and welcome to a special GTC edition of the NVIDIA AI podcast.
This is episode four of five on the road to GTC Live in Washington, D.C.
Bonus conversations about the state of AI you won't hear anywhere else.
This episode is all about science.
In laboratories and research centers around the world.
AI is becoming a core instrument of discovery.
Listen in as scientists and technologists explore how computation is accelerating progress across fields.
Enjoy the conversation and remember.
The NVIDIA AI podcast brings you new interviews with leaders across research business, the public sector and more every week.
Listen and subscribe wherever you get podcasts.
So science used to move at the speed of experiments.
Now it moves at the speed of compute, crunching data as fast as we can collect it, modeling everything from the atom to the atmosphere.
Then this just speeds up outcomes.
No doubt.
And as AI begins to supercharge quantum computing, we're on the verge of discoveries that could redefine physics, chemistry and life itself.
Quantum computing is at the heart of the fastest acceleration that should unlock scientific discoveries.
From molecules to the cosmos, AI is transforming how science models the world.
And leading this conversation, we have George Church, Chief Scientist at Lila Sciences, Matt Kinzella, CEO at Inflection, Marc Tessier-Levine, co-founder, chairman, and CEO of Zyra Therapeutics, and Anirudh Devgan, president and CEO of Cadence.
All right, Matt, let's start with you here.
Big day, big quantum day.
Can you talk about how the mixture of we'll call it classical computing and quantum computing is really going to change the game?
Yes, I can talk about that.
That's an easy one.
First of all, you know, when we...
First of all, thanks for having us.
This is going to be a blast, guys.
This is going to be a lot of fun.
When we say quantum, maybe we should just define some terms, because not everyone in the audience might know what that means.
And so when we say quantum, we're talking about the world of the very small, the atomic and the subatomic levels.
And there's a whole different set of rules that govern the data on there called quantum mechanics.
And so when someone says quantum, it's taking advantage of those very strange quantum mechanical properties.
What?
Hi.
What's going on here?
Jensen Wong, everyone.
Oh, he brought us water.
That's what he does.
Yes, we do.
Thank you.
It's very important to stay hydrated.
What is this mumbo jumbo about quantum?
Superposition.
Entanglement.
Entangling stuff.
This is just mumbo jumbo.
Hey, good to see you.
You know, once a busboy, always a busboy.
Yeah.
From Denny's to D.C.
Yeah, thanks for having us.
It's incredible.
It's been an amazing morning.
And this panel, I think, saving the best for last, Jensen.
So, you know, maybe ask this.
You caused a kerfuffle last year when you made some comments about quantum.
Anything you would like to revise.
You know, as we're about ready to launch our quantum panel here, about how NVIDIA thinks about quantum
You know, I got to tell you, I'm afraid to say a word.
If I just said quantum, the stock price goes up.
Quantum, quantum, quantum.
Buy, buy, buy.
Here, listen.
Listen, the work that we're doing together is really important, obviously.
And I think the message that I was what I was trying to say is that quantum and classical computing really needs to work together so that we could bring in the usefulness of quantum computing.
And it's becoming very clear that the two industries really need to work together as one.
And that quantum classical computing, quantum GPU computing could really help solve a lot of very sticky problems, very challenging problems associated with quantum computing.
And the ecosystem and ourselves, we're doing some amazing work.
It remains incredibly hard.
I mean, the work you guys do, it's deep science.
You know, we do deep engineering, they do deep science.
And so I'm excited about the work that we're doing together.
I won't ruin it by telling you guys what it is, but this doesn't work.
It's okay.
No, no, this is about you guys.
This is about you guys.
And apparently I want to thank everybody for coming to GTC in Washington DC.
So the original plan I'll be honest, the original plan was President Trump was going to be in Washington DC.
And so we brought GTC to Washington, D.C.
The administration's all going to be here.
And then literally two days ago, President Trump said, Jensen, could you be in Korea today?
And I said I would love to be in Korea today, but I came to visit you in DC, And so anyways, we're in two sides of the world.
But anyhow, he wishes all of you well.
He's going to see me in a day.
