Can Europe Build an AI Champion? — Mistral CEO Arthur Mensch

Thank you for having me, Arin.

So the new site that we have is a high availability site. So we're going to be using it to serve our customers. Our customers are in need of tokens. It's actually in short supply at the moment and so we've been at work building more capacity for them in 2026 and there will be much more in 2027. So why do we do that? Because inherently the business we're in is about transforming electrons into tokens. In order to do that, you need to train the models. We're building everything on top of open source models which makes it easier and I think more healthy for our customers. But then once you have the models, you need to put them on GPUs. And we've built that expertise on serving efficiently GPUs and turning them into token generators through the acquisition of COAB that we talked about together a couple of months ago and through those investments that we are doing on the infrastructure. Europe is lagging behind when it comes to the buildout of infrastructure and so we are investing to close that gap. So that's a new site. It's based in France so it's low carbon tokens I should say. And it's going to be there to serve our customers and to both with our studio offering and public cloud services and private cloud services.

Well, you know, we have a few customers I should say that look at their dependency to tokens as a critical supply chain problem. And one of the things that we have in Europe is that we're a European company, I should say. And as a European company, when we build European capacity, we can actually serve APIs that are fully under our control. And so that decoupling from other providers is actually quite important for our defense customers, for our manufacturing customers. So that's the reason why we really decided to invest is that from a product perspective, owning the full stack in Europe is extremely attractive. And what's more is that because we built that software stack that can fit on any hardware we are also bringing it to Asia. We've announced with Singtel the partnership where we're actually bringing our public cloud platform on top of the hardware they're building so that they become a full stack provider. So we're all about building full stack AI providers that are resilient that are not dependent on foreign entities when it comes to the digital services and so that investments when we own the compute is important for us because that allows to have a big product offering. Now we also run compute, we work with hyperscalers in other parts of the world. But in Europe our customers are asking us to actually own the full stack.

So there's really the two categories. So on one side when we work with enterprises among our customers they don't really care about the GPUs themselves what they care is the services on top so the high value services that allows to deploy the agents that allows to connect of course generate the tokens but then use the tokens to call tools, record what the systems are producing in term of data. We use that data to train new models that are going to be more efficient. And so enterprise customers they don't want to see the infrastructure. They want to have serverless offerings and they just want a comprehensive suite of tools that when stitched together allows to build business applications. So that's what they need and that's how we address their needs. Now you also have AI labs and companies that are maybe quite advanced in using their own data flywheel using their own data to build their own models and those may actually want to actually get the compute itself. And so we do offer GPU as a service through managed Kubernetes to AI labs and to some of our customers that have part of their R&D teams that are actually doing AI research. So the two things are available but when it comes to enterprises high value services and token generation is the thing that we provide.

Absolutely. I mean the AI labs are in sore need of compute and we have some of it and some of them are actually asking us for a lot of compute today. We actually need to prioritize the access and so we're giving it to some of the AI labs but more importantly we're prioritizing our customers that see a surge in usage as they are moving toward the systems that are running on the background and agents are producing much more tokens. So we're at work building as much capacity as we can to address the large amount of customers that want compute for different kind of use.

I mean every lab wants to have inference in Europe because you have latency problems etc. So that's one use case. But then when it comes to training the question is it's not really a geographical question. It's an availability question. And as we have availability and as we've have a lot of proof points on training on this kind of hardware because we train on this kind of hardware ourselves. So we've made a lot of improvements on the training software that we expose to our customers today.

