SES AI Corporation (SES) Earnings Call Transcript & Summary

Hello, and good afternoon, everyone. My name is Shawn Severson, CEO and Founding Partner of Water Tower Research and Head of Energy Transition and Sustainable Investing Research. Exciting topic today. We're going to be discussing AI with the CEO of SES AI Corporation and ticker SES. We have Dr. Qichao with us today, and we got some great topics to cover. I would also encourage you to take a look at our website at www.watertowerresearch.com. You can find all of our research there on SES and other companies as well. It's all open to access format. So I encourage you to take a look. You can see fireside chats research notes, et cetera. As a reminder, this event will be available on demand. You can access it by using the same link with the original registration. So should you choose to forward it to anybody or provide additional access, again, use the same link to use that. And today -- today, we'll get started, and I'm going to jump right into it. Qichao, are you ready?

Sure, sure.

So AI, of course, have been at the top of the news for quite some time. You can't escape but it's everywhere. But it's something that you've been working on for quite a while. You've been using it at SES. And I was just wondering if you could provide us a little history how you're using it? How it came about? And then we're going to discuss, I know some of the specific applications, but a little history for us, and I think to start would be helpful.

Yes. So it really started as a request from the OEMs when we went into A-sample JDA. And then -- so the 2 aspects for safety -- the 2 aspects for using AI in battery include AI for safety and AI for science. And then -- so the AI for safety is probably the most important. And then that came about starting around 2017 when the OEMs started realize, okay, when they have more EVs on the road and they're starting to have these failures and that the failures would lead to recalls, pretty expensive recourse. And the traditional [ all the ] manufacturing really just was not enough. So the OEMs needed a way to collect the data and monitor the vehicles to really ensure 100% safety, a more advanced modern way of all the quality control than the traditional quality control that really was developed by the Japanese in the '70s and '80s. So that's one, that's AI for safety. And then another is more AI for science. It's -- that's really to accelerate the development of new materials in the battery space. And then this model, the pharmaceutical industry has been doing this for much longer using AI to discover new drugs. And they've done a really good job in terms of mapping a large molecular space and also learning other properties. So we're beginning to also use that but for batteries. And we actually think AI for science will have faster validation in batteries just because the testing period is much faster than in the [ pharma ] industry. So two things, AI for safety and AI for science.

Can you just walk through the platforms you have, just I think investors can frame up when you discuss, you're talking with investors. I think you have 4 kind of platforms for AI. Just maybe walk us through those before we get into some more of the details.

Yes, the way we develop products and then our product is basically the battery plus this health monitoring software. So we start with what we call [indiscernible], which is AI for science. Basically, we start with the entire database of small molecules. And then to put things into perspective, the entire battery industry in the last 30 years only explored several hundreds of unique molecules, several hundreds, so 10 to a second. And all the batteries that we're using today in everything are based on the tiny space, 10 to a second. But actually, the entire database of small molecules could be as big as 10 to the 11th or even bigger. And then no one in the battery industry has explored this entire space. So our goal is we want to explore this entire database so that we can find better molecules with better performance for the electrolyte, for other materials in the battery. And then if you can -- if you think of all the progress that has been made in the past 30 years, just based on that 10 to a second database, imagine how much more progress we can make if we explore 10 to the 11th database. So that's the first platform basically developed. We come up with new molecules. And then these new molecules, we formulate, we build -- we formulate electrolytes. And then -- that's the second platform called Hermes. And we recently established this electrolyte foundry in Boston. And that platform will test different formulations based on these molecules. And then third is Apollo. That's basically we take the latest largest formulation and then build batteries. So we have the A-samples, the B-samples. And then we announced 2 B-sample lines with 2 carmakers. And also we are converting the A-sample line, the previous A-sample lines to make UAM cells. So that's Apollo. And then last one is basically the product. So we have these cells from Apollo, but then we're going to take Avatar. And Avatar is going to monitor the quality manufacturing data and then also the actual vehicle data and then build this digital twin version of the battery from birth to the entire life on the vehicle and then predict incidents and then try to stop that for the instance happens. So [indiscernible] to Hermes, and Hermes to Apollo and the finally to Avatar.

And when we talk about AI, you're using AI, right? So we -- as a company, you're taking, I would say, off the shelf, but I'm trying to understand how you're taking AI? And are you customizing it or tweaking it to solve for those specific needs that you have? I'm trying to understand the process from how AI as itself getting to a product and how to use it?

