The factory of the future thinks along. It recognizes what is changing. It predicts what will happen next, and it helps people make better decisions faster. But what does that mean for production?
Welcome to Siemens Techmind 2026. Siemens CEO Roland Busch and Festo CEO Thomas Böck discuss the AI revolution in production — from intelligent machines and digital twins to the question of how humans and AI will work together in the future.
Take a look and experience how data becomes real production intelligence.
Please give a warm welcome to Roland Busch and Thomas Böck. A warm welcome. Come over to me.
So, right here in the middle, please.
So, before we really start talking about AI and the factory of the future, we want to marvel a bit — and we've already been marveling here at this flying jellyfish from Festo, Thomas Böck. Sometimes technology starts where you least expect it, namely with a jellyfish. How does that come about? What's this about?
Yes, learning from nature. We've been doing that for many, many years. That means we try to imitate nature, to replicate it technically, and we transfer the knowledge we gain from that into our machines and components.
And now the butterflies are coming along too. These are the Emotion Butterflies from Festo. So here, technology is taking a cue from nature.
You've been doing that for over 30 years now, in principle. And it does find its way into your applications.
So how come these butterflies don't collide? They're moving autonomously and collectively here in three-dimensional space. But why don't they crash into each other?
Well, they have the appropriate sensors and, yes, they naturally know the position of the other butterflies in space and can coordinate accordingly without colliding.
Roland, what's the fascinating thing for you about this technology?
That they don't collide. We still have to manage that first.
No, I mean, I think it's — I think it's truly genius to take cues from nature, because nature contains millions of years of evolution. We had a different topic earlier, in our pre-talk, about hands. So, we're all talking about robots, humanoids, and there are many problems to solve, but one of them, really the most difficult one, is the hand — building a tactile hand. And if you think about how robots are built today, they actually have all their strength in their joints — so a joint with a motor inside. We function completely differently, right? We have muscles. So, thinking about how you can design a hand differently — that's definitely worth doing. There are already many approaches, but learning from that and transferring it — not everything always becomes a product, but in many places you can learn a lot.
I find that interesting too. It's probably nice — if the kids in the room were here, they'd say, where can you buy those? That would be something for home, right? I want one. Thomas, what do these lightweight creations actually have to do with Festo? Why do you build something like that?
Yes, nature has truly managed, through evolution, to develop the most efficient drives. That means energy efficiency, miniaturization, compactness — that's exactly what we're looking for in our drives, in our components, and also the possibility of replicating these movements in industrial automation. And that's fascinating, and that's exactly what drives us.
Roland, how does technology fascinate you at Siemens? Siemens has stood for decades as a company that stands for engineering. Today, the core business is largely technologies that work in the background. How does technology fascinate you?
So, that's a good question. When we spun off Siemens Energy a few years ago, there was a really interesting moment when we said what the new company would be called, and it was called Siemens Energy. And then there was actually a pretty big moment — everyone, before that, everyone from Siemens Energy had their heads hanging, like, we're being sold now. And suddenly they were all cheerful. Why? Siemens Energy — the brand stays, and Energy says what it is. And then people said something I couldn't understand: we're the leftover, they thought. What could that be? Because then we're Siemens — what are we? So how do you get a purpose behind a name? And we wanted to avoid adding some description, because every description falls short. And then we said, that's how you add a purpose. Typically, you go with the big management team and write something down, and then you get ten sentences where everything is in there — nothing wrong, but nothing right either. And then many people said, no, the purpose is actually already there. You just have to put it into words. And whoever it was, it doesn't matter. Anyway, in such a conversation, out came: Siemens — we create technology to transform the everyday of everyone.
And you have to explain it — that's the annoying part — but then it becomes tangible. You just have to look behind the scenes. In principle, every car driving out there has probably seen Siemens technology in some way. Either the components were included with the software, or it was automated. Every third production line in the world runs with Siemens automation. How the second electron is distributed has somewhere run through Siemens technology — through our control systems, our software, how we do grid controls. And you can't walk down a block in New York without passing a building that is controlled by Siemens technology. And then it becomes tangible that we are everywhere. And without us, without technology, the world would probably look different. And that somehow makes you proud of what we do.