And we're going to, we're going to, we're going to go and support, support the president in this, in this tour through Asia.
You know, he's our president.
We want him to be enormously successful so that America could win and America could win.
And so anyways, I want to thank everybody who's here working in D.C.
Thank you for thank you for your service and thank you for coming to GTC.
All right, great to see you.
Thanks, Jensen.
Take care.
We can't wait to hear more.
I mean, we said last night, we are in Washington.
We've heard on the panels this morning, so maybe we'll just jump to this and then we'll come back.
Matt, to that.
But, you know, there's an extreme need to accelerate AI.
I think Jensen's been at the forefront, NVIDIA leading the charge with the president, the White House and Capitol Hill.
But we also see a lot of doomerism.
You know, you guys are in what I would call long-haul investigation, discovery, and execution.
Talk to us a little bit about the importance of having Jensen leading the charge and NVIDIA leading the charge to your businesses that require decades of innovation, and maybe how you're partnering with NVIDIA to make that happen.
Matt, we'll come back to you on that.
Absolutely.
We have looked to NVIDIA as a trailblazer of how you bring deep tech out into the market in a staged, commercialized way.
And so, as we've thought about bringing quantum technologies out into the market, we saw them point their GPU engine at graphics, to start, and then at crypto mining.
And they always found the next thing, until the crown jewel of large language models came around for them.
And so the way we've thought about commercializing quantum is by pointing quantum at areas where there is true quantum advantage already, like timekeeping, and then quantum RF sensors and antennas, and on the trip to that crown jewel of quantum computing that will do things that classical computers can't do.
And so we've really used them as what we've modeled our strategy against.
And so Jensen's led the way in many ways, and including.
One of them is how to commercialize deep tech.
Absolutely.
Do you have any comments on this, George?
Well, I totally agree.
I think that another early stage thing that we can do that roughly involves quantum is natural computing, where the best simulation of a particular system is the system itself.
It's 100% accurate.
And there's certain systems where we have the synthetic ability to make them even more cheaply than we can compute.
But they're definitely quantum.
They're very complex systems are possible.
For sure.
So Anurad, you could have been on any one of these panels, all five, given how broad your company operates in.
But I do want to ask you about how is AI accelerating all the way, going from the chip to the entire system?
And, as we heard from our panelists before you, also into electrical and cooling systems and the power grid.
Yeah, first of all, great to be here, and So for people who are not familiar with Cadence, Cadence basically makes products to design chips and electronic systems.
And we have a long history of working with NVIDIA.
We have worked with NVIDIA for more than 20 years.
And it's remarkable to see what NVIDIA has become and what it will be going forward.
Now I think the key thing for chip design and also now the chip and systems are merging together, is there's a lot of deep science in it.
Basically, our products are software products, which is a combination of CS plus math, plus physics, and applied to chip design and system design.
The whole stack needs to be optimized.
What AI can do is it can provide the next level of innovation, the next 10x productivity improvement, because chip design itself is exponential.
If you look at chips now, by 2030, the chips will be 10 times bigger.
The systems will be 30, 40 times more complex.
So we need the next level of productivity improvement that AI automation can provide.
So to me, it's a combination of AI, the basic science, the ground truth, and accelerated compute.
This is what I've called for a while the three-layer cake.
And all three layers of the cake have to work together with AI, the ground truth.
That's still very important.
And then accelerated computing.
And the interesting thing with accelerated computing especially.
You know I always wanted for a long time a combination of CPU and GPUs.
And NVIDIA has fabulous GPUs, and GPUs have become much more general purpose as NVIDIA has latest generation.
But with Grace Hopper and now Grace Blackwell.
The fact that CPU GPU are close together gives a very great foundation platform to build science and AI on top of that.
Before we go to the next one, maybe just an announcement to everybody in this hall.
Jensen is beating you over to the keynote.
So I've been advised to tell everybody in this hall they need to make their way over to the keynote to grab your seat so that Jensen's not over there alone.
Meanwhile, we're going to continue this conversation.
One of the things that has been a promise for a long time is how technology is going to advance human genomics, advance drug discovery.
