Can AI Create Materials That Never Existed?
Max Welling, one of the pioneers of modern AI and co-founder of CuspAI, explains how his company uses AI to design new materials for uses ranging from semiconductors and batteries to carbon capture and water purification. He describes the pipeline from a customer's request to generated candidates, simulation and synthesis, and says 2026 is the year CuspAI has to deliver materials into the real world. The hosts then weigh whether that timeline is realistic.
Transcript
Nobel laureate Geoffrey Hinton has warned that AI could pose an existential risk to humanity. Turing Award winner Yann LeCun has dismissed those fears as overblown. They disagree on almost everything, but they both back the same company. CuspAI was co-founded by Max Welling, one of the pioneers of modern AI, and he is not busy building chatbots. Instead, he's creating and harnessing AI to discover new materials for everything from carbon capture to water purification, to plastic alternatives, and more efficient batteries. Some estimates put his company valuation at more than $500 million, but not a single material it's discovered has been deployed at scale, yet. So if this works, it could change everything. But the question is, can it? And if so, when?
This week, we talk science, scale, and when CuspAI will be able to deliver with company co-founder Max Welling.
So, Stephen, we aren't talking about building an updated version of the latest chatbots. So what is it that CuspAI is doing that makes it so different from what we talk about every day with regards to AI?
Well, so I'm super excited about this. Super excited that for once we're not talking about LLMs, large language models.
Right.
Fed up we're talking about large language models, actually talking about something practical and useful.
Right.
As you say, a search engine to find new materials, you just type in what you're looking for, lo and behold, it appears. So that's exciting. I love a bold claim, Autria. You know, it says it's going to do trillion dollar material breakthroughs in months, not millennia. I kind of love that, and I love the work they're doing with some big companies like Hyundai to actually bring this into fruition. But, and there's always a but, right, it's just as you said, $500 million is a lot of money.
Right.
So how do we know that it's working, and how do we bring it from theory to reality? And to answer that, I think Laila...
So I think you guys will know that I'm pretty skeptical when it comes to AI facilitating and propelling science forward. I think it's pretty crazy, people who claim that AI will solve all problems in science. Because scientists are like the most intelligent of us...
Or so they say.
Or so they say. So I think it's going to be pretty difficult to not only generate new materials, but actually build them in the real world. And I think the best comparison for me is AlphaFold. And I think because we've spoken to John Jumper, who obviously was responsible for creating AlphaFold, and it was really, really hard, and they failed quite a few times. And it took a couple of reiterations to get to where they wanted to. But I think CuspAI doesn't have all the luxuries that the AlphaFold team had, which was data, which was scale, just reliable, experimental data which they can get quickly and has accumulated over a long, long time. So I have just no idea how they're going to do it.
Look on the bright side for Max. At least John got Nobel Prize, right?
Exactly, exactly.
All right, well, we want to know what you think. Will AI really be able to deliver on the promise of life-changing scientific breakthroughs? And if so, can they do it in months? Tell us in the comments below and don't forget to like and subscribe while you're there. Now, let's hear it straight from CuspAI co-founder himself. Here's our conversation with Max Welling.
Well, Max, CuspAI is really vowing to deliver on some big promises, to deliver on groundbreaking materials in really just a matter of months. I read an article where you said, "2026, the focus is going to be on delivering," and the word is deliver. So to put you on the spot, what types of tangible deliverables do we have so far?
So we are currently building our platform. And so 2026 will be a time when we will work with customers to actually design new materials, but also synthesize them and put them into the real world. And so you can think about materials in the semiconductor space or battery or fuel cell materials, but also materials that capture PFAS out of water or carbon dioxide out of the atmosphere. So that's our goal for '26, to actually design these materials and, with our partners, put them into the real world.
By the end of the year.
By the end of the year, that's right.
Could you tell us a little bit about how you get from the theory into the practical, how you get from the design of the molecules that, say, into actually making something?
So it starts with, let's say, a query. Like, you go to a customer, and then the customer says, "I would like to create a new material that has these specific properties that I'm interested in." And so the first thing that will start is a agent that will then start to do a literature research to see what's out there in our databases as well as in the sort of open source literature. After that assessment, the materials that come out of that can get tested for how appropriate they are for the task.
But it could also be that it turns out there isn't a material out there that satisfies the properties, and then our generative models come into play. They will then generate entirely new, you know, materials that have never seen the light of day in this universe. And we typically generate hundreds of thousands to millions of those, which will then all get tested by our, what you call digital twin. The digital twin consists of very cheap property predictors, which will very quickly assess whether the material could exist in this world, to the point where we do more sophisticated molecular dynamic simulations to assess their actual properties in high accuracy. And then after all that is done, we then typically generate a synthesis recipe. And with our customers, we then start to synthesize these materials and then actually measure the properties of the material.