And at the head of the group stands a technician by training. Are you a physicist? You're an electrical engineer by training. So you bring that expertise into corporate management. But Roland, how has the One Tech Company that you are now developed in the direction of AI? Why do you believe AI is the future technology that must be bet on now?
So, the thing is, we're also a company called Siemens — not the only one, but the idea behind it was that we first wanted to position ourselves as a tech company. That's decisive, because technology is our DNA, that's where we differentiate ourselves. And the 'One' stands for it because we believe we have to work differently. We have to try to scale technology as broadly as possible as a technological platform. I also believe that Siemens' value is driven by our technology and, of course, by serving as many verticals as possible, because we can scale up. So that's the idea. Why is AI so decisive? AI is a key technology that will change everything. It already does, partly, today. But the way we develop products, how we manufacture products, how we operate products in the field — that will change completely. It's a key technology, similar to electricity — I've compared it — or steam. Steam took 60 years to change societies. Electricity, 30 — that was Siemens' coming out. Who would have thought what electricity could do at the time when people started building motors or light bulbs? The computer and the internet took 15 years. You notice it's always half, right? And now AI, we say 7, 8 years — three, three and a half have already passed. So it will take a few more years, and then we'll realize everything has changed. Now many people think, well, maybe it's a bit overhyped. But if you think back to when OpenAI first unleashed its model on humanity, and what has happened between then and now, you realize we're on a curve that isn't linear. Humans can't imagine exponential curves. We always work incrementally forward, and you only really notice you're on one when it takes off. I believe that we are on an exponential curve — you can already see it in the speed at which AI factories are being built, in the speed at which models are developing, in the speed with which, for example, robots — a manifestation of AI — are now being driven by AI instead of algorithms. And now there are companies ramping up from 0 to 100,000 robots a year. The world has never seen that. So all of that is proof, and we're right in the middle of it. One more point, even though it's gotten a bit longer: there are many labs in the world. The bet now, I believe, there are about 50 in the USA, 30 in China, and one and a half in Europe. So the world is already divided when it comes to foundation models. The new world — we're talking about Physical AI, Industrial AI — is just emerging, because the models aren't trained on industrial data at all. Design data is completely different from language. Time series data — the data that manufacturing delivers or that your machines deliver — the models aren't trained on that data yet. If you do that, you get into completely different spheres of how you can manufacture. We'll talk about that more — autonomous manufacturing and so on. That's actually the world in which we — and Siemens — can be in the lead. I believe that our ambition to be an industrial company, or to become one, that we can make that happen. We're not there yet. But we already have very, very good approaches. We need partners for it, we need data for it, and above all, an extremely high speed to market.
Thomas, how do you see that in a family business like yours? You think in traditions, you think in generations, and not for every family business is AI or KI at the top of the agenda. In which areas can you support your customers with AI? What are your application areas that are currently at the forefront?
So, we see first of all the opportunities that AI offers — AI in application — and we also see the chances we can work out for ourselves to be early, to be at the forefront. We come classically from movement. We solve all kinds of movement problems — electrically, pneumatically, with software, with AI — in every manufacturing company, but mainly on the shop floor. And what we can do — we know our limits too. We're not going to start developing world models or our own large platforms. That's what Siemens and others are for, and we're glad we have partnerships. What's in the foreground for us is movement as such, and intelligent movement — making it better and better, up to autonomous movement, so that machines teach themselves, machines optimize themselves according to very different criteria, mainly to save energy or to reduce overall operating costs. And that's our ambition, and we notice that with what we already have on the market, we are at the very front, and we'd like to keep that position.
Then let's look at a very concrete example, and it's right here. This is a welding gun, as there are many of them, especially in automotive production. What role does AI play here?
