🔴 Arthur Mensch (MistralAI) devant la commission d'enquête sur les vulnérabilités numériques

I swear it.

Very well, Madam President. Mr. Rapporteur who isn't here, so Madam Rapporteur. Ladies and gentlemen, deputies, I want to thank you for your invitation to discuss this key subject, which is Europe's autonomy in artificial intelligence and, more generally, in digital services. I think we shouldn't make a distinction between today's digital services and especially tomorrow's, which are essentially artificial intelligence. You have storage, you have a bit of execution of agents on it, etc. But fundamentally, digital services, the cloud, it's artificial intelligence. There is no distinction to be made on this subject. And so that might answer your question. I noted your two questions, I will answer them later.

Perhaps to give a bit of context. As you know, artificial intelligence is redefining economic and geopolitical balances. The question is not so much whether Europe can compete, but rather how it can do so. Because I think that if it doesn't compete, in reality, it completely gives up any form of presence and voice in the concert of nations. So the question is, can it do so without depending on extra-European actors? The answer is obviously no. The second question is, can it do so by gaining leverage over its suppliers? By gaining leverage as an exporter of technology and by rebalancing the economic imbalances that already exist and will increase? I think the answer is yes.

A few elements on Mistral. It's a company we founded a little over 3 years ago, on April 28, 2023, with the intuition in 2023 that generative AI would significantly reshuffle the cards in the field of digital services, and profoundly change the way the economy works because fundamentally it shifts intelligence from labor to capital and machines. We saw in 2023, because my co-founders Guillaume and Timothée and I had a certain experience with the big American players, that the world was heading towards a clear oligopoly, and we obviously wanted to counter this oligopoly. And so we created Mistral in 2023 with about fifteen people in France and England at the start, essentially researchers.

What we did was first demonstrate to the world that we knew how to train language models, which are what's used for generative AI. And then we demonstrated by making an open model, meaning we gave them to the world to show that we could also take them, modify them, deploy them where we wanted, etc. So it was a technique of decentralization and displacement of established actors. And then from there, we continued to invest in models.

Today, we have invested a billion this year in R&D. So we invest very significantly, we build models, we are one of the leaders in open models that can be modified, deployed where we want. And we have a whole economic model that is both above, meaning we will provide a platform for deploying agents, delegating tasks to AI. We will create business applications in contact with our clients, who are generally large companies. And then we have an economic model that is below, meaning we also host the models, we put them on machines and we make them generate what are called tokens, which is the economic unit of artificial intelligence.

Meaning we train models that generate sequences. From these sequence generations, we take actions, we execute things, we think. These sequences are made of what we call tokens, which are numbers, numbers between 0 and 65,000. And so we generate these tokens for our clients, the hosts, we transform electrons into tokens. There, today, a few figures: we are 1,000 employees, a valuation of 12 billion euros, and 1 billion euros in revenue targeted by the end of the year.

Our clients are enterprise clients, public sector clients. To name a few in France, we have the DINUM, the Caisse des Dépôts, France Travail. In the private sector, it's CMA CGM, Stellantis, TotalEnergies, BNP, Paris Bar, etc. We do about 30% of our revenue in France, 70% abroad, and about three-quarters of our revenue in Europe. To give you an idea of where we are, we are fundamentally a technology exporter to both the United States and Asia.

What you need to understand about generative AI and fundamentally what we call AI today is that it's a model that consists of transforming energy into intelligence, into the generation of tokens which are then used to delegate tasks, to do mathematical research, to operate logistics chains, to improve tools, to deploy systems embedded in robots, in drones, etc. So we are in a business where we transform electricity into token generation. And it's fundamentally something that should be treated more or less as a natural resource. Meaning we should think about intelligence as we think about energy.

So our goal is to provide intelligence that is cheap, affordable, let's say in English, that is secure in terms of supply, meaning we don't need to go get it from Americans for example, and which is also sustainable because it transforms energy that is less carbon-intensive than elsewhere because we use in particular French energy. I think we don't think about it enough in this way. Intelligence is like electricity, it's like energy. And besides, we have this tendency, which in my opinion is a rather false tendency, to separate digital services from artificial intelligence. Today, the cloud is artificial intelligence. The growth of the cloud is artificial intelligence.