And I would say that there's been a lot of promise and there's probably been less success than the world might've thought 20 years ago in terms of the transformation that would come.
It feels to me like we've been laying down the tracks, right?
The groundwork, the primitives for these big breakthroughs.
And so my question to you is are we on the verge of really solving problems that we've been talking about for the last 25 years?
Do you think the cycle time on discovery is about ready to change because of where we are in AI?
And, you know, anybody who wants to chime in, we'll start here.
Right.
So I think we're at an inflection point.
Absolutely.
And it's for the reason you mentioned.
The tracks have been laid down and there have been advances just in the last few years that have made it possible to achieve liftoff.
If you think about the problem of drug discovery and the conundrum that we face today, when you start working on a molecular target to make a drug, make the drug, take it through clinical trials and get FDA approval on average that takes 13 years.
Nine out of 10 drugs that enter clinical trials fail.
The overall cost, including the failure, is about 2 to 4 billion per drug and it's not improved in the past 20 years.
And that's because drug discovery is still very artisanal.
A lot of science, yes, but a lot of art, intuition, empiricism, trial and error.
And at the highest level.
The promise of AI is that, with the right kinds of data and the right amounts, we should be able to transform this from an artisanal endeavor into an engineering discipline with much higher success rates and shorter timelines.
I think in terms of the inflection point, if you think about there are multiple applications of AI in the drug discovery platform.
One is at the level of logistics, if you will, designing clinical trials, recruiting patients.
There, the advances with large language models are already being put into action.
We're seeing companies across the world implementing this to improve the efficiency of recruitment, identifying patients, filing reports to the FDA.
All of those things are being accelerated in a dramatic fashion.
The second area is in molecular design that's been touched on in the movie and you've talked about it where, instead of screening for drugs, we're designing drugs.
Just as you can ask Sora to make a movie of two puppies playing on the beach, you can take a protein and say I'd like to make an antibody, a drug-like substance that attaches itself in just the right way, and have the AI design it.
A lot of action there, but the advances are recent.
The Nobel Prize a year ago went to two groups the DeepMind group, Demis Hassabis, John Jumper and David Baker, our co-founder for advances just in the past few years.
AlphaFold, RF diffusion, RF antibody that make it possible.
So we didn't have those technologies 10 years ago, even five years ago.
So that's prompting that acceleration.
And then the last area, in some ways the hardest one, is understanding human disease, human biology and human disease and getting those insights to figure out what should we make drugs to and who are the patients who are going to respond.
AI is just entering that area.
That's also, I believe, going to build over the next few years.
But we're seeing acceleration, very rapid acceleration in the first two.
That wasn't possible until a few years ago and we're going to see increased acceleration in the third.
George, same question to you.
Are we on the verge of a new cycle time in terms of discovery?
Is the promise, you know, I look at discovery and invention and these are nonlinear systems, right?
And it feels like we're at a moment, because of all the technologies that we're talking about, that a lot of this promise is going to come to bear.
What are you seeing through the lens of Lila?
Yeah, so this is definitely an amazing moment.
There's the intersection of exponentials that have been occurring both in computation and in biology.
For example, 20 million fold reduction in costs of sequencing.
And that's something we use not just to study the populations but to guide experiments, to interpret experiments.
So we're just that analytic tool.
But their synthetic tools are also getting better.
I mean even our confidence in having low toxicity and high efficacy is making the clinical trials shorter.
The record now, as far as I know, is baby KJ, which was seven months from birth.
Diagnosis to cure with a gene therapy.
And we are really actively using not just the dichotomy between screening and analysis and prediction, but putting them together.
So you can use AI to design as well as you can, but then you make big libraries because there's some modesty that you can't necessarily design the perfect thing.
But if you make these big libraries, that combination is much more powerful than either one of them separately.
So, for example, at Dyno and Manifold we've made proteins that target the nervous system 100 times better and they detarget the liver, where you can get toxicity and absorption of the valuable drug.
So those are just examples.
This is happening now.
This is not dreamy stuff off in the future.
Exactly.
So before I go to the next question, I have a public service announcement.
And that is the pregame show is actually being broadcast inside of the hall.