Now, I should also say that in '26 we have another goal, which is to connect to high throughput self-driving lab experimental facilities, we have a number of partners around the world, where the feedback that's coming from the physical world, the actual experiment, is much faster. And so we can then take the measurements and the experiments we're doing back into our platform, retrain the models and have a much more, I would say, balanced sort of dance between the digital part and the experimental part of the platform.
On that point, Max, about making the real-world experiments faster, I mean, you spoke about synthesis and synthesizing these materials in 2026, but isn't synthesizing really, really difficult and unpredictable and hard to do fast? I know you mentioned a story about two labs trying to synthesize the same material, and in one lab, because the atmospheric composition was different, they couldn't replicate it. Like, how do you get around that with this situation?
I think this is a really good question. So in the past, we've gone through a sort of a project where basically we wanted to reduce the amount of time for designing new material from years to six months while building the platform. And we had a partnership with a synthesis partner, and then we basically gave this particular material to the synthesis partner and asked them to then synthesize it. And it was quite hard to actually synthesize it because then you'll have to actually invent the synthesis process. And this can sometimes work and sometimes not. So one way around this is we ask the partner to tell us, "What are the constraints under which you can actually synthesize more easily?" And so that's one thing you can do. And you put those constraints into your generative model so you will only generate things which can be synthesized.
But a better way to do this is to synthesize a lot of things in a high throughput setup. So everything you synthesize there, where you just change all the process parameters in a sort of systematic way driven by an AI, then whatever you synthesize is actually synthesized, right? And so it's there, you then have to measure what you synthesized. But then at least you know that, you know, as you're going fast through this cycle, so you don't have to wait until, like, six months after you designed it to see if you can synthesize it, you synthesize much, much faster. And that feedback can then help you make sure that you only search in the space where things are synthesizable.
So, Max, I wanted to ask you and build upon your previous answer about verification and the speed and scale of verification. So you had AlphaFold, which you had kind of immediate verification in the form of existing protein structures that you could compare to. And then you had AlphaGo, where the immediate verification is whether the model win or loses. And same with something like AI coding assistants where they can immediately build something. So do you think the key here is cheap, clean, verifiable feedback to get to where we ultimately want to be with materials, innovation, discovery?
I think you can also call this the reward. So the system needs to have a reward of how well it's doing because that would tell it, you know, how to climb the hill, how to do better next time around. Now, a material science, it is not as clean as for coding where you can run the code and see if it gives you the right answer. We need to do, let's say, a molecular dynamic simulation, if we do it digitally, in order to verify whether the material that we generated has the correct properties. And then we have to go to an experiment, which is kind of the ultimate truth.
But in an experiment, you cannot always do the kind of the measurement that tells you whether the material that you have has the properties that you want and what the structure of that material is. So typically you have fast measurements you can do that give you an indication, and then slow and more expensive experiments like an X-ray diffraction measurement, where you can actually verify that the material that you have is the material that you wanted to have. So it is somewhat more complicated story, but still it is of course key to be able to determine whether the material you have generated has the right properties, yes.
And that brings us to a perfect stopping point. Because coming up after the break, we're going to ask Max Welling, can a European-based AI company actually scale? Stick with us.
Back with Max Welling. And, Max, you're promising some pretty big breakthroughs in several different arenas. So like we mentioned, from carbon capture to eco-friendly plastic substitutes, to even water purification and batteries. I would like to know from you personally, what is your prediction for which of these areas where we will see tangible success first?
It's of course very hard to predict. But if you put a gun to my head, I would say the first success would be probably in the semiconductor space. Now semiconductors are used in both chip manufacturing as well as in solar panels, and I think, you know, we see quite a bit of traction there. Now, the reason I'm saying that is not because our platform is uniquely positioned for that particular material, but it's also because we have partners there that are willing to work with us to make it a success. So you really need a deep partnership because you have the domain expertise from your customers or your partners, and they typically also have the experimental setup to then actually manufacture the materials that you design. And we strongly believe in this interplay between the digital part, which we represent, and then the experimental part and the domain expert part, which sits often with the customer.
So to that point, do you think that a success with semiconductors would have the furthest reaching real-world implications?
