top down and the bottom up simultaneously. The NVIDIA-Lilly partnership instantiates this principle at a company level: bringing together hardware infrastructure, AI platform layers, foundation models, domain-specific biological expertise, and physical lab systems. For any cross-disciplinary initiative, identify which capability layers are missing and whether you need to build them, acquire them, or partner to access them. Partial stacks produce partial results.

Treat Federated Collaboration as a Competitive Advantage

Lilly's Tune Lab model shows that it is possible to improve a shared model using proprietary data without exposing that data -- through federated learning architectures. This is relevant beyond pharma: any industry where competitive sensitivity prevents data pooling should explore federated approaches. The parties who contribute to a federated model benefit from its improvement, and the more participants, the better the model -- turning a coordination problem into a structural advantage for those willing to engage early.

Anchor the Future Case to a Historical Precedent

Ricks' most persuasive move in the conversation is tying GLP-1's potential to the antibiotic precedent -- a class of drugs that ran for 60 years and made modern surgery and dentistry possible. When communicating a large, uncertain vision, grounding it in a historical parallel that the audience already accepts as true lowers the cognitive burden of belief. For anyone building a case for a transformative technology or initiative, the question to ask is: what is the best historical analogy that makes the claim feel real rather than speculative?

Full Transcript

This transcript was auto-generated and may contain errors in speaker attribution, transcription accuracy, or formatting. Long transcripts may be truncated due to processing limits. Confirm accuracy and completeness against the original source before referencing or republishing.

[00:03]

[Applause]

Jensen Huang (NVIDIA)

Hello. It's great to see all of you. You know, once a year I get the benefit of thanking all of you for the incredible work that you do. As you know, we pioneered a new way of doing computing called accelerated computing. And we've been working on this for now 33 years. The big idea is that if you could co-design -- meaning if you understood the problems you're trying to compute for and you develop the entire stack, meaning the algorithms, the computers and the processors within it -- that if you co-designed you solve the problem from top down, you solve the problem from bottom up and inside out, that you could accelerate these applications, which are incredibly computing intensive, by a level that Moore's law would only dream of. Moore's law runs at about 10 times every 5 years, 100 times every 10 years. In the last 10 years we've accelerated AI about a million times. A million times versus 100 times. A million times compounded over 20 years gives us an opportunity to maybe address some of the most incredible, most impactful challenges of humanity. That's you.

And so, we've been thinking about accelerated computing and this co-design idea for a very long time. After some 30 years or so, each different field of computing domains became within reach. One of the first ones was computer graphics. And somebody told me about 10 years ago that it would take 30 years for us to solve this incredibly hard problem called ray tracing for computer graphics -- literally following every photon of light as it bounces around the room, eventually reaching your eyes. Well, that was 10 years ago. And literally today we ray trace every single pixel in a video game at several hundred frames a second at any resolution. We've taught an AI what an image should look like and it generates the image for us -- temporally consistent, spatially consistent, incredibly beautiful.

About 10 years ago we started talking about self-driving cars. Today I think we are at the ChatGPT moment of self-driving cars. There's no question in my mind that within the next 10 years almost every car that gets manufactured will have either robotaxi capability, or if you like to drive yourself you still can. Each one of these domains are within reach. And of course, we now have the ChatGPT moment in generative AI and large language models. We made huge breakthroughs in multimodality -- so that not only do you understand the word "cat" and the letters c-a-t, you can look at an image and hear the sound "meow" and recognize all of that information aligned to the same place. We made huge breakthroughs in reasoning capabilities -- the ability to decompose a problem we've never seen before into routine components and through composition solve problems step by step. The ability to do multimodality reasoning and be grounded by different sources of information is going to be a revolution for your industry too.

[05:01]

These breakthroughs, and the agentic capabilities -- the ability for agents to reason, use tools, work with other agents -- completely revolutionary. And so these breakthroughs are permeating your industry, permeating life sciences. Some of the most exciting work I've seen in a very long time.

We announced a suite of platforms we're very proud of: Parabricks for gene sequencing, MONAI -- an open source platform for medical imaging -- and BioNeMo, a platform that understands geometric structures like molecules and proteins. MONAI has been downloaded some 6 million times. We achieved a Guinness World Record in gene sequencing. The vast majority of the world's structure prediction models are built on NVIDIA today. And on top of BioNeMo are a whole bunch of pre-trained models: Llama Protina for protein design, ReSENT for molecular synthesis, Kermit for predicting toxicity, Evo Two -- a DNA foundation model -- and Codon FM for RNA design. Really terrific progress. And today we also announced a landmark partnership with one of the most important companies in the world -- and a really good friend -- Dave Ricks at Lilly.

