Siemens President & CEO Roland Busch, “How AI Is Transforming Manufacturing”

Thanks for having me.

I mean, what the message was the following that there's a new general purpose technology which is AI, which is transformative as much as electricity was at that time. And I said actually steam took 60 years to change societies. Electricity, interestingly enough, 30 years. And it changed a lot. It changed the way how we manufacture, how we communicate, how we transport. The computer took 15 years to penetrate societies and guess what? AI is supposed to do that in seven, maybe eight years. Three already gone. So, and think about if this is beyond imagination because normally you measure the change of technology in your lifetime. And think about it, steam 60 years, that's basically your lifetime work. 30 years is still quite a bit, but seven or eight years, this is no time. This is super fast. And it's a general purpose technology because it addresses everything that has to do with what makes a human different. It's about rationalizing data and then coming up with conclusions. It's about writing the next word you can think of. So this is so fundamental and it will change everything. And just to make that aware, it's not only manufacturing, it's the way how we design, how we manufacture, how we operate things. And this was basically my intro topic and then I had a deep dive into a lot of things which we're doing. Yeah.

Well, I wish I wish I would. We're working on AI technologies since decades. And actually on this CES show I was presenting a piece of hardware, Synapse. It was, I think, back in the '90s where Siemens was producing the most powerful neurocomputer at that time. We stopped doing that, obviously, and I actually handed over one of those historic boards to Jensen a year before. I gave him that predecessor of what a GPU is today. And so, we know this technology is very powerful. We have it in use already in trains and everything. But the moment, really the moment where the breakthrough or where I always say where the genie got out of the bottle was when with ChatGPT, this technology was made accessible to anybody because you can interact with your own language. And this is where the technology is taking off. Before that, it was meant to serve scientists, mathematicians, physicists to deploy it. But the moment when you really can interact with normal language, it scales. And that makes a big difference, including processing power, data availability. So a lot of things are coming together. Normally, disruption happens when not only one thing changes, but a couple of things coming together and they make an explosive mix, and that's happening as we speak. So, I mean, we are blessed, the younger guys in the room blessed to live in that time because you will experience what this technology will do.

So, the first thing is the models which we define, foundation models which we see currently, which we're dealing with, they did not see most of the industrial data yet. And these are different modalities like time series data. It's a completely different animal. Design data, CAD design data. You have to read them with a software to contextualize them. Any kind of visual inspection data. So all of these data are not really fed into the models. So this is the first part of your question: we want to train models on industrial data. And you need a lot of them. When we hired recently our lead for AI technology for this industrial foundation model, we said, 'No, I give you access to all the Siemens data.' He said, 'Well, not enough.' So I mean, take everything, every 10 years of generations of controls which we designed, 20 years of generations of trains and all that. Giving access to all these data is still not enough. So we need a lot of data to train models on. Number one. Number two is that we believe that you have to understand the real world in order to take advantage of the AI world or the digital world. So the combination makes the mix. Is it that you need to have domain know-how? That you understand why did you design a product as you designed it. So very often when I talk to my AI experts and they say, 'I want to have a new design.' They say, 'Okay, give me your design data from the last generations.' Number one is I throw a bunch of data analysts and AI scientists on it. But I need the people who did do the designs of your last generations. Because they have the domain know-how to input why did you do that. This is super relevant. So this blend makes a difference in the way how you design something newly, but also how you really at the end of the day make an impact in the real world. So this combination makes the trick. And hardware is at the end of the day, this is what I said as well, is more important than ever. Because if you want to monetize all the investment that AI is now sucking in, it's not enough to train models and to throw them on knowledge-based work. It's really about getting impact in the real world, which really has a chance to monetize all the investments. So therefore, it's super relevant. It's difficult because hallucination is not an option on the shop floor because fixing a problem on the shop floor is difficult. Having a robot do repetitive task is one thing, to do any kind of task is another one. So there's a long way to go and it takes a little bit longer than we have imagined. But it will make its way and it will change the way we manufacture. Yeah.