In the semiconductor space, for instance, for chip, for new chips, to keep Moore's law going, new materials are introduced in about one a year. Every node there's about three new materials which get introduced. And so the space, and we can also say that, in the semiconductor space, we are entering a stage where lithography has come to a point where it's almost creating the smallest structures possible because we're hitting a few atom, the size of a few atoms. And we are now creating chips that grow in the third dimension, so they become sort of taller. And in order to make that happen, we really need new materials. So there is a good, you know, ask for the design of new materials. And I would also say that the machines that manufacture these materials, these so-called atomic layer deposition machines, are very advanced. And we can very quickly verify that our new designs can actually, you know, can be made and also have the right properties.
Actually talked about having the right partners, and I noticed, you know, that you partner with people like Hyundai. When it comes to scalability, can those partners be found and can that scale happen in Europe, or is scalability difficult in Europe?
I don't think it's more sort of difficult in Europe. It depends a bit on, are you talking about scalability of the actual material? Can you make the material at scale, or can you scale your company to a very large company? So those are two different questions, I think. But we have a very strong chemical industry in Europe. For instance, we are working with Kemira. And if you find a partner that already has the ability to scale, like BASF, by the way, is another company that was the first company that scaled the metal-organic frameworks to the ton level. And so you see that, in Europe, we actually do have, you know, the right companies to help us scale these materials to the very large numbers, yeah.
Following on from that, Max, I want to get more into kind of industry versus academia. You guys are obviously a commercial company, and you're hoping to do something that will really benefit society and solve some of the most important problems in the world. But to Autria's point and question about, okay, is what you're focusing on right now the most important thing, how do you ensure that you are aligned with the most important problems and not just what the world is commercially interested in? Because right now the world is commercially interested in building as many data centers as possible, on the moon, in deserts, you know, in space. But is that what's actually best for all of us?
Humanity.
Yeah. So how do you, you know, deal with that, that alignment issue?
Yeah, it's a fine line. Because if you only work on things which are, you know, idealistic, let's say, or the best for the world, there might not be an actual market for that. And as a company, you would go bankrupt. So that's not actually helping anyone. And so you need to find the problems out there that both make a positive net impact in the world as well as there is a strong market for it. And so we see things like batteries and photocells and chips are things that create positive net benefit to the world. And we have a ethical framework, and we have lively discussions with the whole company about what topics we should, you know, avoid working on.
So kind of to pivot a little bit from that question but stay in the same vein, you know, we cover a lot of different layers of AI here on Agents of Tech, and primarily when people think AI, they're thinking LLMs and chatbots, and that's pretty much what the general public is using when they use AI. But it seems like what you're doing, your company, you're really one of the few fighting the good fight and truly trying to use AI for the betterment of humanity. I mean, if we're going to find a cure for cancer, it's going to be with what you're doing, not the latest version of ChatGPT. Do you feel any kind of pressure to deliver on that?
I always feel pressure to deliver, definitely. But I would say, yeah, we do use LLMs in our space as well. So our agent actually that designs the whole computation and the whole design process itself for the material is an LLM that either is one from, you know, the OpenAI houses, or it's an internally trained open source AI. For me, the only reason I do this is because I do want to make a positive net impact. And so if that would go away, it would not be worth it for me to put in all the hours we are putting in here. And I think the same goes for all our employees. We carefully pick them, that they are aligned with our mission. And so if times are a little tougher, they will stick around because they know they do it for a good cause, and I think it also creates a beautiful sort of culture inside our company.
Max, you talked about 2026 being the year to deliver, what does success look like? How do you know whether you have actually delivered in 2026 or not?
If we sort of have designed a material and it's being synthesized, and moreover the company that we work with will claim that they will use it into their actual products, that go into the real world, and this material could not have been designed without AI, I think that defines success for us.
And, Max, what do you think is your biggest obstacle? Because you're really innovating in the real world, this isn't just simulations which predicts what a material will look like, it's actually putting it out there. And, historically, innovation in the real world has been really, really hard. As in we've made...
Slow.
We've made a lot of progress in the world of bits, but not so much in the world of atoms as much as we hope. So what is the biggest obstacle you need to overcome in order to deliver?
I think you just said it, which is the real world is, unfortunately, full of friction, which means that when you connect the platform, which is bits, where everything is nice, and you connect it to the physical world, where you have machines that, you know, you need to put things in, that take time to produce things on the other side and need to be measured, all of that goes slower than you want and is more friction. And then the other thing which is tricky, which is once you have even synthesized the material with the right properties, you will then have to scale it so that it will, you know, be a material that's used in the actual real world. And there's many steps there which are quite unpredictable. The way we handle that is that we talk to our partners. So they have collected the domain experts on how to scale things. They can tell you early on that certain things look promising rather than they will not be able to, scalable. So this deep partnership with your customers is really key in order to not waste time in the beginning and produce something that could never make it into the real world.