So as much as I'm sure you want to hear Brad and I and the guests, Jensen is a bigger show.
And you want to make sure that you get your seat right.
Thank you very much.
Okay, now to your regularly scheduled program here.
One thing I want to add on drug discovery.
Yes, please.
And Cadence has a molecular science division.
We are working with a lot of the pharma companies.
A lot of the math and science is similar in chip design, which is very nonlinear to drug design.
And something to watch because is how much of the work is done on the computer versus how much of the work is done in the lab.
And if you look at three big areas that I'm involved in, one is chip design and there 99 of the work is done on the computer and you get first time right silicon.
Then the second part is system design, design of planes and data centers and cars.
Over there, 20% of the work is done on the computer, 80% is still done on the physical work.
And then you go to drug design, molecular design.
I think only a few percent of the work right now is done on the computer, and most of it is done.
That's why the design times are so long.
So what happened in chip design is most of the work moved to the computer, and then you can do modeling, which is accurate.
Then you do simulation and optimization, which is basically design.
So the more the modeling is accurate, more simulation happens.
More optimization with AI can make drugs happen.
So I think it is definitely on the verge, But if you compare it to other established areas, it is still very, very early days in drug defense.
I just want to add, in terms of this transition to go from a few percent to 10 20, 30.
Part of that is having accurate simulations.
Eric Schmidt has a saying that is, I think, very profound, which is AI is to biology what math is to physics.
Currently, for physical physics, objects.
You can model them using the equations of quantum mechanics or other things.
For biology, we don't have equations.
AI is the tool to make sense of biological data, because AI can see patterns where we can't with human eye.
Of course, you have to feed the right kinds of data for that.
So what we're in right now is a moment where we and others in academia and other companies are generating massive amounts of data to train AIs that will understand biology, so we can have more and more of that stack in silico rather than in the wet lab.
That together with the automation that George talked about is what's gonna accelerate this.
So a follow-up question on that, and this is a bit of a challenge.
I saw a recent report that said that China has 10X more drugs in the hopper than the United States does, yet we have 10X the amount of data centers that they do and probably even a lot more compute power.
What is that disparity And I'm not talking about trials.
I'm just talking about the creation of it.
Why are we falling behind and what needs to be done to accelerate this?
Well, I wouldn't say we're falling behind necessarily because there's quantity of drugs and there's quality.
A game-changing drug is one that reverses aging because 90% of us are going to die from that.
Right.
But you can make a whole lot of Me Too drugs that don't really solve problems.
That's part of it.
I think, in addition, they do have some advantages.
Like they have this IIT system investigator-initiated trials which are very streamlined and seem to be dodging anything having to do.
They seem to not be running into problems with toxicity.
Focus on lowering that. toxicity problem.
So I think it's healthy competition.
But I think we're quite a bit ahead, partly because we are integrating computation so smoothly with it right now.
Now, I appreciate the clarity on that for sure.
One of the questions and feel free to chime in there as well, but one of the questions I have around quantum is, I think, for the average investor, for the average observer, They see these stocks flying.
They think it's part of the mean bubble that exists in the world.
They have no idea when the benefits will come.
So for somebody leading a company like Inflection, where you have real customers that are leveraging this today, help the viewers at home understand, like what is the important consequence for making these investments.
Are the stocks today in the quantum universe, are they ahead of themselves?
Do we have some of these mini bubbles where everything is getting pulled up?
Because I think deflating that and providing a real set of context is important for the longevity of the industry.
Well Brad, just for some context, and you know this, but before I came to be CEO of Inflection, I was an investor for 19 years.
And so One of the things I did learn as an investor is never give stock tips to friends, because you're going to be wrong more than 50 of the time.
And so I will punt on any commentary on the stock prices of the publicly traded quantum companies.
But what I will say is the reason this matters is going back to where we started.
Unlocking the power of quantum mechanics and turning that into products will result in orders of magnitude improvements in those types of products.
And so we're not talking 50, we're not talking 100, we're talking 10 10000, 1 million, X improvement in performance.