High value-added services, with high margins that allow for R&D, that's artificial intelligence. So in fact, everything else, from the moment we gain in AI, we can build everything else which are essentially cloud services that are commodities today and that essentially consist of knowing how to operate open-source software because it has become the global infrastructure. Our goal is obviously to anchor this value chain. So this, I would say, our chain, where we are in the value chain between the operation of clusters and servers up to the creation of business applications. And I think it's an atomic unit of the value chain that goes from the construction of semiconductors, so ASML for example, this company that does lithography, to deployment in companies.

So that means that the cloud and AI are really the same thing. What I can say is that we must see sovereignty, we must stop thinking of sovereignty as isolationism. We must think about it as business people think about it. Meaning we must think of sovereignty as a subject of leverage. In a world where you import all your digital services from the United States, you have no leverage over the United States. In a world where you create part of your service, and possibly export it, you have additional leverage. In a world where you sell goods to the United States, you have leverage over them.

We see it in European Union negotiations, we export a certain number of things. That gives us arguments and cards to play when we are at the negotiating table. If we don't have this capacity to not import all our digital services from the United States, we have a real problem. The problem that arises, in fact, is that the total market we are talking about must be compared to the payroll. Today, at Mistral, our AI consumption for our employees is 10% of our payroll. Now, we make this technology, so we adopt it a bit faster than elsewhere.

And if you extrapolate, in 3 to 4 years, if all of Europe uses AI for 10% of its payroll. 10% of Europe's payroll is about a trillion. A trillion, that's digital services because you have to generate the tokens, etc. It's a trillion if we use non-European technology, it's a trillion in trade deficit to add to our existing trade deficit on digital services. It's a trillion that goes elsewhere to be invested in R&D and is not reinvested in R&D in France, or in Europe. So that, I think, is an important point, is that we must think about this sovereignty issue from a macroeconomic point of view. An essential macroeconomic subject whose importance I think we don't realize.

So that's a first subject. There is obviously a subject of economic security, meaning when you know how to deploy AI, you want to ensure in essential services that access to the service cannot be cut off, which is actually quite theoretical. We've seen it in different places. So that means it's important for all sovereign, governmental aspects. It's obviously important for defense because AI is now at the heart of military operational centers and becomes the platform that allows making operational decisions on the battlefield, doing intelligence, etc. So if we don't have access to this technology, we will have to import it from elsewhere and that means we can be cut off from this access. So I think it's particularly critical.

Finally, there is an additional subject which is not a subject we really deal with directly because we don't have a significant consumer activity, which is that AI that generates content shapes cultural representations, shapes language, and also shapes ethics. Meaning these are models that have a certain policy, they activate certain things, they will write a certain type of code, etc. And so all this becomes a mediation of information and it becomes a mediation of action. And so if we have to import models from elsewhere and if we don't have an alternative solution in Europe, we will be forced to use the biases and choices made by others, the United States and China in particular.

So a few recommendations to finish. AI offers the hope of reshuffling the cards. That's true, it goes back to your first question. One of the strengths of Europe and France in particular is its talent base. It's an asset, it's an entry into the system. Meaning if you don't have innovative companies that can attract these talents and convince them to work for us, they will simply go elsewhere and bring their value elsewhere. That's what we've seen a lot. I think it's something that is starting to reverse a bit. We must benefit from it.

Beyond talent, I mean Europe doesn't have many assets. We have heavier regulation, we have a fragmented market. Obviously, a fragmented market means more difficulty selling, but it also means longer-term partnerships. So that, we can say, is an asset. There is a very strong cultural diversity asset in language, etc., which is a real diversity for AI.

What I can say is that fundamentally we only exist if there is a market. And the problem we see is that this market materializes a bit too slowly compared to what it should be. Meaning companies don't yet fully realize the amount of AI they will use in the coming years. And in fact, the consequence of this is that those who can invest before the demand are rather the Americans. And so, for example, there is an energy surplus in France, about 9 GW on average, this energy surplus will be captured rather by the companies that are the most aggressive from a capitalist point of view and who will be able to put money on the table, transform these electrons into tokens before others.