Everything. So, we are trying now. We're building data alliances. This was very interesting. We announced it a year ago, actually, a data alliance with the top machine builders in Germany. And I would if you would have asked me a year before, I would say they never share their design data or their job data. I mean, they talk about the job, they mean this is your digital twin of a part, this is the machine, digital twin of the machine, the controls, and you basically find the sequence with which the machine puts a tool and mills something out of it. So this is what they call a job. And they have hundreds and thousands of jobs they can share. And I thought they never would. They did. Reason is that there's a huge competition from China. The machine builders are under huge pressure. The Chinese built maybe not the most complicated machines with 64 axis and everything, but normal machines, they make it 20, 30, 40% cheaper. So the way to differentiate is take the engineer out of the equation. Basically make an autonomous working machine. You just feed in the part in the digital twin and let the machine go. But for that you need to train the machine. And for that you need data. And so they said we share the data with you guys, Siemens. We have also our own machining data. And now we are training that and give a foundation model for machines, for CNC machines, which they can now use as an application to run the machines. So that's what we do. We get data from our customers, from our partners. We do similar synthetic data. But you need real data. Synthetic data alone doesn't do it. Maybe not yet. And you see it also when you talk about robotics, most of the robotic companies have gyms where they use teleoperators to train the robots. So I thought with my humble assumption was just take a simulation, simulate them and then simulate what they have to do and then let them work. It doesn't work. You have to train the whole technology stack, the brain and the hardware. And train it in the real world to make it work. So you need to gather your own data, you need to encourage to share data, you need to de-silo your data, make them available. And you have to tag them in a proper way to make them really valuable data, which is a hard piece of work to do, I have to say. It was encouraging to see more sharing in the ecosystem.

Yes, which is not in the DNA of the industrial companies. They don't share. They don't believe in sharing their data.

One of the hardest problem is definitely with just we are in your robotics labs is to develop robots which are, and I'm not talking humanoids because we are more focusing on the shop floor robots which are fulfilling any kind of complex tasks on the shop floor. Packaging something, putting a connector in a socket or whatever it is and with high speed, not in this boring slow process. So these kind of problems is hard to solve. That would be definitely one. And the other part is actually train and educate people who can do the job on the shop floor. Yeah.

So we are going all in. We spend a lot of training. Actually Siemens spends 450 million euros on education and training of our people and we are pivoting definitively into AI training. So we are trying to get training to everybody. But also to our engineers, to designers, so that's super relevant to bring them up to speed. Then the point is again having the combination of domain know-how and AI technology, this is really the difference. Our software guys, simulation guys, they say we tend to hire now less software engineers. We go to hire domain know-how expertise because these guys can write the spec. This is super relevant that you have a proper prompt which is in that case writing a spec what you want to know, what you need to have, including the boundaries and homologation which you need to do. And then let's code or let the system code. This is a particular challenge. I guess some people which I met recently at tech companies, they say coding is solved. So don't bother. It's done. So what just think about what you need, what is the specification, and the thing in between is not that relevant anymore. I don't know whether I would put it that way, but we're getting close to that. Certainly advancing quickly and figuring out security and so on and compliance. Exactly exactly. Yeah and that makes a difference. Since Siemens is a hardware company, we are a software company. We have the largest industrial software portfolio in the world. So you can build our software, the most comprehensive digital twin which will behave like the real thing. I mean because we can simulate everything, mechanics, electrics, the software running on it, fluid dynamics, whatever. But on the other side we build stuff. We know how to automate a plant. We know how to automate an assembly line for a car. We build trains. So we have so much expertise and this combination is somehow unique. You don't find so many companies who are so deeply entrenched in the software hard core as well as in hardware. By the way, these disciplines they don't really blend. So it's easier said than done. So you say throw a bunch of software guys on a hardware job and they solve it. It's different. The development cycle is different. The monetization, the business model is different. The go to market is different. So these are two areas which normally don't blend and we try hard to bring this together. Took us a while and we're still working on it, but once you manage that, once you get these disciplines together, I mean that can do really something magic, in particular when you use AI.