And so going back to what I was saying about how we followed Jensen's strategy on how to monetize and commercialize and build the market for ourselves in quantum and actually kind of show the world that quantum has real advantage today, as we pointed this, at areas where it already does have those types of 10 X, 10000 X, 1 million X improvements in performance.
And so timekeeping, RF antennas, things like this sensing, those are real market opportunities, to your point where there's real customers and products.
Today,
So part of it is, I think you can look to other areas than quantum computing to see real quantum advantage.
Now, when will we see advantage in quantum computing?
That's the big question everybody wants to know, right?
It all comes down to logical qubits.
Logical qubits are really the keys to the kingdom in quantum.
And this is one of the ways that AI and quantum are really going to interact and be tightly coupled, because AI helps us get to logical qubits faster.
Because logical qubits are error-corrected physical qubits.
So they're sort of these pristine qubits that you can actually use to do computation.
So up until 2023, the world did not have logical qubits yet.
We weren't even sure if we could get them.
In 2023, we saw the first logical qubits.
Today, Inflection and a handful of companies have logical qubits.
We announced that we had 12 as of a month ago.
And it's generally believed that about 100 logical qubits you'll start to see quantum advantage in areas like material science.
When you get to 1000 logical qubits you'll start to see quantum advantage in areas possibly like drug discovery.
And as you scale those logical qubits, you'll see quantum advantage in areas beyond those.
And so I do think this will be absolutely game-changing technology.
I left a great career at a firm called Maverick Capital to come do this full-time because I see the opportunity ahead is absolutely huge.
So I have to ask George and Mark, I mean, you're using GPUs today.
We'll call that It's hard to call GPUs classical computing.
I think that is CPUs.
So CPUs to GPUs.
How are you looking to quantum to help change what you're doing?
Because, quite frankly, if I look at the algorithms that they're working on and what they do, it would seem like there would be an opportunity.
Maybe not immediately, but maybe in the future.
Mark, you want to start?
Well, in terms of which kind of going from CPUs to GPUs to quantum computing or whatever?
Well, we're going to have a world.
I mean, listen, as we know, nothing goes away.
We'll always have CPUs, GPUs, and we'll have quantum, which will assist it.
So we're like many companies like ours that are applying AI to molecular design, AI to develop foundation models in biology.
We're big consumers of GPUs, obviously, in what we're doing.
There are certain applications where really having the power of quantum computing would come in very, very handy, particularly for molecular dynamics simulations, which right now require basically much more compute than what we like to deploy.
So we see lots of use cases where we could leverage that kind of technology.
For now, we can run with the technology that's being made available by NVIDIA today.
I mean one thing to add on quantum, and we work with a lot of the quantum companies as they design their computers.
And quantum is very promising, especially for certain applications.
It can give a huge speed up.
So we are closely watching when it gets to scale.
But I think the future will be hybrid.
I think CPUs are still important.
Oh, absolutely.
CPUs are phenomenal.
You know, FPGAs, custom silicon.
So I think it's not either or thing.
Quantum will have certain big applications.
It will give dramatic speed up, but all the other hardware platforms will work together to solve.
I completely agree with that.
I think just as GPUs layered into the data center and enabled new capabilities of what we could do with compute, QPUs will start to slowly layer into the data center and just expand what we can do with compute.
And I think it'll result in more CPUs and more GPUs being deployed and sold, because we'll be now addressing problems that just weren't able to be addressed with compute historically.
As we begin to reflect on the day heading into the keynote with Jensen.
In my world, the world of investing, the talk really is about a bubble and whether we're ahead of ourselves.
That's where we started the conversation.
I think it's super important to that we kind of steal people's minds for the fact that this is not a straight line up and to the right as we look ahead.
So, as you take yourself out of the position of running your companies, but just as an observer of technology ecosystems over a long period of time, when you hear all of the deals announced over the course of the past many weeks, trillions of dollars of compute that's getting built.
What concerns you the most, right?
As you're running your company, as you look ahead, what do you think could upend us?
What do you think might be that risk that causes this to slow down or us to get ahead?
Are any of you concerned about?
You know kind of the dark fiber and the GP you overbuild that you certainly hear talked about on CNBC every day?
Well, let me start.
