Paperjam: Prime minister Luc Frieden announced, on stage, a partnership with Mistral AI. Can you tell us more?

Arthur Mensch: This is a multi-year partnership, because there is a lot of work to be done on many government services. It includes both a research dimension and an operational dimension, with the equipping of public servants. In practical terms, our aim is to use our technology to improve public services. Numerous initiatives will be announced as they are implemented. As an example, think of what we have done to equip the agents of France Travail [editor’s note: the equivalent of Adem]. We have deployed AI agents who interact with jobseekers, talking to them and helping them to identify the opportunities that are right for them. We have also touched on the fields of research and defence: an agreement with the defence ministry will also be announced.

To do what?

For internal equipment and productivity initially, but overall there’s still a lot to do.

How does this type of partnership make sense in the development of your business?

What sets us apart from other players is our ability to deploy our technology where our customers need it. It can be hosted in the cloud or installed on a local infrastructure. The choice is entirely up to the customer. We are also committed to knowledge sharing, technology transfer and co-construction. This involves fine-tuning models and systems to specific cultural contexts. States often want to retain control of their tools. They also need support in building the services that will be grafted onto them. That’s why we mobilise teams capable of supporting them, teams who know Luxembourg, the language, the laws and the specific features of its ecosystem.

The prime minister said that the discussions went very quickly.

In Luxembourg, yes, things went very quickly because the government is very close. These are things that have also been done in Singapore and other countries.

Last week at Vivatech in Paris, you announced the launch of “Mistral Compute.” How does this fit in with what’s coming in Luxembourg?

I think it fits right in. Mistral Compute is an effort specifically made in AI. Unlike sovereign clouds, which focus more on storage, we are working more on computing capacity, which requires both specific skills and software stacks, and special operations because these machines consume a lot of energy. But, as soon as there’s an American player in the loop, it’s blocked--and we saw a real opportunity to build an integrated European player, i.e. one that manages the compute and infrastructure management aspects and then all the software above that. That’s what we’ve been building over the last two years. This enables us to deploy on sovereign infrastructures and to offer a service that is completely disconnected from American players. We would be delighted to deploy this in Luxembourg too.

But how do you deploy, in that case, with a local sovereign cloud provider?

When we work with local sovereign cloud providers, we generally just bring the software solution. A certain number of developer tools that allow you to build things quickly. Or we act in a completely vertical way and offer a sovereign cloud.

For us, the two things are quite complementary. We’re delighted to equip them and enhance their offering because it benefits us too. But at the same time, we feel that there is also a lack of skills and expertise in the deployment of high-density GPU clusters. That’s the expertise we have. It allows us to integrate and provide a new offering that doesn’t exist on the market.

Behind the “sovereign” side that you talked about at Vivatech, there’s still Nvidia.

The AI value chain is quite long. There are parts of it that Europe doesn’t have.

What we have enabled to emerge in Europe is, on the one hand, GPU management, which is a new activity; and on the other hand the aspect of creating models, creating software and then creating applications. We started at the end of the chain. That’s where we’re closest to the user. It’s the most interesting and the most critical. But we don’t have everything. Upstream, we don’t know how to make chips. There are suppliers who manufacture servers in Europe but who use chips that Nvidia prints. Nvidia works with DSMC in Taiwan… which works with ASML in the Netherlands. As in all value chains, there is a strong global link. For me, what’s important in sovereignty, where we’re focusing, is that, when it comes to deploying services, the entities deploying them should be able to say: “You can’t cut off our access.” Our customers don’t lose control. If we want to take over maintenance, recover the source code and so on, we can. In a way, we don’t bring our existential risk to our customers.

What are you missing in your spectacular development? What’s holding you back in Europe?

We’ve already largely changed gear. I would say that what helps us, fundamentally, is demand. European funds, investment, capital that crosses borders--that’s working. But the real issue is that European companies need public contracts. Our American partners are much more interested in public demand… European companies only grow through the revenues they are able to generate. The European market is more fragmented, it takes longer, the cycles are longer, there is less conviction and there is less technological expertise in procurement. There are European companies producing excellent technologies, but they need public procurement to scale up. What we are doing today with Luxembourg is an excellent example of what needs to be done on a European scale.

A word, perhaps, about your convictions?

We thought AI was too important a technology for us to abandon. And it’s a technology that would be dangerous if it were only controlled by three players.

For us, the risks that are often put forward by our American partners--of saying we might lose control, we might make agents that will destroy humanity--are very remote, are largely science fiction. The real issue is control. Controlling a huge amount of user interaction data that can be used to create powerful engines of influence. For us, the solution is to promote the most powerful artificial intelligence possible in a decentralised way. Bottlenecks must be avoided. It’s better to disseminate the technology as widely as possible, work with businesses, look at where value can be created, transfer technology and work together. In other words, we deploy teams. We have teams whose speciality is to go and work in companies to understand the processes, what can be automated.

We are convinced that for the world this will allow much more balance, much more cultural representation and also more safety when it comes to defence issues.

How do we deal with the black box issue? We never really know what tech companies do with their algorithms, what they favour, what they penalise, how they process data…

There are several types of issue. Regarding security, i.e. when we build systems that are connected to data, we have to ensure that it’s in the right place and in particular that it’s under the proper control of hospitals. The Americans must not be able to seize it under their extraterritorial legislation. Then, when we build systems that will recommend decisions, there is an evaluation aspect, i.e. we don’t want these systems to have a certain number of biases. Systems are always biased, but we want to control these biases. There are a certain number of parameters that should not be taken into account when suggesting certain decisions. As we are a technology provider, we do not claim to manage these aspects of bias. On the other hand, we do provide the tools to estimate them. In other words, we can assess how the models are deployed, what kind of biases they may have, and then, if they do have biases, they--as we have an open technology--can be modified. Our customers can modify the technology and correct the biases they don’t want to have. So that’s where I think this decentralisation aspect is important. There’s no single ethic--there are as many ethics as there are applications and cultures too perhaps.

There’s a lot of talk about you as the future European champion. How far have you got with your developments?

I’m not going to talk about our results because we’re a private company, but every quarter we multiply by three. Growth is very strong. It’s accelerating in Europe, and also in Singapore and southeast Asia. It’s also accelerating in the United States, where our sovereign and portable offerings have a certain resonance. As of now, we have 250 employees, two-thirds of whom are in France. We have a large hub in London, which is a historic hub, another in Palo Alto and one in Singapore. Today we’re also announcing the creation of a team in Luxembourg with five people to get us started. This is our founding team, with technical expertise and delivery expertise, and who know the public and private ecosystems well. We want to start delivering very quickly for the country.

This article in French.