[10:01]

Dave has been CEO of Lilly since 2017. In the context of the history of Lilly -- a 150-year-old company -- he's been with the company for 30 years. I'm anxious to learn from him how he learned so much in just a short period of time. Please welcome my guest, David Ricks.

[Applause]

David Ricks (Eli Lilly)

Thank you. Great to be here.

Jensen Huang

So at 2017 you became CEO. What was Lilly like and what was its vision for the future of life sciences? And then what is it like today?

David Ricks

Well, in 2017 we were almost 140 years old, a pretty well-established company. We had been through a difficult period. In our industry, as many people in the room know, we are plagued by the cycles of innovation that are actually longer than the reward cycles on average. So you have these peaks and troughs, and in the troughs, of course, you either don't survive, you end up as the back end of a hyphen on another company, or you learn how to survive and you do difficult things. Lilly is the only major pharma company that has not merged or combined itself to get through such a period. We've been through a few of them. So I think we've developed a bit of a DNA for scraping by in these tough times.

One of the things that really stuck with me -- I was a leader on the executive committee during our last cycle, which was 2010, 2011, 2012 -- was how you have to be so relentless on invention because that time cycle is brutal. When you have a patent cliff and you have nothing to replace it, you have to rescale your whole company. And imagine doing this every 10 or 15 years -- it's impossible to get momentum. So we set about to attack that problem two ways. One, you can say, "Okay, we'll just have better ideas." But that's difficult to sustain. The more important thing we did was speed up invention. We systematically took apart the R&D process -- literally every step, and there's hundreds of steps from candidate selection to first approval -- and we squished the time. Still today, Lilly is about 40% faster than the next scaled pharma company on average across all therapeutic areas. So we can invent things inside the patent cycle. Then you can start to grow.

Along the way, we were lucky enough to be early on an idea called GLP-1 in 2006. We launched the first one in the world. And to their credit, our scientists kept working on it for like 12 years while it seemed like not much was happening. And our competitor innovated in obesity -- we chased and executed, I think, quite well to become the market leader last year. So we not only now have an engine that's faster, we have a tremendous success cycle with this one idea. The challenge before us now is: how do we find another success cycle before that one runs out, hopefully a lot sooner?

[15:01]

Jensen Huang

You said something that I hear myself say all the time: the fundamental purpose of the CEO is to create the conditions by which great ideas can systematically arise. Our job isn't to create ideas; our job is to create the conditions for great ideas. And what you just described is essentially Lilly building the conditions by which great inventions can happen. One of the most important things we did at NVIDIA was investing in infrastructure -- creating the conditions by which amazing AI researchers can come to our company. I built one of the largest supercomputers in the world for NVIDIA to use. At the time I had nobody to use it. But if you don't have the AI researchers, and therefore you don't have the instrument -- and if you don't have the instrument, you'll never have the AI researchers, because scientists need the instrument necessary for their science. I came to the conclusion that the right answer is to create the conditions for great AI work. And we attracted some amazing people. Today we do frontier work in physical AI, biological AI, and robotics AI.

And so that led our conversation to the idea that AI is making enough progress that we might be able to apply it to tackle some of the most extraordinary challenges of humanity: biology. And the two of us thought it would be amazing if the largest computer company in the world partnered with the world's largest life science company -- as if we created a lab where we are essentially one company combining the best things together. Which led us to our announcement today.

[Applause]

David Ricks

Today we announced four parts to this. We bought a bunch of chips, and we're building -- I think it'll be finished this week, actually -- the largest dedicated on-prem biology supercomputer in the world. That'll be in Indianapolis, which is the world's center for biological discovery. We're going to put together a research team here in the Bay Area -- a joint Lilly-NVIDIA AI lab. We're going to develop new data, because if the data is an inch thin you don't get very far, and biology is largely unknown. We need to create massive experimentation for the purpose of training machines, and then we need to build on the work NVIDIA has done with the models to ladder up to even more sophisticated predictions.

