Member of the Managing Board, Chief Technology Officer and Chief Strategy Officer, Siemens
ET in Dialogue - Peter Korte CTO Siemens India
Fireside chat of Peter Korte CTO Siemens India with ET Editor on Industrial AI.
About Peter Koerte
Peter Koerte, Chief Technology and Strategy Officer at Siemens, participated in a panel discussion on March 30, 2026, alongside KION CTO Rob Smith and OHB CTO Christina Wagner, focusing on industrial AI, digital twins, and data ecosystems. Koerte stated that the major difference between industrial AI and consumer-oriented large language models is the data challenge, noting that industrial data, unlike language found on the internet, is not publicly available. He argued that no single company has enough data to build the next generation of frontier models for the physical world, and that the only way forward is to build open data ecosystems through partnerships where companies exchange data while protecting intellectual property. Koerte described the partnership with KION and OHB as a way to combine domain expertise and datasets, with Siemens helping to encapsulate that knowledge in models that can be scaled universally. He expressed the belief that Europe has a global advantage in domain knowledge, and that by combining it with trusted partners to build an industrial foundation model, the region can scale it faster than others. Koerte also noted that space and aerospace companies face challenges in designing systems for flexibility and making them software-defined and AI-capable, as assets in deep space are not easily adaptable.
Transcript (52 segments)
N
Narrator0:10
Transform Siemens' Innovation Day 2026 is Siemens' flagship innovation forum in India. This forum brings senior leadership together with customers, ecosystem partners, media and analysts to explore how emerging technologies, especially industrial AI, are reshaping industry, infrastructure and mobility.
I
Interviewer0:30
My guest today is Dr. Peter Koerte. Dr. Koerte has held a range of business leadership roles in Siemens. He is currently member of the managing board of Siemens AG, the chief technology officer and the chief strategy officer. He's been a frequent visitor to India and this is his third visit. So I guess my first question is: what's the difference between industrial AI and the regular AI that we use?
P
Peter Koerte0:55
Yeah, there's many differences but let me pick three. The first one is industrial AI where it's being used. Obviously it's not in the consumer space but in the industrial space where we talk a lot about physical AI, meaning making the physical infrastructure better, more productive, more resilient. The second one is that the AI we're using very often is mission critical, it needs to be very safe, reliable, trustworthy. It must not hallucinate. The third one is the way we develop this AI because industrial AI is not usually based on words and language that you can train on by going to the internet and downloading web pages. Actually it's trained on engineering data, production data, and usually you don't find that easily available. So the way you develop these models is very different than in the consumer space.
I
Interviewer1:50
So Dr. Chris Seammon says it's powering the industrial AI revolution. I mean this is all good but where are your customers seeing results first?
P
Peter Koerte1:57
Yes. The number one thing where they see this is usually in energy and energy consumption and where we can cut a significant amount of energy that's being consumed. One of the most obvious ones for that are the energy heavy industries such as buildings as we're sitting in here right now.
I
Interviewer2:17
So it turns out 30% of all the energy worldwide is consumed in buildings because we humans spend 90% of our time in here.
P
Peter Koerte2:24
And so we want to feel comfortable as we do right now and there's a lot of energy wasted today that goes into cooling of buildings. We've developed an application or an AI that is called Comfort AI that takes a reading every 50 minutes, learns, updates itself and cuts energy cost by a third and it's very easy to deploy. That's a great example of how we make infrastructure more intelligent thanks to AI. And that's not limited to buildings. You can go into cement industries where we look at energy and where we can cut it. So usually this is very stark but it's also in industries where time to market is very important and where we can help to significantly accelerate the development and the production cycle.
I
Interviewer3:14
Okay. So you spoke about buildings and manufacturing in your example of cement. One of the things I imagine you need for energy savings is AI scale at deployment at scale. Give us a sense of how Siemens is planning for this kind of scale that industrial AI is going to use.
P
Peter Koerte3:26
Yes.
I
Interviewer3:26
Give us a sense of how Siemens is planning for this kind of scale that industrial AI is going to use.
P
Peter Koerte3:31
Yeah. So the AI that we provide is usually through a platform called Siemens Accelerator. Think of it as a portfolio of software that enables our customers to achieve their results. This is where the AI is embedded. This is where we would be working with applications that are all AI powered and that help us to design better ways of how to produce cement or how to operate cement mills or buildings. For example, here we got a building. The application is called Building X and it's part of a SaaS platform where the AI is embedded and has algorithms to optimize operations.