So first of all, we're not a manufacturer of welding guns, but we do the entire motion control. And because we know the motion control and thus the data space, we were able to use pre-trained AI algorithms here that monitor the movement in the normal case. That's many cycles that are run in normal operation, and then the system itself can decide when maintenance intervals are necessary. That's what we call predictive maintenance. At the same time, it's also true that these welding electrodes wear out over time. That means it's also about production quality, which we can monitor, and in the end it's also about energy use — if too much movement with too much energy is taking place. Whenever that's the case, it makes sense to plan a maintenance interval, and we can significantly extend maintenance intervals here because we know the machine much better. And you have to imagine, there are thousands of these welding guns in a system, they're all monitored simultaneously, and you get to planned maintenance intervals, save costs, and are considerably more productive overall, and can, for example, also prevent downtime — where you used to have to shut down the whole production first — and now you can proactively repair it.
Roland, that's interesting. So AI collects data and makes a forecast out of it. How does that change production for you at Siemens?
So I believe AI as such doesn't collect the data. AI processes the data. Collecting data is an exciting topic. There's an incredible amount of sense in going to the edge — really to the manufacturing sensors where the data is generated. By the way, most data today runs off into nowhere and isn't stored at all. And when it is stored, it lies around somewhere and is barely used. And then you also have to put an ontology behind it to really be able to use it. So just throwing AI at data only brings so much success — it only really becomes exciting when you do it with intelligence, with domain knowledge. And so this interplay of data, of industrial data and AI — but the third dimension is just as important, and that's where you come in with domain knowledge: knowing what it takes to have such a gun really productive and with quality. Those are the topics where you really get ahead with deep domain knowledge. Our customers help — of course our customers understand what they do, but really every expert who gets involved, how movements are organized and so on. And Siemens — again, we are involved behind the scenes in the entire value chain. We automate, we have sensors, we also manufacture ourselves, which means we generate our own data that we can use. We can also try out technology ourselves, and that puts us in a very, very good position to really bring industrial AI into use and to train models.
And there too, let's look at an example — namely how AI in production, more precisely in the area of packaging, can help. And we have a short film for you from the Erlangen plant F80. Take a look at what's happening there.
So, pick and place, I believe, is what this packaging process is called, and there the robot decides with the help of AI which package belongs in which box, how efficiently you can pack, and so on. What exactly is the added value for the customer in the end?
So in the end, hopefully, the customer gets what they ordered — no more, no less. I mean, in the end it's actually a pure productivity topic for us. That means we can then be extremely flexible — products simply come out, they have to be packed. Of course, saving space, they have to be packed, the right products have to go in — that's mostly a combination. And depending on, if in the end you deliver, for example, a package with various components, then the lot size is one. That means every time someone would have to stand there and think about what goes into this box now, and then a new one comes with a new order — what goes in there? And a computer, or a robot, or automation with AI can do that better. And in the end it has to fit into the box. So the robot is also playing Tetris. And these machines are sometimes already better than we are.
Thomas, at Festo there's a very similar idea with the Virtual Assistant. The engineer doesn't just get any answer from the AI, but context-based technical information, and is then guided to the appropriate documents and tools. What's the background on that? How exactly should one imagine this?
We had the approach of making the entry into the product world as easy as possible. We have a relatively large range of very different products, and it's not entirely easy to find your way around. In the classic catalog, we'd probably need weeks. Even in the online configurator, you have to know where to click. With the Virtual Assistant, I start a dialogue and end up even in the concrete designs. That means the engineering tools are called up directly from this assistant. So it's not a classic question-and-answer game where eventually you get an answer that you may or may not like, but I go directly into product selection, also into the pre-configured product design, and with that I can immediately — if I say I like what the AI has proposed — order the entire bill of materials.
So two companies with two very specific strengths. Roland, what happens when you bring these strengths together — those of Siemens and Festo? Where are the overlaps?