Well, I think the two things may be slightly true in that European companies are adopting and we help them adopt the technology and increase the ROI so that they can justify more spending. And then on the other side there's effectively pretty high aggressivity when it comes to deployment everywhere. What Europe has is a very good grid. So availability of energy is high and it's actually fairly easy to build data centers that are in the hundreds of megawatts. So that's an asset that we have. Now what our customers are telling us is that AI is becoming so important that they actually need to think about where they actually source the technology itself. So the same way in energy you actually import energy sources but you also produce your own energy. AI is really looking like energy at this point in time. You do need to have affordability of energy. So building on an open source foundation is the way also to do more customization to create your own models so that they can run on smaller hardware. You need security of supply. If your provider is actually under certain legal constraints that may come from foreign entities, you never know what can happen. And so when it comes to your resilience plan if you're in the board of a company globally well you do want to make sure that your tokens may come from different places of the world. And so the fact that we can provide fully independent token generation to all of these companies is actually useful not only in Europe but in the entire world.

I mean it's hard to tell. What we see is there's a very big increase in demand. And today there's actually not enough chips, there's not enough memory. There is enough electricity in Europe, there's not enough electricity in the US. So today the demand is actually way above the supply. Now we all anticipate the rise in demand in slightly different ways. We're at work with our customers to make sure that when they spend a euro in tokens, they actually get like two euros in return. Because if that's not the case, the entire thing is going to collapse. They will have too much compute because at some point everybody is going to ask I'm spending 10% of my opex in AI. Is it bringing me more than 10% in growth? So at the end of the day what really matters when it comes to estimating whether we're overspending or under spending is whether the enterprises are effectively adopting the technology to build real world use cases. What is working today and what is driving the demand is coding. Where if you're a developer it brings you a lot of productivity but at the end of the day if you want to have an impact on the real economy well you do need to bring the technology to the real world objects, to manufacturing etc and so that is also something that we're really investing in making sure that AI systems are affecting engineers but not only the software engineers the industrial engineers as well.

It's not only a need for Europe. I think it's a need for every state that wants strategic autonomy. I think the main reason is economical in that this is a technology that is going to be a line in the budget that is maybe 10% of the wages. So if you look at the wages in the world it's around 50 trillion. So we're talking about something that is worth 5 trillion tokens I would say in the next five years. It all depends on how fast it goes. But I believe that with the right enablement and with the right models, we can actually get there. Now, if you take Europe, Europe is around 9 trillion in wages. So we're talking about roughly a little north of 1 trillion in spending in AI in the next 5 years. The amount of money that goes back to the US because of digital services today in Europe is 250 billion. That's a lot because all of this money is actually reinvested in R&D in the US and not in Europe. So in a way you have some compounding effect of depending too much on technology from one region to another and that compounding effect is going to increase if there is no alternative. That's the reason why you see states like India also start to think about their full stack strategy. That's the reason why we work with the Singaporean states to develop a full stack AI strategy in which if we are to disappear well they can still produce the technology. And that's the reason why Europe is starting to look at AI as a strategic asset the same way it has looked at gas and that's I would say there is a realization even in the policy maker side that something needs to be done but really the companies I would say are the ones that are making it happen. What we see with all of our customers in the US, in Europe, in Asia is that the kind of proposition that we bring which is centered around open source models that can be customized is resonating with them and that brings demand and we believe that that window which is fairly short actually because there's only a limited amount of chips, a limited amount of memory and limited amount of electricity. We believe that the demand we see allows us to take a very meaningful position in everything that is related to mission critical AI deployment. So that I think is the hope that we have but really what I'm regretting and that's the reason why I was asked to actually go to see lawmakers in France and I wanted to share that this is not only a technological problem it's actually a macroeconomic problem. You can't afford to have a commercial deficit of a trillion if you actually want to stay competitive and in the race and so that's something I think that people are realizing that we're talking about something that should be concerning for any one of us.

