And a very good afternoon, ladies and gentlemen. I'm Amena Baker and I'm the head of Middle East insights for Kepler, based here in the UAE. It's my absolute pleasure and delight to be moderating this session today. We'll be talking about a very exciting and timely topic, very popular these days. It's building intelligent and resilient, sustainable industries using AI. So I have a stellar panel here to kind of discuss this topic and a little bit more if we want to delve into other topics. Sitting next to me is Girish Saligram, president and CEO of Weatherford. We have Roland Busch, president and CEO of Siemens AG, and Uljan Shakra, the founder and CEO of Domian. Gentlemen, welcome. Girish, I want to start with you and to talk a little bit about AI applications. We hear a lot about AI being integrated in operations, integrated in fields, for example, to improve efficiency, to increase resilience. How does this work? Is it measurable? Is it worth it?
Amina, first of all, thanks so much for having me. It's a pleasure to be back at ADC. Look, I think to state the obvious and what everyone's feeling, AI is the most profound change that I think all of us are experiencing in our lifetimes. And it's not just a technological change. It's a societal revolution. It's changing pretty much every aspect of what we do. So, from an energy perspective, an oil and gas standpoint, we are seeing some tremendous applications across the board. What is exciting about it though is I think we're still very much in the infancy of this. So we're seeing a few things come to fruition. We're being able to tangibly measure them, but we've still got a long way to go and I think that's what's exciting about it. So we're seeing applications in drilling, for example, where the ability to predict exactly what you're going to see when you're drilling, you know, problems you're going to encounter and how to mitigate those, where exactly to drill, reinterpretation of seismic data. All of that is being enabled through AI and giving tangible results for operators, for service companies like ourselves.
We are seeing a tremendous potential in the design of products and services. How do you actually do something that used to take multiple weeks, months, sometimes years? That time frame is getting condensed enormously. We're seeing a tremendous application from a production standpoint, especially around production optimization. So the amount of data we have collected over decades with a variety of different fields feeding that into algorithms. We have, as Weatherford, we've got different offerings that we provide to our customers that have integrated AI models built into them. We are providing them in our service operations things that used to take four to eight people on a rig are now being done autonomously. So tremendous applications across the space which are providing productivity and better production outcomes for customers along with what every other company around the world will eventually get to, which is a better way to run the organization itself.
Girish, I think it was a line from a movie where they said, before you revolutionize my industry, understand my industry. So do you feel like the AI partners that you work with understand your industry?
Well, the good news is that's why we exist as well, right? So I think it's a combination. They bring a tremendous amount of technological know-how, but we bring the domain expertise, and I think the combination of that is incredibly powerful.
Roland, I want to come to you and talk a little bit about pilots because when it comes to AI, we always hear about pilot projects. So how do you move from that kind of pilot stage to expanding that further and, you know, just running it across different applications?
That's a good point, and let me double click on what you said. If it comes to industrial AI, so the application of AI in the industrial world, we are just at the beginning. Is it in the design, in the manufacturing process, or the way how we run assets? And this is the reason is because we have very strong models which are trained on language, but the models didn't see the industrial data yet. And this is a completely different language. I mean, design data, they do not look alike any kind of language, time series data, very special animal, visual data on inspection data which coming from your operation, from your shop floor. So therefore, that's about to come. So this is the first one that not before you train models depending on whether you really want to put a foundational model in place and a second, then you have to go deep, but you can also tune a model and make it work for an application. So not before you have that really done, then you can start working on it.
Then comes the other point. What you have to do is, let's take in your operations or in the asset management, if you put a co-pilot on top, it stays shallow. You really have to rethink your processes deeply, change them, and then throw AI on top of it so that it's naturally coming. So even people don't recognize that they use AI, but it's naturally coming. This is when it starts scaling because then you really get the benefit. Before that, you just argue, I mean, now I spend more tokens, get a little bit more outcome, and finally it's a net zero sum. So this is the other one. And then very often it comes also from the leadership. Companies where the leadership says, I want to go all in, they are much faster in scaling and using it. And I can give you one example from, I think we can release it, from PepsiCo when they were asking us to improve a line and our guys in four to six weeks, they drove productivity by 20%. And then the CTO, she says, okay, now I want to roll it out to all my plants because that makes sense. It's a commitment from the top and then you work on it, obviously you improve it as you go.