[20:00]

Jensen Huang

One of the important things to understand about AI is that it's easy to think about as a multi-layer cake. At the lowest level you need energy for the computers. Then you have the chips and the infrastructure layer -- the operating system of AI, if you like, covering how you cause the computer to learn, how you guardrail it, how you fine-tune it. That layer is kind of like the HR department of AI. Above that is the models. And above that is the domain-specific data and flywheel.

One of the most exciting areas is the idea that we would train models to synthesize proteins or chemicals, put that into a robotics lab, collect more data, take that data back into our model -- this flywheel of data generation and model improvement, using AI, collecting ground truth, experimentation, putting the scientific flywheel on steroids. That future is just incredibly exciting. So we're systematically bringing together the brightest minds in drug discovery domain expertise and the brightest minds in computer science -- and together, in our co-innovation lab, we'll have essentially every layer of that cake and hopefully a blueprint for what is possible in the future of drug discovery.

40 years ago -- 43 years ago -- I was the first generation of engineers where the concept of computer-aided design was invented. Before my generation, engineers would design a chip and it would be a hit, then they'd go several years without another hit. My generation came along and we were able to represent the transistor, the logical gates, the functionality in software. A computer could represent your domain and accurately simulate its behavior. Today we design massive chips -- 15,000 people coming together to design a computer system. We just taped out Vera Rubin: billions of dollars of R&D, 3 tons when finished, a million and a half parts, six new chips, 220 trillion transistors. And that computer is going to work first time, exactly as expected, building on every previous generation.

[25:00]

I think your time has come for that. Finally -- a million times, a quadrillion times of computing power later -- I think we might be able to represent one of the hardest problems in the world: human biology. I'm really hoping your industry moves from drug discovery as "wandering around the forest looking for truffles" to computer-aided drug design.

David Ricks

Sidebar -- we used to have a department of soil discovery. Seriously. We would send people into tropical forests, scoop up soils, bring them back to Indianapolis, and refine out what became antibiotics. In fact, the most important antibiotic in the world -- vancomycin, a last-resort antibiotic in hospitals, the one without which we couldn't run hospitals -- was discovered in Borneo in a soil sample by Lilly. Fungus produces tools to fight bacteria, so they'd dig in soil and discover these naturally occurring defense mechanisms and refine them into drugs. That was in the 1950s.

Jensen Huang

Now your job gets to be just like my job. You never have to leave your office. We do it all in silico. I really think this announcement between us could be a blueprint for something. Computer-aided drug design -- still a decade yet, but I really believe this.

David Ricks

We need it, right? If we think of all the unmet needs at a conference like JP Morgan, you're reminded of the incredible pipeline of the industry but most of it is sort of chipped out of stone -- hard work. Occasionally we harness systems, like when monoclonal antibodies came around. But mostly what we do is really empirical and difficult. Each small molecule discovery is like a work of art. If we can make that an engineering problem versus an artisanal drug-making problem -- think of the impact on human life. The first chips you worked on were quite empirical, too. No clue how they'd perform. Today you know the functionality of your systems down to every single bit. I really believe that could revolutionize the future of drug discovery.

Jensen Huang

So tell me -- GLP-1 is really in its third phase already. Where does it go from here? How could it change everything?

David Ricks

As it turns out, obesity is like a key lynchpin driver of chronic disease in adults. We evolved in a world of scarcity. There was no need for an off switch on hunger. And so that's why we have obesity, and that leads to many, many chronic diseases that shorten life. By having a pretty effective set of tools -- going to get more effective -- we can down-regulate more than 200 chronic diseases.

[30:00]

The two big evolutions in the next 24 months: first, more choices. GLP-1 is one peptide, but it's in a family. By combining two -- GIP plus GLP in tirzepatide -- we see better performance and better tolerability. Within this super family of proteins we can harness combined strengths for more tolerable, more effective drugs at different phases of treatment. Our triple-acting one is coming next year, 2027. Second, expanding use cases. We know you lose 23% of your body weight on average on Zepbound. Conversion from prediabetes to full diabetes drops by 93%. Imagine treating the 70 million Americans with prediabetes -- we'd have 93% less diabetes. That's a profound finding.

