Dr. Andreas Liebl: Back Europe's AI frontrunners and create a pull effect for everyone else

appliedAI started in 2017 as a department inside UnternehmerTUM, the Munich innovation and entrepreneurship centre. Dr. Andreas Liebl had taken the concept to the EMEA leadership at Nvidia, to Google, to a string of other companies, and the answer kept coming back the same way. We need this, let's do it. 

Nine years later, appliedAI acts as two organisations that work towards the same goal with a different focus. appliedAI Initiative GmbH supports companies, from large corporations to large mid-sized enterprises, through their AI transformation, from strategy to AI solutions their teams can run and develop independently. The non-profit appliedAI Institute for Europe offers orientation in the age of AI, from learning about the AI Act, and makes knowledge more widely available to business, public administration, policymakers and society. 

Liebl is unusually blunt about the arithmetic. Europe does not have the compute. It does not have the electricity. It does not have the model providers. On his own team's projections, none of that changes inside three years. What interests him is the question that follows from accepting all of it. So what do we do now?

His answer starts with pace. Building world champions, he says, means not focusing on the slowest ones in order to keep everyone together. It means selecting those with the highest potential and giving them the support to get there. With 27 member states and 27 sets of strengths, he sees that as something Europe can work with rather than a problem to solve.

We spoke to him about the adoption gap, the four trajectories his team tracks, and why he thinks the hunger now sits with the younger economies rather than the established ones.

The Recursive: AI sovereignty is one of the biggest topics right now. What does it actually mean for Europe, and in 2026, how close or how far are we from it?

Dr. Andreas Liebl: To understand this, we developed something we call the Observatory. It maps growth trajectories of AI across a range of dimensions, and we think it gives us the best probability estimate of what is going to happen.

For all of those estimates, the picture for the next three years is the same. We do not have the compute capacity in Europe to keep up. We do not have the electricity to keep up. We do not have the model providers to keep up. Everything happening inside that three-year window happens under those conditions. There is no competition coming from Europe. Whatever we do is a factor of a hundred to a thousand smaller than what is in China and in the US. Even if you are ten times more productive, that still leaves a factor of at least ten.

So that is the framing condition, and the real question becomes how we can be sovereign inside it. Even if we decided tomorrow to build the next OpenAI and gave them a hundred billion, it would take one to two years before they were growing and had access to the compute. Possibly longer. So what do we do now?

We split the trajectories into four disciplines. In at least two of them we can still do something, although it will be hard.

The first is model capabilities, meaning the intelligence of individual models. The scaling laws still hold, more is better, and we cannot compete there. Maybe we talk about world models, what Yann LeCun is doing, what others are doing, and maybe we get to different architectures that are more efficient. But even then I would assume the US leads, and once something is there, it is globally there. It grows outside Europe, because there are simply more resources available to drive it. On the capability side, we have to take whatever is there. Luckily, there are enough open-source models that keep up.

The second is agentic orchestration, and this is the interesting one. It is the ability to build complex multi-agent systems that can do much more complex work. Think of the difference between a single human genius who develops a car, and an organisational structure where average humans work together to develop a car. The organisation manages it because intelligence lies in the structure: in the processes, in the roles, in how tasks get distributed. That is what we now need to do with AI systems. Not the single most intelligent system, but how to separate and split tasks so that organisations can produce very complex outcomes. That is the application side. We can be good at that. We can build these orchestrators, there is a lot of innovation happening, and we should want to be in that game.

The third is recursive self-improvement, and I doubt we get there, to be fair. There are a few players. Sakana in Japan, one or two in the US, and the large providers like Anthropic. It will obviously be a massive driver, but potentially one that sits outside Europe.

The fourth is robotics, where the Chinese companies are the major driver of innovation. There are a few in Europe, and maybe we compete on at least some components, and maybe on the infrastructure side.

So when we look at the major drivers of AI capability that are coming, the places Europe can play a role are the agentic orchestration layer and the coordination and infrastructure layers, with a little on robotics. The other dimensions we have to accept. We take what is there. And strategically, we should focus hard on model access, so that we continue to have access to these models until we genuinely have competing ones. Right now the general intelligence gap between the large providers here and there is roughly one year. Or we focus on niches, where the models we have in Europe can compete. But those are very specific niches.

If Europe cannot compete on model capability, can it compete on industrial adoption instead? 