It will not change short term. And it's not that much of a problem. I mean we live in a globalized economy. That is the reason why we've been growing so fast in the last 50 years and it's great. The question is how do we maintain the equilibrium that we have in order for every part of the world to actually thrive. Today goods are being exchanged between the US, Europe, China, between South Korea, Taiwan. When it comes to building the system that I'm used to deploy AI, I would say on the good side, this is fairly balanced in that the semiconductor chain is completely intertwined. You have ASML which is a critical piece of it which is a European company. You have TSMC, you have Samsung, you have SK Hynix, you have a bunch of fabs in the US as well and then you have Nvidia of course and more and more I would say chip designers that are trying to disrupt the space. So on the good side it's of course Europe could actually build more but for this it needs a market and for it to have a market it actually needs to have cloud providers. So we go where we think we have an edge which is the digital services, the deployment of AI serverless systems that allows to build AI applications, the deployment of a high value skilled workforce that allows to turn those serverless services into AI applications that deliver value and we build value for that. Then we reinvest in R&D. We actually buy chips. And eventually we think that the tech ecosystem in Europe in particular can grow to a point where it becomes a good idea for a fab to set up for a company like Samsung for instance or a company like SK Hynix or a company like TSMC to set up a fab in Europe. But for this to happen, you actually need for those companies to have a market and the market is the infrastructure that is getting built. So let's start where we are strong. That's what we shared. And then let's create something that allows every country of the world to get enough leverage and to participate into the AI revolution in a way that is not creating unfair dependencies.

So we don't do it yet. This is of course interesting. No, we're not ruling it out. Because if you look at the chip design space, there are really it's not low hanging fruits, but there are fruit. So you can really lower the cost of deploying tokens to a meaningful extent. And so we're working with a few chip designers that are really building custom ASICs. They do much better job than we do. They take our models and they try to make it work. We deploy our systems or serverless infrastructure so that we can actually create tokens with a multitude of chips. We think it's going to matter for our customers. It's going to matter for the total cost of ownership. Today we're really focused on making the models and operating the chips. We're focused on turning the models into things that are useful for enterprises. So with the right enterprise context etc. That's already a significant part of the stack. That means we have multiple business units. But really when we look at our markets and when we look at our opportunity of building on top of open source models really owning the infrastructure owning the product is very important. Owning the chips may come I think it should come at some point but for now we are relying on Nvidia which is a great partner to us and we're testing a few things here and there.

Yeah it's this is not something we know how to do you know. We know to spec the things that we want. We know the kind of memory bandwidth we need. We know the kind of network requirements we need. We know the kind of memory chips that we need. But then we think other companies actually do a much better job than we do.

So I would say maybe one day but really optimizing for cost is going to matter in the entire industry.

I mean we found Mistral already which was which I think is pretty cool. But finding others is we're leaving to others to do it.

So it's an agentic platform that is based on our open source models. So I think that's already a pretty key aspect compared to others I would say. What is an agentic platform? It's first of all it's a place where you can well use the models with chat interaction the same way you've been doing it for the last three years. But really where you start to get value out of AI systems like the one we're doing is when you start delegating some tasks to it. So if you're a software engineer it means downloading one of your connecting to one of your pull requests and making it better. It means taking a PRD and turning it into a

It means orchestrating a multitude of agents that are actually doing your jobs on your behalf in a way where the models are well connected to the entire context of software engineering. So that means the product interface where the product gets designed, that means the documentation of the company, that means the customer requirements, etc. So we are betting on the fact that really the job of delegating tasks as a developer is going to be a very similar job of delegating tasks as someone who is actually not technical. So that's the reason why we brought together two products. One of them which is our coding agent platform with a command line interface which is the thing that developers like to use, and then we took Le Chat, which we named in a French way, but if you pronounce it in English you get various results. And so we realized that we could actually turn it into something quite unified.

So it's an agentic platform. What is there in an agentic platform? You have the models, you have the business context that is constantly updated, you connect them to your various systems of records, and then you launch agents that are building intermediary representations of what's actually happening in your company. And then you have an execution layer so you can say to an agent to actually run tasks all the time with triggers on a regular basis, etc. So that execution layer matters. What is kind of specific to what we do? I would say the first thing is it's built on open source models which means that you benefit from the cost optimization that we brought. The second thing is that you get way more control than with other providers. All of the state, all of the data, all of the customization to the user and to the organization can actually be hosted on our customer tenant. We can connect what's deployed on our customer tenant to our token generator and we can connect them to deployment on their GPUs. It's sometimes hard for our customers to deploy on their GPUs. So we've combined a stateful hosted component on our customer tenant and a stateless GPU hosted by Mistral to have the best of both worlds: control over your data and then efficiency and cost efficiency and latency improvement. The third thing is really around the customization.