There are also non-obvious use cases. Inflammation is a primary one -- being obese causes excess chronic inflammation, which is not good for your cardiovascular system or joints. Spontaneously in our trials, people would report that they could stand up from a sitting position for the first time in 10 years, that their knees didn't ache, that people with Crohn's and colitis saw resolution of symptoms. So we're taking that on medically with use case studies. And we're also starting a number of brain health studies, with a specially purposed GLP-1/GIP formulation for brain conditions -- particularly addiction. These medicines down-regulate appetite, which turns out is not really driven by caloric need but by other self-learned behavior loops. It can do a similar thing for other harmful self-learned behavior loops, like gambling, alcohol consumption, smoking cessation. We're exploring all of these.

[35:01]

And on the horizon: much longer-acting formulations, and we'll launch our oral GLP-1 this spring -- a major breakthrough not just for convenience but for global reach. Injectable systems are some of the most complicated drugs manufactured in the world, and we have to produce them at hundreds of millions of units. We just launched in China and India and Southeast Asia. There's a lot of obesity there. We cannot possibly meet all that demand with injectable systems. An oral form -- just a pill -- we can scale massively. That will launch this year as well.

Jensen Huang

Has there ever been a life sciences technology that has benefited humanity as broadly as this?

David Ricks

The only other one is a class of drugs -- antibiotics. The discovery of penicillin invented modern medicine. If we subtracted antibiotics from our lives, we couldn't have modern dentistry, we couldn't do surgeries. Fleming self-dosing this weird mold he grew in his lab -- that was the start of a major wave that lasted into the 1990s. A 60-year cycle with 100-plus new medicines. I think this could be like that. But this is not for an acute infection -- it's for chronic disease of the modern homo sapien. That's what we're affecting.

[40:00]

Jensen Huang

Tell me about the history of GLP-1. How did this get discovered?

David Ricks

It goes back to the 1970s. A German scientist discovered what's called the incretin effect -- he observed that if you give sugar orally versus intravenously, you get a very different insulin response. That led in the 1980s to the discovery of GIP first, then GLP -- the two parts of Zepbound. The native human form of GLP has a half-life of about 7 minutes. To give that as a medicine you'd have to walk around with an infusion all day and night -- not a product. It wasn't until we figured out how to extend that half-life -- first to twice a day, then once a week, now working on monthly -- that we could make it an effective product. One thing in AI drug discovery that is really working is the ability to create massive combination possibilities and test them in silico, then filter out good ideas. We're doing that at scale now and can do it at an even bigger scale with our collaboration. The other side is discovering the lock -- can we find more biology targets using AI? That is the holy grail. If we put those two things together, we can model the whole system at once.

[45:01]

Jensen Huang

GLP-1 beyond obesity -- what about longevity and aging?

David Ricks

Longevity as in extending the maximum shelf life of a human is not really a problem we work on directly. There's a lot of biology on why individual cells expire, and you can extrapolate that to organs and the whole organism. But I think it would be a weird world if people lived forever -- why would we reproduce? How do we evolve? We're more focused on a more practical problem: if the maximum shelf life is 100 years, how do we get more people a better shot at that full life? The way we do that is by knocking out disease.

Life expectancy in 1900 was about 46 years for a male -- the biggest factor being child death from infection. Antibiotics and vaccination changed that quickly. From 1960 till now we've added about 10 years mostly through chronic disease extension. HIV/AIDS conquered, cancer care improved dramatically. We're up to almost 80 years now -- it would be 84 if not for opioids and traffic deaths, which are exceptionally high in the US. Self-driving cars can help with one of those. GLP-1s might help with opioid addiction, which we're studying.

The next frontier beyond GLP I'm really excited about is dementia and diseases of the aging brain. As the rest of our bodies carry us into our 80s and 90s, this is something we haven't figured out how to address. There's inflammation plus protein misfolding happening in the aging brain. Perhaps we can discover something really exciting. AI-assisted simulation of brain tissue and how to look for new targets and the interaction between targets we don't understand -- that's the work. We're spending a lot of effort there and have some exciting programs running in studies now.

[50:02]

Jensen Huang

If you had a direction you'd like to set the co-innovation lab researchers on, what would it be?

David Ricks

There are two basic problems. The first has a lot of momentum already: drug design optimization -- engineering drug designs using AI. You know the target because it's been around and well-profiled, and you're customizing the key to the lock. New modalities like RNA and gene therapies really haven't been optimized because they haven't been around as long. I think that whole suite of drug engineering and optimization problems is pretty tractable, and the NVIDIA scientists coupled with ours could make a lot of progress quickly. That solution would also be highly amenable to our Tune Lab interface for our biotech partners -- better drug making through in silico design.