I mean, what is promising for me to see is that there is still a lot of demand from our customers, the big companies, to build more and more compute.
And the way I think a lot of people don't quite grasp that there are multiple phases of AI in my mind.
And we are in the horizon one or first phase, which is the infrastructure build out, build out for on the cloud, on the edge.
But there is horizon two and horizon three phases which can be even bigger.
So the first phase is infrastructure LLMs And if you look at that first phase, I still see, talking to all the big customers, years of growth in compute and AI capability.
But what is even more encouraging to me is phase two, which I've always said for years is going to be physical AI.
AI is not going to be restricted to the cloud and to software applications.
It will move to the physical world, which is cars planes drones robot, and that could be trillions of dollars of monetization.
And then phase three, Even though we are already doing science now.
But science is AI, to me is the Horizon 3 application, with drug discovery, even material science.
I mean, these are trillions.
So I think we are only in Horizon 1.
Yes.
And talking to customers and partners, Horizon 1 still has ears and legs.
And then you add Horizon 2, which is physical AI, Horizon 3, which is science is AI.
I mean, this is a long way to go.
It's a great framework.
And I think we're seeing the same thing in our field.
In phase one.
The availability of the algorithms and the compute means that it makes sense for us to try to generate data at scale.
Previously, we couldn't have made use of those data or interpreted them in the right way.
So we're going to have that build up.
We're going to see the permeation of AI in all aspects of the laboratory.
That's how people will analyze their data.
It's going to become just routine and baked in.
So I see, you know, steady growth, and especially with automation of the kind that George is thinking about.
But also I think it's important to try to set expectations right.
It currently takes 13 years to go from starting a drug discovery program to having FDA approval.
We're not going to get it down, that whole 13-year process down to two years.
Can we get it meaningfully down to 10 years to?
I think we should give ourselves the ambition that over the next 10 years we get down to we cut it in half and we cut the attrition in half.
Those would be huge gains in terms of bringing life-saving therapies to patients.
But we have to be realistic.
Some of the AI companies that started before we did are already going very quickly from starting a project to getting into clinical trials.
There's one that announced recently.
Within two years they went from starting a project to getting clinical trials.
Normally it takes about five years to do that.
So I think we're going to see meaningful improvements, but you also have to.
But we already have an example of seven months, so let's not say that it's that far off necessarily.
And to the hybrid strategy, i think we're we're we already have.
Energy is going to be a big thing, and if we look at biological intelligence the, the energy consumption there is 12 orders of magnitude better if you talk about petaflops per watt um, and so we need to look at that as an example.
So we'll have hybrid systems.
We've already made single molecule transistors in a CMOS system.
We published this three years ago.
We're going to see more of that molecular electronics.
So hybridizing with quantum, which is quite different.
Well, looping back to Brad's initial question, which is the bubble bears, it sounds like and I say this tongue in cheek because i'm an optimist is you could do some pretty useful things with a hundred x or a thousand x more compute.
I mean, you could make life-saving drugs out there and things that can literally cure diseases that have never been cured before.
I mean, i think the way it's my mental model is is this is is an enormous investment cycle going on here and it's a return will need to be earned on these investments and, And so I think it all comes down to do you believe a return is going to be earned on these investments.
And I guess you could say, are the investments leading to useful things?
And it sure seems like they're going to be useful.
And so the question is, what kind of return are they going to earn on these investments?
And these multiple phases reinforce the previous phase.
So if there's infrastructure AI, physical AI sciences, AI in physical AI, when we deploy AI to robots or cars, of course the car itself has an inference chip, like in a Tesla or a robot.
But the new world model still has to be strained on the data center.
So not only the infrastructure AI is strong because of software application.
When physical AI happens, it reinforces the data center.
Same thing with sciences AI.
It will reinforce the data center and also in medicine there'll be a lot of robotics that helps.
So these layers not only are good by themselves, they reinforce the previous layer.
Yeah.
Gentlemen, this has been an amazing conversation.
Science and quantum.
We got a special guest, Jensen Wong, come up here and literally add some serious value here.
So thank you very much.
Thanks.
Great job, guys.
Thank you.