You make it better and better and then you can roll it out and scale it. So that happens more and more, but it's a combination of having good training of the models, having the data available. By the way, this is also sometimes very, very difficult. Then if you have a proof of concept, you prepared all the data and said, now roll it out, and they're running into other problems because they don't have the data in the other plant, it's not there yet. So bits and pieces are coming together, but we get there.
Correct me if I'm wrong, but why are companies still hesitant to share data related to how much efficiency, for example, they've gained from applying AI, or do you think this is going to change?
So I mean, sharing data on the efficiency you gain is, I think nobody would dare sharing that. I mean, that's, we have another problem. The problem is about, so now come back to my point, when you want to create an industrial foundation model which is similar like an LLM where you can say you can throw an LLM on many, many different tasks and you get a somehow reasonable answer, an industrial foundation model would do the same. It would, so for example, you're training one on mechanical engineering. So what it does, it really allows you to, the model to really create a design of a new product based on your requirements which you can really build. It does not look good, you can't really build it, and that's something where you have to really go deep to do that. You needed to really share the core engineering data of different designs and then train a model on it. Here you see the concerns. I mean, who would share, would you share your design data with me?
Yeah, right. Or at least it's a perceived edge.
So I mean, no, in order to encourage that, first thing, we have a lot of data on our own. That's one thing. The other one is you not necessarily have to get the data on your servers. You can also train a model on your data. So I don't have to own them. And the third one is, I don't need the latest data of your latest design. If you have others which are already on the market which are eventually already re-engineered and I take all the generations before, I rather learn from the delta between one design to the other. Why did an engineer change? This is where I make my learning rather than from the latest design. So, but it's always a concern that you give IPR away once you open up your data.
Thanks, Roland. Uljan, I'm going to come to you now and talk about something that, I mean, when we say AI, there's a lot of talk about the sovereignty central to that and how companies could make sure that they have sovereignty. But what does that really mean in practice?
So to Roland's point, this is a lot about protecting your IP. It's more about owning the data than anything else because we have this concept of AI being software. Okay? But actually AI is data. So when, for example, we bake our data into an AI model, we are exporting that information and we cannot get it back. It's a bit like putting sugar into a cake, you bake it, you cannot extract it.
So this change of software going from physics where you have two separate objects that you can legally contract and, you know, as an owner of the data, you can use some software, terminate the contract and then go to another one. You cannot do that with AI anymore. So it challenges actually how law applies. So the concept of sovereignty is not a concept of borders control like it is often being mentioned, but it's more about how can one company protect its own data and control its own future, especially with these models becoming a bit like electricity for software. So if software is the operating system of an organization, whether that is public or private, so this applies to governments, it applies to industrial AI, it applies to banks, mission-critical infrastructure. So if that software is the operating system, we can think then of AI as electricity. So we can build our best, you know, products and factories, but if we externalize the electricity, then we can be shut down and remote controlled by somebody else. Now, I'm not saying we should talk about alliances here and, you know, there's a lot of conversation around Europe versus US and China. I think it's more, this is more of a strategic autonomy play where every organization should control its own future. So it doesn't matter if your dependency comes from Europe to US or China. If you don't control it, your company's at risk. So sovereign AI is about vertically integrating models, agents and infrastructure so that you can control your own company.
So it's possible to control it. It's possible to protect, let's say, the sovereignty of... And because I've heard also again the opinion that once you just feed this beast, there is no way you could control, you know, how it uses it or how it teaches other agents.
Unfortunately, we have designed as an industry the wrong business model, and at Domian we're changing this over and over time. So for companies to actually be sovereign, you need to own the models.
And training these models, it's much more accessible than you can think. For example, we released a 4 billion parameter LLM last week which was trained on a public European supercomputer. It took 45 days and it's an edge model for industrial AI that, you know, Siemens, for example, can download for free on Hugging Face and can train and tune for their own use cases. But within open, just the weights. So there is a concept of openness in these models that is very opaque. Opening the weights, which means having a copy of the model, doesn't mean you have an open model because you need also the code, the data sets. You need full transparency for a company to really trust it. So we have two aspects here. One is we need open models for sovereign AI to take place.
And the second one is we need a commercially sustainable business model which doesn't necessarily mean that we need to have four or five companies to subsidize intelligence for the whole market, otherwise they will become the market, they will eat the market...
They'll be a monopoly, yeah.