And so for us, our subject, we manage to be quite aggressive but we would like to be more aggressive and for that, we need visibility on the market. This market is obviously the private market, companies that choose to work with us for 5 years, for 7 years, and we have plenty. But it's also public demand which forms nearly 50% of Europe's GDP. And so being able to ensure that this public demand flows across the entire value chain starting from where we are strong, where we have added value, which is AI, and doing it in a concentrated way, meaning not deploying with subsidies, simply making demand at the end of the chain on digital services. That's an essential asset.

It's an asset that our American and Chinese partners use massively. It's an asset we have always been afraid to use in Europe. That must absolutely stop. And finally, in general, we seek to structure markets, ensure we have relatively unified demand. So unified markets are useful. Unified capital markets are useful. We never do enough. And then finally, from a purely infrastructural point of view, France indeed has this chance of having surplus electricity. We must continue to do so because AI, we can consider it's about 1 kW per person by 2030. 1 kW per person by 2030 means 40 GW to build in France, about 400 GW to build in Europe. That means about 20 trillion in investments to make.

So we really need to be aware of this and build electricity but above all ensure that electrons are transformed into tokens with European companies. And to build 400 GW in Europe, we need to accelerate projects. That means faster permits. We also need a notion of reserving energy capacities. There. And then on your second question, I think I've addressed it a bit. Beyond AI, digital services. Yes, you can do managed databases, you can do hosting, you can do virtual machines, etc. Those are services with extremely low margins. You can't build a business at scale starting from a place where the margin is low.

If you want to build a business at scale, you start from where the margin is important and then once you are big enough, you use part of your R&D, but you don't even need to be big, to go create all the services you use. If you start from where the margin is low and you don't have volume, if you have neither volume nor margin, you have no chance of reaching scale. So either you are in a place with low volume, high margin, or you have low margin but high volume. That was the Chinese strategy and you move up the value chain. But if you are in a place where you have high margin and not much volume at the beginning, you can only move down the value chain. So I think that's really important to take into account. You must start from high-level services, go back down to low-level services because those have less margin and that's how we approach the problem from a strategic point of view.

It's the economic unit of artificial intelligence. It's something that is billed per million tokens. A million tokens on one of our models, for example, will be 1 euro. So you send me a million tokens, I generate a million tokens for you. You will pay 1 euro for the input, 3 euros for the output. So that's it.

It's a few letters in a word. It's a number, essentially. It can also be a few pixels in an image when it's input. So we manage all multimodal data. It can be a bit of sound as well. So that's the unit. Our economic model is partially a hosted model, so consumption-based. Meaning you use technology, you host technology with us and you pay per token, you pay per unit of time in CPU. So when you do something other than AI, you pay for storage. We also have a model that is more of a platform model where we bring our software, deploy it on client infrastructure which then pays for its infrastructure, its GPUs, its CPUs, its storage, and on that it's more of a license model on which we have a service layer to build value and transform companies.

It's generally used by developers who will make applications with it.

The token is the client unit of the digital service. So when you buy a digital service, you buy tokens. The unit for us and the way we finance our projects is rather the megawatt. The megawatt is the electricity we will transfer into intelligence. 1 MW... so I'll give it in gigawatts because I know the figures well in gigawatts. When you invest and create a data center, you put in graphics cards that will generate tokens when you host the model. To create 1 GW, you will invest 50 billion dollars over 5 years. So it costs you about 10 billion per year. With that, I'll give you the margins a bit. With 10 billion in data center per year, you can make about 20 billion in revenue in tokens. You can sell 20 billion tokens if you do things correctly. If you are roughly at scale, 1 GW is roughly at scale. So that's roughly it, meaning you have 50% gross margin between the digital service and the investments the digital service provider makes.