You get some benefit by deploying personal agents that are doing things on your behalf. You get much more benefit if you're building a procurement system with the right interface and with the right orchestration that is pinging multiple people at the same time. And so we've built Le Chat in a way where you can host your custom application. So your procurement agent, your customer service agent, all hosted in the same place with the same observability, with the right governance. So connected to the same data and you can control what kind of employees get access to what kind of data. And we've built it so that our forward-deployed engineers can actually quickly build applications that are business applications and that are easily accessible by the customers we work with.

So autonomy is the one thing that matters because you get more and more leverage when you actually can delegate a long task to an agent. When you're talking about autonomy, one critical aspect in terms of infrastructure is how do you connect the GPUs that are producing the tokens to the execution layer. And you want to do it in a way where you have supervisors on top that are making sure that your agent is only doing the things that it is allowed to do. So for this we're using the sandboxes that the core team has built for us and we're deploying them. It will soon be available in Studio as well. These sandboxes make sure that you basically have a serverless interface to spin up and down a lot of agents at the same time. So autonomy requires the right infrastructure and the infrastructure is both the GPUs, the CPUs, and you want serverless deployment there, and also the state management. When you're deploying an agent, it's going to create some state, it's going to learn things, it's going to write memories, it's going to use file systems to organize the things it is learning whenever it has a query. And so the management of state and personalization, the persistence in between two sessions of delegation matters. And so Le Chat is actually doing it for you.

Now the second thing that matters when you're deploying end-to-end process automation in the company is that in general it's not only about autonomy, it's also about validation by humans. And so you need a combination of dynamic deployment of agents that are dealing with dynamic inputs, but you also need deterministic gates. You need a human validator in certain places. You need your procurement and your build to be validated by humans so that someone actually takes the responsibility of validating an agent process. And so what that means is that you need durable process execution that can interrupt themselves and go and ask for human permission, go and send a message to a human so that they can pursue the process with the right indication. And you need to do that in a way where the system, the agent, can actually interact with multiple humans. So that's why we are using what we call workflows, which is the combination of deterministic behaviors and dynamic behaviors in a way that makes the CIO happy in that they know what's happening and they know that the process is being followed, but it also makes the business users happy because it works. And we are all making that through in a way where when you're deploying an agent in Le Chat, the ground source of how it's working is code. So you can actually take that code and modify it. And a developer can actually understand what's going on, which means you can maintain it. Because if you build agent platforms in a way where agents are only defined in the proprietary language of a provider, you're not going to be able to maintain the thing over time. So we've made that Le Chat is allowing to go from a deployment for a non-tech user to deployment for a tech user so that you can maintain the thing over time.

I think two things. First of all, it's disrupting completely the SaaS business. In that overall, once you have the execution layer, once agents can actually operate a lot of things for you, the one thing you should care about as a buyer in an enterprise is to make sure that all of the SaaS providers that actually own some of your knowledge are making that knowledge accessible to the agents. So there is effectively a growing realization that the connection of models to the systems of records of a company is the one thing that matters. So once this is done, you can actually start doing a lot of automation. Now a big problem in enterprises when they are deploying systems like this is that pretty quickly they cease to be bottlenecked by the intelligence of models and they become bottlenecked by their own organization. So back to the point I was making, you need to orchestrate agents with humans in a way that is observable and in ways where the systems evolve over time and get auto-corrected over time. It means that you need to readapt the organization you've had centered around agents that are orchestrating the processes. So the way to do it is basically to look at your core processes and to think about how you would make them faster, how you can automate them more, and where should be the human gates where you maintain the quality and you maintain the innovation. So that means thinking about the organizations in a way that is quite top-down in that you want to take every function, every core business, and think about how to re-orchestrate all of the people that are involved in that process around an AI system where the agent is the orchestrator and the coordinator. So that's I think a very important aspect. And maybe the second or third is that it's changing profoundly the way information is actually being shared in companies.