I
Interviewer4:22
Dr. Curt, you describe an ambition to build an AI operating system for the industry. So it's not an after the fact insertion. It's the very core of the industry itself if it's an OS. So if industrial AI is the OS of the industry, what are the building blocks of this OS?
P
Peter Koerte4:39
Yeah, the building blocks are essentially lifeblood of that is having access to data. If you think about it today, if you go into a factory, in particular a brownfield factory that's been there for many years, usually there's very little data because these machines are manual, mechanical, and there are no sensors. So the first step is: do you have enough data to understand the shop floor? Can you digitize that? Very often we apply sensors, vibration sensors that tell us about the condition of the machine or machine vision that looks at material flow. Once you have connected all the sensors, then the question is aggregation of that data into a data layer with its own ontologies and semantics. Then comes the model layer where we put the model on top. Usually large language models but not always; sometimes traditional machine learning. Then you have the reasoning layer that makes sense of the data based on a human machine interface. These are the things that have to happen. At Siemens we believe if you want to be successful in the industrial world you have to have the whole stack. You cannot just do one layer. The key word is domain knowledge. You need to understand which data you need.
I
Interviewer6:31
So you need to understand which data I need.
P
Peter Koorte6:34
I give you an example. Trains, we also produce trains. On the train you would think maybe it's the propulsion systems or the brakes or whatever. But it turns out it's the doors because the job of a train is to transport people. It has to stop, open doors, people go in, doors shut, train goes to next station, open again. Very often these doors fail and are major cause for delays. Today we connect to all of these doors. We put sensors into the motors that drive the doors and we can see on the voltage curves whether the motor degrades and we can tell you 10 days in advance whether they're going to break. So these are very good examples that show you need to understand the domain because the use case drives the kind of data you need and that drives the application.
I
Interviewer7:37
So these are complex layers. The sensor layer, the data layer, the reasoning layer, maybe a dashboard layer as a consumer interface. I imagine one of the most important things is that all these layers stick together and work together. How do you make this happen? How do you make this into one coherent architecture?
P
Peter Koerte7:50
Yeah.
I
Interviewer7:50
How do you make this happen? How do you make this into one coherent architecture?
P
Peter Koerte7:53
It is by having pretty much as you outlined it's building blocks. I think the word exactly the right word is we tend to think of microservices. You have a microservice that does the ingestion of data, a microservice that ensures the data lake is accessible and stored, a microservice that ensures data security with regards to cyber security. All these are components that we stitch together. They are part of what I call Siemens Accelerator where you have microservices. We build them together because we reuse them over and over again. You need to have a data layer no matter if it's buildings or cement mill. However, the environments are very different but the general logic is pretty much the same.
I
Interviewer8:42
Dr. Curt, digital twins are very exciting. It's been the stuff of science fiction. We've seen spaceships and planets and digital twins of that. You've spoken about the need to evolve digital twins from passive simulations to active intelligence. So question number one, how do you make a digital twin actively intelligent?
P
Peter Koerte8:57
So the idea and this is in the process where we are. In the past, how was a digital twin created? Usually created by greenfield applications. You wanted to build a new building, new machine, car, plane. You had the digital replica first, the twin, and then the physical part. They coexisted next to each other and you had versioning issues because one of the two would change and they would come out of sync. What we see now is that the two worlds, the digital twin and the real object, are connected because of data and sensors. You have a continuous update. With that you can take a factory as an example. You always know almost in real time the condition of the machine, where the material is, where the worker is, what job is being manufactured. Now the AI can do time travel. It can go backwards and forwards by simulating what may happen. It can go backwards and show all failures with material shortages and predict that at station 8 the screws are getting low, raise an alarm. That's the predictive power. You can react or manually intervene. For example, semiconductor shortage a couple of years ago. Now you can run what-if scenarios: if this type of material is missing, can I still run the factory? Can I run something different based on the material at hand? All AI powered in real time. Thereby you have higher throughput, higher uptime, higher productivity.
I
Interviewer11:31
Dr. K, I'm going to ask you to wear your customer hat. Assume you're a Siemens customer. What would be your biggest challenge here? Data readiness integration or proof of ROI?