So it's a significant margin, it's a margin that allows investing and doing R&D. In particular, it allows financing what costs extremely much in AI, which is the training of models because we use GPUs ourselves, we use today tens of MW to train models. So those are costs only for us that we bear. So that's roughly what we can keep in mind, 1 GW. I mentioned that if you have just 10% of the European payroll that becomes AI, we're talking about roughly 400 MW because we're talking about 10,000 euros per year and 10,000 euros per year rents you roughly 1 kW of GPU. A GPU is 2 kW. So you rent half a GPU with that.

So finally, you are in a situation where to build this, you need to make 20 trillion in investments. These 20 trillion in investments will yield about 8 trillion per year in revenues that can go back... So we are on scales that are quite significant if you want. And what you need to take into account is that there is this 50% gross margin between the service provider and what makes the service provider. In fact, in the chain that transforms the electron into token, there is about 10% that goes to the electron. So if your only role, if Europe's only role is to be the energy provider, that means 90% of the value is elsewhere. And so it leaves Europe and will be invested, recycled in R&D, used as a vector of power by other actors. So that needs to be taken into account.

Very well, I will take your questions in order. So on the ethical subject, we make systems that help humans be more efficient, that will automate a certain number of processes while keeping humans in the loop on important decisions. And so we have a whole team of designers, about ten designers at our company who will think about human-machine interactions, ensuring that when AI summarizes information to help with decision-making, it provides enough elements so that the human remains informed. Now, I'm not telling you it's an easy subject because in fact, humans tend to be a bit lazy and so if the thing does it for them, they think less and less. So there is this subject of managing to keep them in the loop, you have to stimulate the human when they are facing the machine. And so those are subjects we obviously think about.

Then, you have a sub-part of ethics which is the subject of engagement in defense. So we work with the Ministry of the Armed Forces, we also work with certain allies of France and it happens that this technology is considered a dual-use technology and so fundamentally it is regulated in terms of exports, in terms of export control and so we have the entire framework of export control regulations that applies to us and so that allows us to have... we don't claim to have an ethical framework, we don't claim to have the legitimacy to have a sufficient ethical framework, we don't have democratic legitimacy as a company. On the other hand, we fit into frameworks that have democratic legitimacy and that allows us to work correctly and not to do what we don't have the right to do.

Then, when we work with the Ministry of the Armed Forces in France, we don't claim, unlike other actors, to have a right of oversight over end use. Because we don't claim to have the legitimacy to explain to the French armed forces what they can and cannot do with technology. We can tell them and we can do it by project, we can explain the reliability of things, etc. But then, who is sovereign in the end in the use of technology? It's the army. The army has a democratic legitimacy that we don't have. And so in that capacity, we have a duty to advise, we explain how it works and I can tell you that today it is used in a perfectly reasonable manner and within the limits of what AI can do. We don't... you have to be aware that AI allows for defense, it allows having much more operational armies obviously, it improves operations, etc., it allows for defense but it's also absolutely critical in...

Dissuasion. That is to say, you have armies like the Russian army that use artificial intelligence massively in their drones. If you are not capable of having counters that are themselves activated by AI, then you don't have sufficient deterrence. And so, the deployment of AI in autonomous systems, in particular, I think we absolutely must not skimp on it. It is absolutely indispensable for having conventional deterrence today. This also means we must master AI end to end. That's on the first subject.

Then, on the subject of cyber defense and cyber attacks, one of our American competitors knows very well how to do fear marketing, if you will. But the fact is that models are excellent programmers and are absolutely capable of orchestrating attacks, discovering vulnerabilities, and proposing exploits. This has been starting to rise for 6 months, it's rising linearly, predictably, it's rising everywhere at the same time. You have another American competitor with a model that does exactly the same thing. You have Chinese actors doing it too, and our models are capable of discovering all the vulnerabilities reported by Mythos, for example. So, we work with our clients to help them on these cyber issues. It's obviously a sovereign issue, and so it's another argument for saying we must have control over this technology. You can't have the French army's code bases scanned by Mythos. I think it creates a dependency that is so irremediable that we absolutely must find solutions.