Closing the adoption gap as far as possible is a legitimate European strategy. Even if you cannot provide the state-of-the-art technology, you can build the support system that keeps general industry as close as possible to what is technologically possible.

The gap itself is the thing to understand. The technology advancement curve is still exponential. The adoption curve of general industry using that technology looks more linear. There is not only a gap between the two. The gap widens, because the technology accelerates while adoption does not. That is the status of global adoption right now.

The layers where we can genuinely invest are concrete. Transform businesses, rethink processes, train people to use it. We have very well-trained people here.

But this is not a nice, cosy field where we can work one step after the other. Look at the speed at which China adopts this technology. There is hard competition here too. You really need to get things going if you want to be as fast as China. And I am not even talking about being faster.

There is the classic technology advancement curve, which is still exponential. Then there is the adoption curve of general industry using that technology, which looks more linear. There is not only a gap between the two. The gap widens, because the technology accelerates while adoption does not.

That is the status of global adoption right now, and the question is how we minimise this gap. Can we, in Europe, on a broad scale, enable organisations to sit as close as possible to the technology that is available? Transform businesses, rethink processes, train people to use it. We have very well-trained people here. There are a few layers where we can genuinely invest.

Closing the adoption gap as far as possible is a legitimate European strategy. Even if you cannot provide the state-of-the-art technology, you can build the support system that keeps general industry as close as possible to what is technologically possible. That is something we could do as Europeans, and something we need to work towards.

To be clear, though: if you look at the speed at which China adopts this technology, this is not a nice, cosy field where we can work one step after the other. There is hard competition here too. You really need to get things going if you want to be as fast as China. And I am not even talking about being faster.

China is one country. Europe is 27, all with different priorities. How does Europe align around a common vision and then execute on it at that kind of speed?

I like sports analogies, and they fit quite well here. In sport, as in AI, we need to talk about fitness levels, or maturity levels. If you take my general fitness and my ability to run a marathon without ever having trained for one, it would take half a year or a year before I could do it. If I were generally very fit, it would be much faster, because I have trained for other things.

What we need to do is train the general economy here to be more mature in applying AI. If that maturity is there, then when new technologies arrive we adapt quickly. If we are not fit, it takes several steps just to get to the level where we can begin to adopt.

So the question is how you train for the Olympics. You can do it the Chinese way, with a state-driven approach. Or you say, well, we have 27 different countries, we have different strengths in different fields, and we can manage this because we combine those strengths intelligently. That is what we have to learn in AI. What the strengths of the different regions are, how to combine them, where we build our own technologies and providers and support systems.

And building world champions also means not focusing on the slowest ones in order to keep everyone together. It means selecting those with the highest potential and giving them the support to get there. You can do that with 27 countries. It does not have to be one.

On a practical level, what actually grows AI skills and adoption in emerging markets like those in Central and Eastern Europe?

A general observation first, and this does not apply to every company. If you look at Chinese companies, they are hungry. They know this will benefit them, that it lets them compete globally. If you look at established companies here in Germany, many of them come from a leading position, and they now struggle with implementing AI because they are in the defender position and not in the attacker position.

It is what we always said coming from the entrepreneurship side: startups can only win, established companies can only lose. The incumbent has a brand, customers, employees. The founder has no brand, no customers, no product. They only work to win. It is a completely different risk profile.

You can apply that across a whole range of companies, and the useful move is to select the ones who say, okay, we can only win. In every country there will be some with that mindset. Find them, and then think about how to build these types of winners, because they are much more risk-loving. They try things out. They want to capture markets. Played intelligently, that is a real strength, and in those countries there are many younger companies that want to win, enter markets, drive things forward. The question is how to catalyse that energy and accelerate them onto the winning track.

In Germany, we are having discussions with the workers' unions where there is deep scepticism towards all types of change and transitions. So it is genuinely hard and slow to drive the transition that is coming, because they come from a very high standard and they fear losing it. When you are at the top, every change is potentially a change for the worse rather than a change for the better.

Everyone in the startup sector is already using AI heavily. But the broader economy, manufacturing, and the traditional industries are still very far behind.

Exactly, and that is why I said it is a general observation rather than something true of all companies. Now you have that situation. There are the traditional companies, and there are the younger ones.

What works for the established ones, to close that adoption gap, is matching with the startup environment. That can mean buying startups. It can mean hiring the talent. If a startup failed, get those people in. But generally the approach has to be: with this technology, we can actually win.

