Mistral CEO: Why OpenAI & Anthropic Will Come for Your Business | Are Chinese Models Truly Safe?

Arthur Mensch 0:00 Sovereign AI is going to be the most important category. The cost of energy, at the end of the day, is not the one thing that matters in the economy of a project.

Harry Stebbings 0:08 He is the last hope in Europe for us to build a frontier AI lab. Yeah, you guessed it. I have Arthur Mensch, Co-Founder and CEO of Mistral, on the show today. It is now a $21 billion company. It is the company that is keeping Europe relevant in the race for frontier models. This is Arthur like you've never seen him before. He's unleashed, he's funny, and he says some incredible things. This was such a joy to do, and I'm so excited to welcome Arthur to the show.

Arthur Mensch 0:35 It's pretty clear that there is not enough compute. I mean, the inference market behaves very much like commodity market, which is pretty high margin today.

Harry Stebbings 0:42 Ready to go? [upbeat music] Arthur, dude, I am so excited for this. I have interviewed some of our, our favorite counterparts in the US. You are our remaining hope in Europe. No pressure, but thank you so much for joining me today, dude.

Arthur Mensch 1:05 Thank you, Harry, for having me today.

Harry Stebbings 1:07 Now, [laughs] I was doing some research, and I say some quite opinionated things, Arthur, but it seems I'm not the only one, because you, the other day, said that leaving frontier AI to the US and China will end up costing Europe what remains of its place in the world.

Arthur Mensch 1:25 I did say that. Uh, and I truly believe, uh, that this is such an important technology that it needs to be distributed in the world. Uh, we're talking about a technology that is comparable to electricity, uh, a technology that is falling more and more under export control. Uh, and so AI is going to be a big enabler of, uh, economic growth, and if there's no AI being produced in Europe in particular, uh, Europe will struggle to actually produce the growth it needs to produce.

Harry Stebbings 1:55 I'm going direct. To what extent does the success of Mistral depend on US overreach? You know, o- obviously, when we have a Trump administration, we see things like Fable being banned, and it, it causes a lot of questions and the need for sovereignty. In a world where there's not US overreach, does that threaten sovereign providers?

Arthur Mensch 2:13 I'd say this is structural. This is not really linked to, um, any way of, uh, of operating or of showing power. It's more linked to the nature of the technology itself. Uh, if you start as an enterprise or as, as a state to use AI in all of your core processes, uh, it means that you want to make sure that the money you're spending on those processes actually accrues pretty close to you. Uh, if you don't, it means you have a trade balance problem. It means you have enormous flows of cash that are going from one place in the world, can be Europe, to another place in the Eur- in the world, which can be the US. And so there is a macroeconomic problem to not being able to produce artificial intelligence, to do R&D, uh, on, in Europe territory. Uh, that is not linked to a certain way of, uh, saying that, uh, Europe is falling behind or a certain way of, of, uh, certain way of using soft power, let's say. So it's structural. I won't say that, uh, some people are not making it a little bit emotional. Uh, but, uh, when we started Mistral in 2023, we started with this conviction that this was a sovereign technology, in that, uh, as a company, as a state, wherever you are, whether in the US, in Europe, or in Asia, uh, you need to own the technology. You need to be able to customize the systems to run them very close to your data. You need to be able to control the cost. Uh, you need to work in true partnerships with companies, uh, that understand how to build AI systems. And sovereignty precedes, uh, 2025 election. So, uh, it turned out to be true, uh, what we said in 2023. Uh, it turned out to become quite emotional when we start to speak with Greenland. Uh, but, uh, it's mostly an economic problem. It's not really a political problem. It's just too important a technology.

Harry Stebbings 4:01 When I hear you speak, I think of Lyn at Fireworks, who came on the show and said actually very similar things. We need companies to have their own models, to train them very close to their own data. How do you feel when you hear me say that, and i- is that the evolution of the company, allowing companies to own their own models and work with their own data?