Next, on the subject of time. On the subject of time, I would say we don't have time. I think, notably because it's linked to physical constraints. We're talking about natural resources here. We're talking about using natural resources to transform electricity into tokens. There isn't infinite electricity in the world. The United States are saturating their electricity to transform it into AI. They are deploying a trillion dollars next year. When they deploy a trillion dollars, it means they want two trillion dollars in return. So you can see a huge market is being created, and you can see that whoever possesses the resource, whoever supplies the chips, whoever possesses the electrons, whoever possesses the power plants or has access to energy on a massive scale, is the one who wins.

So today, you have 90 terawatt-hours per year available in France. There are many things we can do with it. Part of it, we will use for AI. You have actors with huge balance sheets who have plenty of money to put on the table before seeing the demand to transform electrons into intelligence. But they are going to do it in the next two years. Once they've done it, you have no more electricity. You have to wait for new power plants, create new ones, etc. So there is a form of monopolization of the European energy resource that is happening, that is accelerating at a crazy speed, and of which we are not aware. And it's truly irremediable.

We already have a demand problem. Because the Americans arrived on the cloud before us, they provide all the European industry, they provide the French state, obviously. The market share of cloud and AI tends to compress for non-American actors. Not only do they grow with the market, but their share also increases. So I think this is something that can be reversed with AI, and what we are starting to see with our clients, with our revenue, with our growth. But the problem is that even if you solve the demand problem, you have a supply problem, because once the supply is monopolized by American actors, suddenly we have no more supply and we can no longer transform electrons into tokens. And this is happening in the next two years.

What needs to be put on the table today in terms of investment is tens, hundreds of billions of dollars. And to do that, you need visibility on the market. And the visibility on the market must be indicated by European actors, by European states. Today, it's not indicated massively enough for us to take the risk of deploying 100 billion dollars. So I think we need to become aware of this urgency, and unfortunately, I don't see very well how we will become aware of it. So we are going to manage to build 1 gigawatt by 2029. That's our objective. It's not enough. We should build more. But for that, we need a form of visibility on the market and on the demand that we don't have today.

And finally, perhaps on the last question, you asked a question about distillation. What is distillation? You take a large model and you use that model to make a smaller one that will be more efficient. So we use it massively because we train large models, which costs us a lot. Then we use our large models to train small models that we then serve to our clients because it costs them less to serve. So what this means is that it's absolutely indispensable. If you want to know what's in your models and control them, you must be capable of building the largest models. So that means you need to know how to train them, which means you need a lot of computing capacity of your own, which is pure R&D that allows you to train them. So what this allows us to catch up... no, it's a technology that allows us to reduce costs essentially internally. So we take large models and use them internally to make small ones, and we need the large models. Which means we need to invest a lot in them.

So I will answer, I can explain our role in Campus, which is very minor. We have a very small participation. We took a participation because in the value chain, we are capable of building clusters, so we buy the servers, we plug them in. It looks like physical racks that will be put in buildings. But we don't do the buildings. We don't do the buildings, we don't do the transformers, we don't do the cooling. We will rather seek partnerships to be able to do it. So in that capacity, we need a data center developer as a supplier. So in that capacity, we took a very, very minority position in this project, and above all, it's a supplier, that is to say, it will become a supplier for us for part of the construction, another part will be used to supply hyperscalers. Because you start from a situation where 80% of European digital services are imported from hyperscalers. And so naturally, the starting point, if you start building infrastructure and you look at infinite demand, you end up with 80% hyperscalers. With the objective that France and Europe must have, it's to ensure that this proportion decreases over time. So that's why Campus IA today will serve us in part and will serve hyperscalers in part. So that's what I can say.