Where to start? Actually, the power sits really in the two largest economies in the world, the United States and China. I have to say Europe has a strong starting point. There are so many really top-notch companies. The Mittelstand is small and medium-sized enterprises in Germany, hidden champions. They are world market leaders in certain disciplines, machine tooling, machine building, and the like. But in Europe, and I come to we can talk about that as well, Europe is regulating too much. We are regulating before we're innovating. Is it the AI Act, the Data Act, the Cybersecurity Act? That's not really encouraging speed of innovation. So therefore, what I see, the adoption rate of technology in Europe is just too slow. There's a reason why I put my AI team now to the West Coast and another one to China, obviously. So these are the two most powerful economies when it comes to new technologies and AI. But there's a difference, though. I think just because China doesn't have access to the latest 3-nanometer GPUs, they are lagging behind compute. So the models in United States, I would say they are ahead, but China is playing the open models card. Most of their models are open. DeepSeek, Maxima, whatever, Quen. These are open models which are learning extremely fast. Then what you have in China, you have the adoption rate of technology is super fast. So Chinese people are really saying, 'All right, if there's a tool around the corner, I just deploy it and I embrace it.' They're also fast in their development cycles. Some of the new innovations which some German suppliers for the car industry they were supposed to go in a German car first. Now they go to a Chinese car because they are just faster in developing, adapting their car. Germans have their cycle and say, 'I can bring it to the next in 2 years.' China says, 'In the next 9 months it will go to a car.' So you go. So this is an advantage for China. United States has a huge advantage in just the capabilities, the quality of the people which you find here which are working on these foundation models or also new technologies. It's amazing. In particular in the IT sector, this is the superpower of United States. Less on manufacturing. If I really want to go for industrial manufacturing base, I would look to China first. 30% of the global industrial GDP comes from China. 20% from Europe and 15 from United States. So still a way to go. And there's another advantage. So it's a balance obviously what I'm saying here. Another advantage in China is if you want to produce something at low cost, that's the place to go. So in summary, I think the best you get if you get the hardware from China and the brain from United States.

Yeah, we try very hard. And we are for example, we are counting just in our space in our industrial space, we are counting something like 60,000 companies in the United States, 250,000 in China, another 60, 70,000, 80,000 in Europe. And these are companies which by definition if you total this amount of companies, I cannot serve with feet on the street. I cannot send a sales guy to bring them technology. So there's a reason why we created a platform called Siemens Accelerator. It's a marketplace. It's a marketplace where you can inform yourself about whatever kind of technology is available and you can transact. You can buy not only hardware or software, you can buy a whole stack, which is unique. There are not so many marketplaces. Actually, I don't know where the others are where I can really buy a complete stack of technology. And we put a lot of effort into that one in order to reach small and medium-sized companies. Here comes the problem, and these are twofold. One sits on us and one on our customers. On us, we have to develop technology which is easy to deploy and easy to use, so that it should be plug and play, which is difficult because it's such a diverse universe. And on the receiving end, we need certain capabilities of the small and medium-sized companies to digest new technologies. I can give you a small anecdote. When I'm traveling, I very often try to meet customers. I mean, it's either our distribution partners who are selling our products. And then it was roughly a year ago, I thought I would like to meet one of the small and medium-sized typical companies in the United States. And when you talk about small and medium-sized companies feeding the supply chain, you're ending up in either a milling or a molding because when you produce hardware, either it's metal or it's plastic, so you either mill or you mold. And it was a molding company, I don't know, 150 million revenue, and I was just asking whether I can visit these guys. They were wondering why I am going there, and they never saw a person like me, and then the whole management team was there. CEO, CFO, the two sons of the founder, the CIO, and so on. And they gave me a tour, and then we're sitting in the room, and everybody introduced himself, and then the CIO was saying, 'I'm the CIO of the company.' And I said, 'Oh, how big is your department?' Then he said, 'I am the department.' So here comes the thing: if I want to deploy latest greatest AI technology, this is hard. So he might have to double the head count. Yeah. Not the first which AI model to buy. And that's therefore it should be easy to deploy, but we also have to see how can we raise the education level of small and medium enterprises.

So we are watching that space intensively in two dimensions. Number one is we have a venture capital company N47. So we're spending roughly a billion over time horizon of 4 to 5 years invest in startups. Which is arms length. So it's not meant to be that whatever company we invest, we eventually want to buy. This is another, then I would give it to my M&A department. So this is really a venture capital company who's measured on their return. But it gives us insights into the startup space. What we do is for N47, we define the search field. So we don't want them to invest in a new burger hub. But it's more about technology fields. But other than that, we just let them go. And the advantages they see, I don't know, something 800 to 1,000 startups a year where they talk to these guys, they know where the money is going, what's the new trend, super important. And the other dimension is really that we are working with our R&D people, we're working with startups. We are trying to get their technology into our technology stack if it's relevant for us or we give them an opportunity to go with us to customers if it's complementary. And in some cases we are just sending our M&A department and see whether we should not buy. So it's super relevant and it's going so fast. I mean, today there can be a startup with one person or two and they make 100 million revenue in 9 months without having, in some cases they raise money like 100, 200 million without even having a product. They have just an idea. So that's very hot. I think it's a little bit overheated, but you know, some smart people will throw money on it. So let's see how it plays out. Portfolio strategy, I guess.