So we're trying that out now, right? We want to jointly shape the factory of the future — the autonomous factory of the future. And I think you've already noticed we come from very different corners. The topic of movement — that's what you do. You can do pneumatics better, or electric movement better than any other company in the world — the market leader. We come from automation, control, digital twins — digital twins of the entire factory. And the next thing, when you build such a factory, we first build it digitally and then in the real world before you do it. You're working more and more with our tools to build digital twins of your products, which by the way don't just look like the real one, but they also behave like it. That's the next step. Many companies can't do that — even tech firms that can represent many things very well lose the physics behind it. It looks good, but it's not the real world. And we're bringing that together, and that definitely has added value for the customer. That means the customer can now, if he wants to design, for example, a welding factory, design it once digitally with everything that goes with it — including the components, the actuators, the sensors — and then build it at the end. And it's really the complete stack we're talking about, the technology stack that starts from some sensor in manufacturing automation all the way to the cloud, where all the topics are worked out, and bringing that together and then realizing it in the industrial world. You really have to let that sink in. Industrial AI means bringing a non-deterministic technology — if you ask ChatGPT twice, you get two different answers — into a deterministic world. That means what happens there has to work, and it has to work to the point exactly. And that's the art — setting the framework so that it's safe, and repeatable, and works, and bringing that into the real world. I think we can do that quite well together with the technologies we have. And the application is infinitely broad — that's the nice thing about it. You work on... I don't know how many verticals we have — anyway, we meet in the market very, very well, and that's why it fits together quite well.
Yes, Thomas, you were the first, you said, to be part of the Siemens Accelerator. That's the digital platform — open, interoperable — where products, infrastructure, software, and services are all bound together into a single ecosystem. How does Festo benefit from that?
Yes, Siemens is broadly distributed in many places, and that's naturally also an opportunity for us to join this ecosystem, because many of the factory controls are based on Siemens, and we can dock from below via the motion control and place — but not just for Festo components, but also for any other components — exactly this intelligence on the platform, and can then deliver predictive maintenance or other hints for the productivity of the system there. We benefit from the data pool that Siemens has, and we combine that with the information we get from our data.
So, both companies are far ahead on the topic of AI, but if we widen our view, we know the competitors aren't sleeping, and they're strong competitors. Roland, the EU member states host only about 5% of global AI compute capacity, it says. Are we in Europe sufficiently prepared for this rapid, this exponential AI development?
Partly, partly — you get the feeling that Europe is doing everything to slow us down a bit more. Because if I think about how much time I've spent with my governmental affairs team and our network to defuse this AI Act so that it doesn't treat consumer AI and industrial AI the same — that took a lot of effort. Now we're in the second wave. Now it takes even more effort to defuse the Data Act in the same way, so that consumer data is treated differently from industrial data. So from that perspective, you could say, the tailwind looks different. The second thing is — and this is a legitimate question, and now I come back to the point — we're talking about data centers, but let's look at this AI stack once. It starts at the bottom, if you go really deep, with critical materials and GPUs — so basically breaking down the silicon, so to speak. We have nothing to offer there. It would be good if Europe — hopefully in Germany — a 3-nanometer fab would emerge, so that we have a certain independence. I think we're working on that. The next layer — I won't go into too much detail — the next one, we could start even lower, is the energy demand. Try today to get a permit within, say, three months for a new site that can connect a 1-gigafactory to the grid. It won't work. So that's the next topic. Then come the AI factories. We have to build more. I think right now we're relying on other models, on other data centers, also from hyperscalers that have invested in Europe. But it really makes sense to bring more and more compute capacity here, to Europe, for various reasons. The next layer, which is usually forgotten, is the layer that ultimately connects the semiconductors with the data. Middleware — that's almost all American. Can you rely on it? I don't know. But I don't see a big strategic lever there. The next are the models, the foundation models — I already talked about that earlier. It would also be good if we had some. There are a few approaches, but one thing mustn't be forgotten. Right now, you always have to spend, I don't know, 500 million or a billion, just to be number one again for 3 months. So that's an incredible burn rate with massive capital investment. We should really have more than one and a half in Europe. Maybe there are a few approaches that build world models — maybe we can still play along. And then comes the next layer, and that's the one where we can really have an advantage. That's industrial applications, industrial AI. There I would probably go full throttle. So if someone asks me in this whole stack where I would invest the money — must-win is industrial AI. We must not sleep on that, and then see what else we can build. But it would already be good if we had a bit more own value creation in Europe across the whole stack.