The other side is targets. We need our robotics effort to really kick in. Why do we have more targets, and why do we understand the targets we have better? A lot of drug development fails because we think we have the target but we don't really understand it -- we need massively more experimental data. We'll build more wet labs running 24/7 experimenting around a target space to really fully profile it, then couple that with engineered drug designs. Machines are made to work day and night to solve this problem, and that flywheel is going to fly.

Jensen Huang

That synthetic data flywheel is incredibly effective. You eventually have a model testing against another model -- a world model or target model. Eventually you have to ground it again by synthesizing the proteins and collecting real data. But you get yourself in a synthetic data flywheel that dramatically accelerates the pace of discovery. I think it's singularly true that finally we're going to have a lab where the expertise and scale is sufficient to attract people who really want to do their life's work at that intersection.

[55:01]

David Ricks

There are a lot of interesting companies doing great work right now -- robotics labs, AI scientists companies, agentic healthcare assistants. The work that Abridge and Open Evidence are doing is really cool. At the healthcare service level, the most tractable problem is replacing human-level service with agentic AI service. It seems so obvious for what is arguably the least productive part of the economy to adopt AI, but there are a lot of frictions. Insurance companies don't welcome change. Physician groups are frightened by some of these technologies. But we really do need to make the US healthcare system more efficient, and in doing so we can have more room for innovation and adoption of cutting-edge medicines.

Lilly Direct -- our direct pharmacy platform, only a year old -- is annualizing close to a billion dollars a quarter right now, 4 billion a year. And there's a lot of interest in how Lilly can help force adoption of AI in healthcare services at scale. On the drug side, new modalities in the last 10 years -- bifunctional medicines, targeted warheads for cancer, novel drug delivery -- are accelerating, and AI can help accelerate that further. Lilly has a reputation 10 years ago of being a sleepy Midwestern company. But fun fact: the first approved biotech product made by a living organism -- Humulin, made in collaboration with Genentech -- was a Lilly product. And 12 years later we bought the first supercomputer in the pharmaceutical industry: a Cray, called Big Red, in Indianapolis. We designed the first approved medicine designed on a computer -- insulin lispro, Humalog -- on that machine. So this partnership is a natural evolution of who we are.

Jensen Huang

I am so excited about the work we're about to do together. I can't imagine a more worthy, a more complex, a more exciting field to apply computer science to. And for someone who's dedicated a whole career building computers, the opportunity to work with you and Lilly and the scientists there in service of the next breakthrough can't be more exciting. I want to thank you for your partnership and for personally driving this. Hopefully the two of us teaming up can bend the arc of history and make a difference.

David Ricks

It's exciting to hear you say that. Thank you for helping make this come together. We're just getting started, but this could be a new day. If we can accelerate progress -- even with modest goals -- in a use case for computer technology that has to be the most personal and impactful for our country and the world, and for an industry that really needs to find more ways to be more productive with R&D dollars, it's a worthy endeavor. We're going to do more and win.

[Applause]

AI Master Prompt

The AI prompt on this page is auto-generated from the transcript content and is intended to support further exploration of the topics, concepts, and conclusions discussed. It is provided for informational purposes only. The user is solely responsible for all outcomes resulting from its use.