Then on the name, I didn't choose it. On a number of technical aspects, I don't know. In any case, what is certain is that part of the revenue is indeed returned to public research. Then on the subject of capital, you mentioned it, it's real infrastructure. You build a data center, you rent it out, it gives you an equity return, and it's essentially a pretty good business. If you have a lot of funds, deploying them in data centers today, knowing that everyone wants data centers, gives you a pretty significant return on investment. And so there are plenty of actors who want to do it. When you have a site of this size, you need actors with strong backs, who can deploy tens of billions of euros. And on that, naturally, because we forgot to create pension funds in France, for example, we turn to foreign financiers, because there is no capital market, the capital markets in Europe are completely devoid of providers of solutions for this type of investment. So we go look elsewhere, and I think we should benefit from it. That is to say, foreign investments, if they are under control, if the governance is good in the investments that are made, and I think today on this subject, BPI France plays a role and has a seat on the board, etc. These foreign investments are beneficial because they allow us to accelerate. Because obviously, we lose a certain number of economics because when you have foreign capital coming in, they take a return. So we lose the return, but on the other hand, we build something. So we can draw something from it. What we propose is to capture the margin from the operation of servers up to digital services. So that's as much margin as we can reinvest in R&D. So we don't offer, we are not capable today of making a completely integrated model where we operate from the ground up to digital services. Maybe that will be the case one day. But there is a sharing of the value. We are capable of creating a lot of value with this. We will be able to reinvest it in R&D. Obviously, there are foreign investors who will get the return somewhere. But on the other hand, if we didn't have these foreign investors, we couldn't build. So you always have to think of sovereignty as a path. You start from a state of the economy and balances in the value chain and you say this state is not satisfactory. We must change this state, it will take time. And how do we change this state? Well, we must put the resources in the right place and then it will change little by little.

On environmental subjects, obviously all this costs money. There are externalities. If you go up the value chain to the semiconductor, you need a lot of inputs, you need rare land, you need... land, which I think is not the main issue because it's super dense. So if you want 1 gigawatt, it fits on about a hundred hectares. So we don't really have a land issue, I would say from a point of view... well of course if you look at the city level, if you look at the European level, there is no issue. So it's not the main issue in any case. Then, you have an electron issue. That's a real issue. The interest of building in France for us in particular is that 70% of the electrons are nuclear. So that means the carbon footprint of the electron is largely reduced. And so building in France is fundamentally better than building elsewhere because it's tokens, people need them. So if someone builds 1 gigawatt of capacity in France, they won't build 1 gigawatt of capacity in Texas, and ultimately the atmosphere will warm up a little less. So I think we need to think about it from that point of view.

Then, you have other inputs, you have water in particular. There has been a lot of technological progress to avoid waste heat and ensure that cooling is more efficient and that the majority of electrons are used for computation rather than for everything around it. And then you have biodiversity issues, obviously, which seem to me, having observed the sites, everything has been done and implemented to preserve these aspects. So for us, it's important. Obviously, from an ecological point of view, it would be better if there were a forest, but from an economic point of view, it has a role. So it has an ecological effect. We must think of AI within ecology. We transform energy into electrons. It's a sort of natural resource of electrons into tokens. You consider the token as a sort of natural resource. For natural resources, you need a sort of mine. The mine is the data center. And what I think we must do, because if we don't, we have additional issues, and I think we must do it. The more we do it ourselves, the more we have a say in how we think about externalities. That is to say, if we don't do it, we will just import what is offered to us from outside. And I can tell you that outside, they think about biodiversity and global warming quite differently from how we think about it. So if we want to have a say, in any case, we must be capable of internalizing production, and that is absolutely essential. And so there are obviously compromises to be made. In the same way that we accepted to build lithium mines in France, if we have to build a mine, we must accept to build data centers. But what would be a shame is to build data centers where the bulk of the economic value completely escapes France. That's the issue. So rather than taking these externalities, let's ensure they can be recycled into R&D and into benefits for France.

Just today, our new generation of models, we train them in France, and I think we have the largest computing cluster in France. So it's a source of pride for us because from a carbon footprint point of view in particular, it's largely reduced compared to what we could do before when we weren't in France. So that's what I can say. Finally, on the subject of calculating carbon footprints, etc., unlike some of our competitors, we think we need a life cycle approach and not an approach with offsets, because obviously it's a bit too easy when you have very large balance sheets to do offsets. Whereas in fact, you need to think in an integrated way about everything that goes into projects and have transparency per project rather than having global transparency and imbalance via quotas.