Thomas, you nodded in strong agreement. You see this issue very similarly, certainly. But let's turn to another topic for a moment. We've talked a lot about opportunities. The risks of AI — we have to speak about those. Are we able to control AI, and especially AI agents, in the long run? Do you share the concern of some experts who say we might need to slow down this whole development or tighten control overall?
Yes, you maybe have to look behind the scenes a bit. Some of that is marketing — also negative marketing to stir up interest, especially when there's an IPO behind it. Even those rogue AIs that hack other websites — those were actually security problems that hadn't been solved on the other side. The AI can find these known open vulnerabilities relatively quickly and very persistently, and eventually breaks through. So that's one story, and it's real, but it's not so dramatic that it's suddenly an intelligence performing supernatural things. I believe we still have limits there. Does it need to be regulated? Probably, especially when it comes to ethical boundaries or security questions. And exactly what Roland said — in industry we have machine directives. We know exactly what we have to do to make a machine safe. Now I don't have to additionally secure the AI, and that was exactly the discussion that was held over the past half year, also with the data — because if we regulate this now, we take away the only field where we can still win and that we can still defend for years. Especially because Germany still supplies 60 to 70 percent of the world's installed base of industrial automation. That means theoretically we'd have the data, we also have the knowledge of how factories work, and we must not let that be taken out of our hands. So it's not a one-sided answer. Security in that area, yes — but industrial AI in application, that's our domain, and it would be a shame to simply give that away.
How do you handle this topic at Siemens, and how do you see the fear of the super AI?
First of all, I fully agree with what you said. If I take one of the best models and incentivize it to crack a firewall — and wisely do it in a sandbox, an enclosed box — what will the model do? It will try day in, day out to get through this firewall. And what will happen? It will succeed. But if I then connect this firewall or this sandbox to the outside world, the probability that this AI simply keeps going is extremely high. So that's just careless behavior. I compare it to when we approve a train — we're also responsible for the train meeting safety standards, right? And there, something simply happened that shouldn't have happened. So I think it's exaggerated, but still, you have to be careful. And the example of ethical topics — it absolutely makes sense that if you type in, 'AI, help me build a chemical weapon,' the answer should be, 'I won't do that.' Such guardrails are definitely needed, and it has to be regulated. We tend in Europe to over-regulate so that nothing can happen at all, and thereby completely slow ourselves down. So I don't have that concern. It may well be that eventually models learn to learn models — right now, for example, with Anthropic, they're developing to 100% AI-based. AI models are developing AI models. So that becomes a loop that reinforces itself. And you can then wonder how fast such a model can develop, what could happen there. It's breathtaking, but it also still has to be fed, by the way, with all the data centers that go with it. So there are other limitations. I believe you have to be careful to treat the technology properly, so that you regulate it in a certain way. But braking it, or these doom scenarios — that the probability of AI wiping out humanity is at x% — I think that's pretty exaggerated. How do we handle it? We try to put all our applications into a box, so that data doesn't go out, that data we want to keep here, and that no outside models are trained on it. We have a very, very good security team. When we have an industrial application in manufacturing, we handle it exactly as it is already regulated today. Software is already regulated for the markets, and we also have an inherent responsibility that when we go into industry with our customers, it actually works.
So, if in the human-machine interaction, Thomas, the human is supposed to keep control — how do you enable an organization to deal properly with AI? How do you approach the whole topic of employee training, further education, and so on?