We have a couple of quick questions for you that we ask all of our guests 'cause we like to get different perspectives on the same questions. So my question for you, Max, is where do you feel people will draw the line when it comes to bringing AI into their personal lives?
Well, it's very different for different people, and I cannot speak for everyone. I think I do know a few things that we don't want out of AI. We don't want AI to invade our privacy. We don't want AI to be used for mass surveillance. We don't want AI to get us addicted. We don't want AI to manipulate our opinions. So I think we need to be keen in order to make sure that that doesn't happen. And we probably have to regulate AI to make sure that those particular pitfalls of AI are not going to happen.
And my question is what is something that's widely accepted in your field that you actually disagree with?
So I think, right now, perhaps people bet too much on this superintelligence to arrive. And for our field, I don't think superintelligence will, you know, basically solve all our problems. You know, superintelligence, however you define it, will come at some point. It will be better than humans in many things. We are already creating, you know, AI that does that. But I don't think it will solve, for instance, the design of new materials. And so I think what we will have to remain humble in the sense that the real world creates all sorts of obstacles that we have to go around, humans have a very particular kind of intuition and knowledge that will remain very useful. And so the way I view our platform is more as a tool that gets increasingly better, but it is being used by humans to design materials rather than completely replace them, and I think that might be a view that might be a little overhyped at this point.
Max, question from me, which is if you look far out in the world of AI, let's say 18 months away from here, what can you see that is going to happen that maybe people aren't talking about now?
If we go back, there was the Industrial Revolution in which human labor was replaced by labor, by kind of machine labor. And then now we are in the digital revolution where human sort of intelligence is replaced by machine intelligence, or at least enhanced by it. But what we are going to see is, you know, a sort of virtuous cycle where AI is being used to generate new materials, and these new materials are going to be used to enhance AI again. And I think this is kind of the third step maybe in the sequence of three revolutions. And I truly believe that, I don't know, you know, how long in the future, we can design materials that feel quite exotic right now, with properties that you could not imagine, and it could completely change the world, and hopefully for the better. That's what we are shooting for. And so that's what I will be fighting for, at least with CuspAI.
All right.
Awesome. Thank you so much.
Great spot to leave it. Max, thank you so much for your time today. We appreciate it.
All right, so 2026, he seems pretty confident that there will be something tangible to come out of this year. It's a bold claim.
I feel like 2026 has already been so turbulent, so, like, I don't know how he's going to ride through that turbulence. Like, this is probably one of the hardest things ever. Like, one of the hardest challenges ever. So I think the idea that, like, you could deliver something in a year is just mind, like, I still can't get my head. Like, I would love for it to happen, I would love it, but I feel like there's so much uncertainty at every single step and you can't simulate for that in a computer.
It makes me wonder, though, if things are a little bit further along than we know about, if he's making that type of a prediction.
Yeah, I think it's interesting, isn't it? He talked about, to one of Laila's questions, I think he talked about scalability and scalability in two ways, right? One, he talked about sort of scientific scalability in one; and two, the actual making of stuff at scale. It seemed to me that he had some good answers to the first set, which is the kind of scientific scalability. I was less convinced of the answers to the real world, when you actually get to make things that you actually have to sell and make a profit for, that I was maybe on the sidelines. Maybe '26 is going to come too soon.
But I do think the theory checks out, as in I think his answer to your final question, which was about AI feeding the energy side of the equation and the energy side making AI better, thus reducing the entropy side of the equation. That self-serving loop, like, it adds up, it makes sense, it leads to great things for humankind.
Right.
But how we get there and...
Right, but materials are real things, right? You can hold materials, you can touch materials. And he said you actually need to put them into machines to make them. Now, are those machines going to be profitable? Are they going to be in a place where they can make the things that are needed? Those are whole sets of questions which is more than feeding AI in the loop.
But maybe the market will just, you know, make it happen.
We really hope that they're successful because I do feel the work that he's doing and that his company is doing will be where we see AI actually have a positive impact on humanity, so...
Yeah.
Although...
We're rooting for you, Max.
They are using LLMs, so...
They are. Well, there's a place for everyone. There's a place for everyone. But I do feel what they're working on is hopefully where we're going to see the biggest positive impact come from AI. So we shall see. A huge thank you to Max Welling for joining us. If you have thoughts on whether AI-driven science can actually scale to the real world, we would love to hear them. Drop your comments below and then be sure to like and subscribe. Thanks for joining us, and thanks for joining the conversation today. We'll see you next time right here on Agents of Tech.