And so you no longer need to ask your colleagues about what's happening, provided you have the right context. We call it the context engine that connects the models to the different systems of records and to the documentation in companies. And so that allows to gain a lot of time because you have much faster turnaround. What it requires though is something that is not technology. You need to ask as a business leader, you need to ask your reports and their reports to actually document what they know. Because if you don't document what the process should look like, if you don't say what is in your head and what is easy to transmit to your co-workers that are nearby, your agents are going to be a little lost. And that's actually extremely important. It means that as humans, as employees, we do need to give as much context as possible to the agents that will become orchestrators of core processes.

So we do, because our customers are more and more asking us for multiple things on the cybersecurity side. Cybersecurity and AI has multiple pillars. So the first is that you can use AI to detect vulnerabilities. You can use AI to do faster updates of your code bases and you can use AI to guard against penetration attacks. It's not only about the models. Models matter of course. And we've seen huge improvements in every category. And I should say our models are a few months behind but they're really catching up very quickly. Those models are able to detect vulnerabilities, to propose exploits. And those models are also able to be used for heuristics to guard against cyber attacks that are penetration attacks against networks. So you need the models but you also need the harnesses that use the right tools, the right network attack tool. You need the red team harness to test your systems with pentesting. You need the blue team harness to defend your systems. And so those things are products that we're working on with customers that are asking us to provide them urgently with a solution. And we are building that on open source models. And it works very well. And we think that as anything cybersecurity, at the end of the day you want the systems to be open source. It's the case for encryption, it will be the case for open source. It will be the case for AI as well. So we're betting on open source technology. We're betting on making our models available to everyone so that everyone can understand the capabilities they may have, even for attackers. And we think that's going to lead us to a safer system. So that's the first thing, that's the first pillar.

Now cybersecurity is also about the way you're deploying agents because it's very easy to vibe code when you have an IDE. Models are great at dealing with a lot of different heterogeneous data sources. So if you use even open source tools as just a single user, you can create pretty powerful deployments and agents that are doing things on your behalf, your personal assistants, etc. That's easy if you're a single person and if you're comfortable giving your data to a closed source provider. It becomes much harder if you are an enterprise because you have inherently a tension in between making your employees and your managers able to build applications that are very adapted to the business processes you want to automate, but you want to make sure that they are not doing things that are weakening your security posture. So you want isolation, you want monitoring, you want guardrails, you want to make sure that models are accessing only the data that you're giving them access to. And even if you're respecting the access control that the IT is providing, you want also to make sure that you don't have need-to-know problems. When you're connecting your models to all of your systems of records and documentation, you're bound to find places where certain employees actually have access to things they shouldn't know about. And it was already the case, but because you're reducing friction and because agents can just look for everything, you actually need to worry about those things and you need to have dynamic access control systems in place. So you have a variety of primitives that need to be set for an enterprise to be comfortable doing strong delegation to AI systems. And so that requires a combination of systems and model capabilities. And so we are at work building them with our customers because often times you need high levels of customization. You need the models to deeply understand the legacy systems, the overall just the pure IT architecture that is often times quite messy in enterprises. And so customization, deployment of our security engineers is actually something that enables our customers to go faster. We'll have a few more announcements in the upcoming months about that thing.