Yes, on different levels. There are of course the specialists who can develop and train these models. That's probably a whole league of its own, but it starts at the base — how do we bring reading and writing, these literacies as we call them, into the broad mass of the workforce at all? We've had awareness months, we've launched hackathons, we've had agents developed from the application base — and not by IT, but by the users themselves, so that they can inspire each other and lift each other up in this world of development. We have, similar to Siemens, a kind of sandbox — we call it Festo GPT — where we've released different large language models, and depending on the use case, whether it's in the direction of software development, coding, or in the direction of automating data, for example, internal company processes, we can use different applications. And it's developing, and we try to keep fostering and fueling the healthy curiosity that prevails in the company, and the communities keep growing.
So now I'll take a short break at this point and would like to give you the opportunity to ask questions. And we'll kick off with two young talents. And we'll start with Bianca Bacher. This here is a microphone that I'm going to throw over to Bianca. I hope we manage it. Bianca — oops. Yes, I'll do it from a sitting position. It's difficult. So Bianca Bacher is a production planner at F80 at Siemens for the new G220 line, an industrial engineer. And what is her question?
So at our plant we already have many useful AI applications. One of them we just saw, right on the shop floor. How do we manage to scale these solutions to other plants and also make them usable for customers?
Yes, so that's exactly what we want to do. Pilot it internally, because we can use the technology ourselves — Siemens as Customer Zero — and then scale up. Now you have to know, bringing a product requires more than just an AI that works in one spot. Productizing is also a promise to customers that you can not only sell it, but that you also get updates, security updates, patches, a roadmap behind it, and so on. So the step from 'this works for us' to a product you can sell is bigger than you'd think. And ultimately it's about putting a product management behind it that first checks: is the use case scalable? And it definitely is. That's why it was so smart to start with exactly such a use case. So is it scalable — then how can you productize it? Meaning, how can I now take this AI that runs in a certain environment for us — it has to abstract a bit, it should also run in various other environments. It's actually more of a software solution, an application. That's the second step, productizing it. The third would be: what does it need in terms of training? How can I deploy — meaning install it at a customer? Does it need to be trained, or does it practically work by just saying, here's the link, let the software run, and it works. So you see, it's a bit more complex, and then you have to hand it to sales, who also sells it. By the way, that's where Siemens Xcelerator comes in, because if it's really pure software, the customer can really buy it digitally. And now comes the really exciting part about Siemens Xcelerator. It's a platform that doesn't exist in this form yet. That's why it also took us a bit longer until it really works, because you can then buy hardware and software with one click, with all the different...
…contract terms and conditions and so on. One thing goes online, the other has to be provided logistically. So we've reached the point where this works now. We have this platform, and with it you can scale such solutions up. But the path there — that was a long answer to a short question, but I just wanted to show once more how complex it actually is to build such a product.
Now let's look at the practice at Festo, where we have Jan Seiler, Head of Advanced Development Analytics and Control. Mr. Seiler, what is your question?
Yes, I have a slightly longer question — I'd like to look a bit into the future. You see a lot of mashups of games right now. I don't know whether you've noticed, but Mario Kart in Modern Warfare, Nintendo 64, Mario 64 in Elden Ring and so on. And that's very, very funny. But what's fascinating about it is that the whole game has been reverse engineered, right? The entire game has to be rewritten in order to bring two games together. And as a result, the value proposition of software is currently eroding extremely. I can replicate software extremely quickly. And my question now goes in the direction of: how do we manage to protect that when it comes to hardware? When we have digital twins that are essentially a specification that keeps getting better, that represents the physics better and better — it was mentioned earlier — if we can use them as a test bench, if we can integrate them, then copying is no longer expensive. Copying used to be expensive. Reverse engineering took a lot of time, cost a lot of money, and was therefore an intrinsic protective factor. And that is eroding more and more. And that's exactly where I'd be interested in your opinion: how do you protect yourself against that?
Yes, exactly, Thomas. A super interesting question — but I would have expected nothing less from Mr. Seiler.