P
Peter Koerte11:40
You don't start on it if you don't think there's an ROI. If you don't see this clearly, you should not start on it. Proof of ROI is a good point. As Siemens, we would tell you that we expect you to save a third of your energy cost on the building example. That's the key driver. As a customer, I would trust Siemens that you know what you're talking about. The hardest part is usually the connectivity. It's really to connect your real world to the digital world. Very often a technician is missing who is able to translate the IT world to the operation technology world. So usually it's the data availability.
I
Interviewer12:42
I want to ask you about the digital twin composer. Now this is positioning integrating digital twin data, simulation, and real-time operating signals. These are three distinct pieces of information. What's the real technical breakthrough in integrating all of this together?
P
Peter Koerte12:57
Actually surprise it's the AI.
I
Interviewer13:01
Okay.
P
Peter Koerte13:02
And the reason is because why we haven't seen it beforehand was the sheer amount of data in real time to be processed was simply not possible even by machine learning standards. It's compute power plus AI. We've seen a significant acceleration of compute with GPUs, parallel computing is now more the norm. We can handle all of this in real time. That's where the true value comes. If you have to wait 10 minutes for a what-if scenario, that's a bottleneck. Now it's possible.
I
Interviewer13:19
It's compute power plus AI. Exactly that.
P
Peter Koerte13:21
And fair enough, we've seen a significant acceleration of compute with GPUs, parallel computing is now more the norm. We can handle all of this in real time. That's where the true value comes. If you have to wait 10 minutes for a what-if scenario, that's a bottleneck. Now it's possible.
I
Interviewer13:46
There's a phrase that you seem to love: a managed high-fidelity environment. What does this actually mean? It sounds like jargon to me, but what does it actually mean?
P
Peter Koerte13:52
A managed high fidelity environment. High fidelity means that the difference between consumer AI and industrial AI. I said it needs to be reliable. High fidelity means that you actually trust that whatever I told you is happening in real life. High fidelity is confidence that it's true. You only can do this if you know the physics behind it. It's a more deterministic way. AI is probabilistic. The hard part for engineers is to have probabilistic models that are almost as high fidelity as deterministic models. Short answer: more confidence.
I
Interviewer14:57
Okay. So the digital equivalent in real life, you can be pretty sure by looking at the digital equivalent that this is what is happening in real life. Right. And you can manage both those realities as it...
P
Peter Koerte15:08
Absolutely, absolutely.
I
Interviewer15:09
All right. So one of the most exciting things about a digital twin from a consumer point of view is photorealistic visualization. Now with the Teamcenter digital reality viewer and I believe it's powered by the Nvidia Omniverse and paired with your composer with these three components, what is now possible beyond photorealistic visualization?
P
Peter Koerte15:16
True.
The importance of photorealistic visualization is for one the way we humans interact with that world. The more natural it feels, the better. I would not underestimate the human effect of having a photorealistic representation of a factory. It feels natural and intuitive. In the past it was clunky, pixelated. But the real value is that we retrofit a lot of workstations with machine vision. That machine vision needs to be trained on picture data. Very often we are able to create that picture data synthetically. If you have a photorealistic representation of this room, I could simulate the sun shining, monsoon coming, shading changes. If we had a camera sitting there, I can create all these different images based on different illuminations and train the camera to work in every environment. That's the real benefit: you can create a lot of synthetic data to feed your vision systems and make them much faster.
I
Interviewer17:13
Dr. Ker, there are multiple things you're talking about: collaboration, engineering validation life cycle, insights, closed loop execution. Could you put this in some sort of priority for us? What's the biggest value here?
P
Peter Koerte17:24
The biggest value and the biggest challenge is closed loop. Closed loop means exactly what we talked about earlier: you always have the two worlds in sync, knowing of each other. With that you can make real-time decisions and enable seamless productions. But that's a long shot because it's a lot of infrastructure, data, and expertise. Usually you start somewhere along those chains. First you create a digital twin and do offline simulation. Then more sophisticated simulation: computational fluid dynamics, acoustics, electromagnetics. All of a sudden you can build twins that behave like the real world with high fidelity. Your confidence level rises. Then you feed operations data and build the closed loops. The loop is usually the last thing you do. It's the highest art.