So we actually build our own models, too, just because a lot of the models -- so yes, there are these open source models that you can use. But then a lot of the models have not been sufficiently pretrained. And so the output you get are not so meaningful. I mean, our goal is not to use AI for the sake of AI, our goal is to use AI to solve problems that humans can not solve. If this tool can solve problems that humans can solve, it's not very meaningful. It has to solve the problem that humans cannot solve, then it's meaningful. And a lot of the open source models can only do things that human scientists can do. So not a meaningful. We want to pretrain these models and then also build our own models so that they can actually do things that our human scientists cannot do. So the AI for science and AI for safety. AI for science, one is, actually -- so the entire battery community actually lacks data, so the molecular data. And so we talk about the industry only looked at 10 to a second. So the industry has a lot of data on 10 to a second database. But when you get to 10 to the 11th, no one has that data. That data does not exist. So we actually have to build a database on the 10 to the 11th, and then pretrain this model that we are making so that this new model can actually build a better electrolyte molecules. And then for the AI for safety, the Avatar also, the data for lithium metal, all the manufacturing and then also vehicle performance don't exist. But we have a lot of data from the A-samples and the B-samples. We're building more cells. And also, we have more quality control points per cell. So we have more data. And then we're using this data also to train this model. And no one has a model that's specifically build and pretrain on vehicle battery performance. We have that. So we actually do build our own model. And then -- so we build our own model. We collect the data and then we pretrain our own model with the data. And then you really have to do all 3 things to get a model that can have a chance of solving the problem that humans cannot solve.

And that's, I think, the unique aspect is you have that data, right? So you're starting with -- you have a starting point that makes us truly proprietary in the model itself because the data is exclusive to you and in building the model. So this is something that was very unique to SES. Is that the correct understanding?

Yes. Yes. So database, both in terms of molecular database and also all the quality manufacturing from the A-sample lines, the B-sample lines and all the testing. So we have proprietary database. And then we have [ domain ] experts, basically battery scientists and engineers to train the model. And we also have this ability to synthesize new molecules and then build batteries and actually test that. So when the model comes up with new solutions, new molecules, you have to synthesize that and the test in the battery to prove that, yes, it has higher from efficiency, higher cycle life, all of that. And also, when this Avatar model gives you a recommendation that when certain thresholds are hit, then this battery is going to have an incident, you have to prove that. You have to have these bunkers to test the batteries so that to actually prove that, yes, this battery is going to have an incident when these parameters are hit. So we have all the database models, domain experts and also the ability to synthesize and actually test the batteries and then to verify the output of the model.

Let's take it into an actual -- into the application and say, how does Avatar actually improve the site and safety of lithium metal batteries, both for EV and UAM? So trying to translate all of this into how it works in the field, let's say, and what it would be able to do at commercial scale that's -- that would be unique because of the AI?

Yes. So AI for safety in EV or UAM, currently, when a large battery manufacturer produces a lithium-ion battery, there is no data on that battery. There is only data on the badge of batteries that this battery was produced in. And then -- and the threshold for quality manufacturing and the screening is quite binary, basically good or reject. But then within the good, you can still have cells that barely meet the requirements. So we have this false positive. And then when these false positive when the cells that barely meet the good requirement end up in the vehicle. And then you get a pilot or a driver that's abusive to the vehicle, so fast charge at low temperature, then you have basically a recipe for an incident. And then -- and so what AI does is instead of having this binary output from the manufacturing reject or good, it's going to assign you a very detailed value. And then, for example, say positive 0.2%, positive 0.5%, negative 0.7%, negative 0.9%. And the cell that has, say, positive 0.3% will have shorter cycle life than the cell has a positive 0.5%. So it's a lot more detailed. And then also, one of the batteries are on board, the vehicles it collects data from the battery 24/7. So if we have a nice driver, if you have an abusive driver, the model takes into account the actual battery life performance data. And so the model is going to consider both the quality data and the live performance data and then give you a very accurate prediction of when something bad may happen. So that fundamentally is very different from just traditionally just relying on traditional statistical auto manufacturing and then putting this in a car and basically have no consideration for the actual quality as well as how the user uses the car.

So wasn't this have a lot of implications for warranties, for example, or resell value in these? I mean thinking about practical application to OEM in the field, there's obviously safety, right, is the most important. But I assume that this has meaningful economic impact to an understanding life cycles and costs, right?

Yes. Yes. And especially if you think about the economic value of EV or UAM, basically the economic value of any electric vehicle is actually primarily the health of the battery. And this model, this AI will give you a very accurate assessment of the health of the battery at all times when it's fresh from the manufacturing line and also at any time inside the vehicle.

So this becomes something if I think about use cases again, it's going to be something that has real feedback -- real-time feedback potentially for the driver, right, but also the data collection never ends, I assume. So as these cars go out into the field and the batteries are being operated, the data set is just going to get more and more robust and we'll continue to learn, correct?