Master Prompt

You are a research partner and thinking companion helping me explore and apply the ideas from a conversation between Jensen Huang (CEO, NVIDIA) and David Ricks (CEO, Eli Lilly), recorded at the JP Morgan Healthcare Conference in February 2026. The conversation centered on a landmark partnership between the two companies to build the world's largest dedicated biology supercomputer, a joint AI co-innovation lab in the San Francisco Bay Area, and a shared research agenda targeting drug discovery, protein modeling, robotic wet labs, and closed-loop synthetic data flywheels. The central argument is that accelerated computing has compounded AI performance by roughly one million times over the past decade -- far beyond Moore's Law -- and that this scale now makes computational simulation of human biology a realistic engineering discipline rather than a speculative ambition. Jensen Huang draws a direct analogy to the arrival of computer-aided chip design in the early 1980s, which transformed semiconductor engineering from an empirical craft into a deterministic process where 15,000 engineers can design a 220-trillion-transistor system expected to work on its first tape-out. Huang argues biology is now at an equivalent inflection point. David Ricks frames Lilly's strategic posture around two principles: first, systematically compressing R&D timelines (Lilly is now approximately 40% faster than its nearest competitor), and second, creating the conditions for great inventions rather than trying to prescribe which inventions to pursue. He positions GLP-1/GIP combination drugs like Zepbound as potentially the most systemically impactful class of medicine since antibiotics -- capable of down-regulating more than 200 chronic diseases driven by obesity and chronic inflammation. The next major frontier for Lilly, beyond GLP, is dementia and diseases of the aging brain. The partnership model introduced includes four elements: a massive on-premises biology supercomputer in Indianapolis; a joint research team in the Bay Area staffed by NVIDIA AI researchers and Lilly biologists; a massive experimental data generation program to train machine learning models at scale; and Lilly's federated learning platform, Tune Lab, which allows multiple biotech partners to jointly train models without commingling proprietary data. The most technically interesting concept is the synthetic data flywheel: AI generates candidate proteins or molecules, those are synthesized and tested in robotic wet labs running 24/7, experimental results are fed back into the model, and the cycle repeats. Eventually the primary model can generate synthetic data to fill in the experimental space, with real-world experiments grounding the model periodically. The two leaders argue this is how drug discovery becomes an engineering discipline -- not by prescribing which drugs to make, but by systematically closing the loop between prediction and ground truth. Key principles to work with: -- Conditions before capability: build the instrument before you have the scientists who will use it; talent follows infrastructure -- Compressing innovation cycles: the strategic moat is speed of invention, not quality of any single idea -- Closed-loop flywheels: AI-generated predictions grounded by physical experimentation, iterated continuously -- Federated collaboration: multiple parties can improve a shared model without exposing proprietary data -- Target discovery vs. drug engineering: two distinct bottlenecks in drug development that require different approaches -- Obesity as a systemic driver: GLP-1 interventions work through self-learned behavior loops, not just caloric regulation, opening addiction and brain health applications -- CAD analogy: biology is approaching the same inflection that made chip design predictable and scalable What this is not: This is not a conversation about replacing human scientists or automating discovery in a simple sense. Both leaders are explicit that biology remains extremely complex, that the physical-world grounding step is non-negotiable, and that the co-innovation lab is designed specifically to attract and support human researchers working at the intersection of computer science and life sciences. It is also not a conversation about extending maximum human lifespan -- Ricks explicitly frames the goal as getting more people to the full life expectancy already biologically available, not pushing beyond it. How to use this chat: 1. CLARIFY -- If I ask about a concept from the transcript (GLP-1 incretin biology, federated learning, BioNeMo, synthetic data flywheels, computer-aided drug design), explain it clearly in plain language and ground your explanation in what was actually said in the conversation. 2. APPLY -- If I describe a problem, a business context, or an industry challenge, help me apply the frameworks from this conversation to that situation. Translate the principles -- conditions before capability, flywheel design, timeline compression -- into specific, actionable thinking for my context. 3. EXPLORE CONNECTIONS -- Help me draw connections between what was said here and adjacent ideas in AI, drug development, platform strategy, or industrial history. If I ask how this connects to something else, bring in what you know beyond the transcript. 4. CHALLENGE ASSUMPTIONS -- If I appear to be misreading the scale of what was claimed, conflating drug engineering with target discovery, or underestimating the physical-world grounding requirement in a data flywheel, flag it and help me think more precisely. 5. SCENARIO BUILD -- If I want to think through implications -- for investors, for biotech startups, for public health systems, for AI researchers -- help me think through second-order consequences of this partnership model and the technologies described. 6. PROMPT REFINEMENT -- If I want to take ideas from this conversation into another AI context (writing, research, modeling), help me construct a precise prompt that captures the right framing. Tone: Grounded, direct, technically honest. Do not oversell the implications or treat the partnership as a fait accompli -- both leaders acknowledged this is a decade-level project with genuine uncertainty. Help me think clearly, not be inspired vaguely. To get started, I have a few questions for you -- answer them one at a time: What is the specific problem in drug discovery that the synthetic data flywheel is designed to solve, and what does "grounding" the model in physical reality actually mean in practice?

Creator	NVIDIA
Speakers	Jensen Huang (CEO, NVIDIA) & David Ricks (CEO, Eli Lilly)
Source	youtu.be/zbiEYMapsvw
Published	2026-02-02
Transcript	2026-06-07
Context	JP Morgan Healthcare Conference -- fireside chat announcing a landmark NVIDIA and Lilly partnership