I do not know the national strategies of the various countries in detail, but that has to be the target picture at country level. We use this technology to attack, to win, and let's build that. China makes a very good narrative on it. Singapore has a good one too. I have not seen too many convincing narratives here in Europe.

Again, in the sports analogy: it is a national training programme. We want to train the winners of the next Olympics. We do not just build gyms and see what happens. That is not the winning approach. The winning approach is to take the companies where they are, identify their strengths, define the target, and systematically support them over a period of time in getting from here to there. It is a different thing entirely from building a bit of infrastructure and hoping. That very strategic, systematic build-up of company competence is, from my perspective, where Europe could actually win this race.

From your work with companies, what has genuinely changed for an organisation deploying AI in Europe?

We started in 2017, and in 2017 it was all about machine learning. The nice thing about machine learning was the clear value case behind everything. If I have a computer vision startup for quality control, I can talk to the people in the plants, discuss defect rates with them, and say that my solution reduces that rate, that it gives you this much value, so let's train the models. Training was the hard part, but then you deploy and you create value. A lot of products and startups were built around concrete value cases and concrete applications, and we were very good at that in Europe. Very particular solutions on the machine learning side.

Then GenAI came, and we had a hard time answering it, because it was a general technology for everyone. You call an API from Microsoft or Google or OpenAI, and everyone is using the same thing.

For companies, what was left was mostly organisational. How do you educate thousands of people to prompt correctly? How do you stop them from putting confidential data into private accounts? How do you deal with hallucinations? A lot of training, and very simple on the application side.

For startups, it was much harder, because there was nothing to differentiate on. You could do consulting services, or some technology on top, but the major platform providers made you outdated very quickly.

Now, with agents, we are entering a different era again, because we are rethinking processes. Process knowledge comes from domain knowledge. We are rethinking value creation in a way that cannot be done by the tech providers, because it is domain-specific and context-specific. So there is a reason again for startups, for the ecosystem, for the experts to build applications that differentiate from the general tech providers. Suddenly there is something we can build again, after three or four years in which it was very hard to do anything that properly differentiated a company.

What is new with agentic AI is the speed. Every two or three months a new technology comes out and we have to deal with it. So the question becomes: who is fastest at applying and translating new technology into my field? That translation capability is what we need to train people for, and we have to accept that things are constantly changing. That is where I see potential strengths, differentiators, and an ecosystem being created.

What is next for the appliedAI Institute over the next one to two years?

There are a couple of things where we can play a role. The first is using exactly what we just discussed. With agents, we could build marketing workflows that every startup with a marketing function could use. Or the same for software engineering, for HR, for finance. With fewer headcount, you can create a much larger company, which means you do not need as much growth capital. And growth capital is Europe's major disadvantage compared to the US, where there is plenty of money and you can grow extremely fast. What agents let us do is grow without that money. That could really change the overall game.

So appliedAI is now concentrating on building that agentic pipeline, starting with software development. If we provide that to the ecosystem, together with the training, because people need to get used to that way of working, we change something for Europe. That is one place where I think we can make a massive difference.

The second is that I honestly believe what is coming with AI will dramatically affect our societies. The education system, the tax system, the labour market, digital trust, cybersecurity. My feeling is that Europe is not prepared for what is about to come. We need to educate people. We need to tell politicians and others how society works with the capabilities this technology will have in the next one or two years. If we decide, if we shape, it can go in a very positive direction. If we do not, it can go in a very negative one. So let's do everything we can to move it in the most positive direction possible.

Anything you would want to leave our readers with regarding the European AI strategy?

In Europe, let's try to build on each other's strengths. The first time I talked with people at the European Commission, I asked them: What is the strategic plan for Europe? Do we want everyone to follow, in which case we are at the speed of the slowest? Or do we select those who really want to lead and drive things forward, and create that pull effect, in which case we are at the speed of the fastest?

With this speed of technology development, we should be focused on how to be fast. Help everyone else to follow, but do not wait for the ones who do not want to move, and do not settle at that very slow speed.

Every country, every ecosystem, every economy needs to decide the speed it wants to be at. And I really want to emphasise that everyone can define this for themselves. The field where they want to lead, where they want to be fast. Then focus the strengths there and be fast, because it is about the speed of application. That is the most relevant thing that counts. We talk about sovereignty across all these different topics, but the first question is: how can we be as fast as possible?