I
Interviewer18:51
At Transform Innovation Day you're speaking about AI factories. First let me ask you: what is an AI factory and how does it look different from a regular factory? You spoke about cement factories a little while ago. So I would ask the question slightly differently because what's the difference between an AI factory and a data center?
P
Peter Koerte19:11
If you allow me that comparison, it becomes obvious. A data center today has many aisles with server racks and cooling. Air cooling. With the next generation of GPUs, power density is so high that air cooling won't work. White space (racks) and gray space (cooling) blur. You have to have liquid cooling. As engineers, we get scared when liquid enters electronics. But that's what we're doing: we immerse racks in liquid or use cold plates to remove thermal load efficiently. The boundaries blur. They feel like factories with plumbing and liquids moving. You build them out in modular ways conceptually: power goes in (material), input is the prompt, output is the answer from LLMs. They operate like a factory.
I
Interviewer21:05
I'm going to ask you a couple of India specific questions. Now as you well know, this is your third trip to India. India has been scaling up compute pretty rapidly. What do you think is going to be India's biggest constraint in this compute scale up? Is it going to be GPU availability?
P
Peter Koerte21:19
Usually it's energy and grid connection. You need to build new capacities. When a data center comes up, you also have to build energy sources and the grid next to it. The good news is that's happening in India as well. Access to chips is certainly an issue, everyone is struggling with that. Today India has 1.6 GW of installed capacity, which is not a lot globally. They want to get to 8 GW in 2030, which is also not a lot. The bigger question is why is it so little? I think it's because of demand. The manufacturing sector in India is only about 15% of GDP. The government wants to get to 25%. As that share rises, demand for AI will increase. Then you need more data centers. The projection of 8 GW in 2030 might be on the lower end, depending on the pace of manufacturing sector build-out.
I
Interviewer23:06
That brings me to my last two questions. I'm going to use a Siemens term: grid to GPU to factory flow. What does this phrase mean and how do you pull it off in this technically integrated orchestrated way?
P
Peter Koerte23:12
Mhm.
I
Interviewer23:12
What does this phrase mean and how do you pull it off in this technically integrated orchestrated way?
P
Peter Koerte23:17
Essentially as Siemens we electrify, automate, and digitalize across 30 different industries. Specifically for data centers: from grid to chip means we look at the energy that comes through the electricity grid. We help data centers determine where to put their data centers in the first place. 90% of all transmission grids in India are designed using Siemens software called PSSE. We help operators pick the right spot. Next step: build the grid connection, simulate and calculate how it works. Then bring electricity from high voltage to medium to low voltage all the way to the chip. In the future it will be direct current for efficiency. Then you scale many times to become the factory floor, the AI factory level.
I
Interviewer24:38
Okay Dr. Curt, final question. To a note of caution though on the final question. Now if you're successful and we hope that you are incredibly successful, AI is going to get embedded in every piece of industrial and infrastructure. It's going to get embedded. Does this pose any threats and what guardrails do we need? What safety guardrails do we need?
P
Peter Koerte24:59
So it will be embedded. There are many ways to go about this. First, it always has to be designed for cyber security. At our conference there's enormous emphasis on cyber security, a concept called zero trust: segregate networks, authenticate all the time, encryption at rest and in transit. We design solutions with cyber security in mind. It's one of five key requirements. Second, we think about AI as going to do everything potentially, but I don't think so, at least not in the next five years. AI will help us as humans, but there's still going to be a human in the loop. The word is agentic. An agent is many AI applications working in concert to develop a job. For example, updating a database based on material data. Once the job is done, a human will check if that tedious work is correct. That goes back to high fidelity conversation. We should trust the AI to some extent, but trust and verify by a human. We'll see how quickly this evolves. If AI evolves quickly, we'll see higher productivity. But we will always have somebody running the AI. The idea of a dark factory with autonomous systems is the ultimate vision, but it's going to take many steps to get there.
I
Interviewer27:18
All right, Dr. Curt, final question. What's your favorite science fiction film? Something AI related or Fritz Lang's Metropolis old school?
P
Peter Koerte27:26
No. Odyssey 2001. I think it's definitely a visionary film. I like that very much. But I also like The Matrix in a way because it's kind of a digital film. There are a lot of good movies that inspire in terms of what's going to happen in the future.
I
Interviewer27:46
All right, Dr. Peter Carter, thank you for talking to us at the Economic Times.
P
Peter Koerte27:50
Well, thank you for having me.