Well, we like to think that we're resilient to this kind of euphoria. I mean, there are reasons to be very happy in that the models are already getting stronger. We see in certain domains in enterprises that it's starting to pick up. There's still a lot of work. There's still a lot of viscosity in adoption in enterprises which means that there's still let's say a lot of value creation to be had. Software engineers, they can use AI systems. Industrial engineers, they cannot use AI systems. And so if you actually want to go and tap the 30 trillion market of manufacturing with artificial intelligence, generative AI, you actually need to solve a lot of new problems that we haven't solved yet. You need the models to understand tools that are very complex. You need the models to understand physics. That's a recent announcement that we've made, investing in acquiring a company that is building models that understand physics. So you know when we talk about AGI, superintelligence, it's all going to be simple. I think it's too simple an idea. You need strong intelligence in language but you also need a very strong understanding of the physical space. We are not there yet. So viscosity is high. Enterprises will make strides in the coming years but they need to change their organization. There are multitude of domains where AI is not having an impact yet for lack of capabilities. So we need to build those capabilities. The recipe is there. You need more data. You need more compute. You need expertise. And you can do it but there's a lot of work to be done.

Now on the question of whether there might be things on the market etc and where to invest etc, we think that as long as we're focused on the creation of value for our customers and as long as we take the right bets in terms of providing them with the compute they need, in terms of making them happy and making them understand that this is not just a technology, it's actually a true industrial revolution, we think we're good. We think we're doing the right bets. We think we're resilient to anything that may happen in the market. And of course we tell our customers that they should care about the supply chain. Because it's effectively getting more expensive because there's a frenzy of buying. And so that also means we need to accelerate on our compute facilities and that we need to make sure they get good cost and that they get good prices, that we get good cost so that they get good prices. So still a lot of work to do. Again it needs to be full stack. It needs to be on the model side, on the infrastructure side, on the product side. Still a lot of work, still a lot of work for our customers as well. But overall the future looks bright because you can actually go to, we are going to enter a society that can grow faster. That can solve problems that we have been unable to solve like global warming. That can solve a lot of health issues that we're running into because of longevity. That can actually build better planes. We've announced the partnership with Airbus today. That can actually better lithography machines, that can build better cars, safer cars. We've announced the BMW partnerships today as well. And the future is bright, but we all need to work together and we need to make sure that everything is built in a way that is fair. And to build fairness, building on open source is probably the right bet.

Does it mean something? I think of course IPOs are always looked at. We as a company are a private company and can stay a private company for longer. But of course that's something that we'll look at with interest for sure.

You know I think AGI was a poor point concept. And so the reason why we've heard about AGI for like 15 years in between 2010 and 2025 is because the thing was not really working yet. But now it's actually working and we can turn it into something that has value for customers. And so suddenly it becomes, it's no longer a pure term, but suddenly the thing becomes way more complex because if you're talking about something that does not exist yet, you're talking about something abstract, you would rather have like a single term instead of saying it's all going to be very complex. It's going to have to understand many different things etc. Now we're at the stage where actually you just need to take the models and to connect them to business data and to understand what the people want to do with it to actually build value. But it means you run into a lot of things that were not anticipated in the PowerPoints of 2010. So you need them to understand physics. You need them to understand human behavior. You need them to be connected to all sorts of the 50 years legacy of software that we've been building. So it's all very complex. It's all very messy. It's all linked to organization and people which is even messier than what you can build with technology. And so at the end of the day I don't think intelligence matters. I think what matters is empowerment. It's a technology that can do many things but there's a lot of plumbing to be done to make it work. So suddenly you move from a very abstract concept and what I think is a little bit messianic as well into something that is very concrete and where you need the combination of people that do not understand technology but understand their business with people that understand the technology but do not understand the business of their customers. So they need to share our knowledge and that's the way we can build AI that is actually useful. That's the way we can accelerate technological progress. So if you take AGI as the definition of accelerating technological progress which is rather a direction than a point of arrival, then really we are building AGI but it's really a direction and it's a very messy direction. There's a lot of things to be built and the frontier is enormous.

Thank you so much for joining me.