Arthur Mensch 4:22 And what-- The category we define and the reason why companies work with us is that we bring them sovereign AI technology. And so in sovereign AI technology, you, you need three things. You need to have, uh, a source of model that is secured. Uh, so you need a company that can train the models itself. So we have that model factory. Uh, Mistral Large 4 is actually leading the open-weight models in agent encoding. It's leading in, uh, uh, in cyber defense and cyber security. So that aspect of being able to constantly produce improvements on models and to tell-- to be able to tell our customers that in two years from now you'll get better models because we are constantly training them and spending billions on it, is super critical to anything sovereign. So it's the reason why Mistral has been created, and it's, it's the, the main, uh, path where we're investing money. Now, you also need to have control of your, your data. Uh, and so in general, you don't want that data to flow back to, uh, the AI providers. And so we've made sure at the technological level that our platform to run the agents, uh, that our platform for customizing models actually sit on, uh, our customer tenant. That's, uh, super important in many cases. When work with manufacturing companies, they don't want their IP to flow back to a sub-processor. Now, that's not enough in general. Because AI is consuming a lot of energy, uh, and because when you move at scale, you actually need two m- uh, you need a lot of GPUs that as an enterprise you have difficulty to procure. Uh, you also need to have, uh, token generation services, and you need to have token factories, basically. Uh, you need, um, a combination of data that sits in your infrastructure and tokens that are produced in managed infrastructure at a lower cost. And so what we do as a company is those three things. We train the models We deploy the platform and the data plane with our, uh, customers, so on their, on their tenant. And then we host what is energy, what is intense in energy, uh, and that's the reason why we've been building our own infrastructure. So we have these three things, and those are the three ingredients that are necessary to serve sovereign AI workloads.

Harry Stebbings 6:27 Can I ask, if we start on one, you said by... the model is secured. Um, I ha- I have the world's best experts on the show, Arthur. Very lucky. But they tell me different things, okay? Some people say, "You know, the Chinese open source models, you download them, you run them locally. Hey, they're totally safe." And I'm like, "Wow, great." And then next week someone says, "Hey, but we have no idea whether there's prompt injections that you don't know. There's backdoors. Absolutely not. You cannot be sure that they're safe." Help me out here, Arthur.

Arthur Mensch 6:58 Well, at Mistral, we, we have a-- I mean, one of our business is to really run the inference for open models. So we run ours, but we also run others. Uh, and today-

Harry Stebbings 7:09 Why, why, why do you run others?

Arthur Mensch 7:11 Because customers want choice, and when you have infrastructure, you want to provide as much choice as possible to saturate the infrastructure. Uh, so that is the reason why that brought us to actually serve more and more of, uh, open models. 'Cause in different domains, you may want to pick a different, uh, model. You may want to pick smaller ones. You may want to pick bigger ones. So the more choice you have, the better usage you can make of the infrastructure, the lower your cost can be. And it's, in a way, it's a community market, so you want the cost to be as low as possible. Uh, so that's the reason why we sell those models. But what you say is true. Uh, models are compression of knowledge. Uh, if you don't control the data supply chain, if you don't know what goes into the models, you cannot guarantee anything on the output. Uh, and so when we serve our models, we can give further guarantees to our customers than when we serve third-party models. Precisely because theoretically, you can make sure that in certain, in s- in a certain environment with a certain context, the model behaves in a certain way. There's no way of excluding that. There's no way of measuring that. Uh, we can monitor that online. That's what we do. Uh, but you can't exclude that it's being used as a, as a vector of attack. I think it is very theoretical because I don't think it's the case, uh, today. Uh, but on the other hand, this is, this is part of the threat model. Uh, and so our customers, they are asking us to be able to provide them with all of the open weight models that are available. They're... We're helping them do it in a secured way where we can help them monitor what's going on. Uh, but in many cases, they are actually very keen to use our models because we're able to tell them that the data supply chain, everything we've trained them on, is something that we have validated before.

Harry Stebbings 8:50 What percent of token traffic, roughly, is your own models versus other open source?

Arthur Mensch 8:55 I would say around eighty percent comes from our own, uh, models. Uh-

Harry Stebbings 8:58 Wow.

Arthur Mensch 8:59 Yeah.

Harry Stebbings 8:59 Okay. That, that's a lot more than I thought. Okay. On the twenty percent, everyone tells me that, "Oh my God, American CIOs and CIOs would never use Chinese open source models," and actually that fear really is there. You have them as customers. You have amazing customers. Are they fearful of using Chinese open source models?

Arthur Mensch 9:18 Uh, it depends how critical, uh, their industry is in general to government. Uh, if you work with defense, for instance, we work with defense customers in Europe, uh, they would never use, uh, Chinese technology. If you work in the finance sector, again, because it's critical infrastructure, they are very reluctant to do it. If you work in retail, uh, if you work in certain domains of manufacturing, again, depends where, uh, the reluctance is different. Again, I think there is a common understanding that open weight models, if they are being deployed on sovereign infrastructure like ours, uh, are... The, the, the threat is that the only valid threat is that they go rogue and have weird behaviors in certain domains because you may have a supply chain attack on it. Uh, there is a common understanding that, uh, this is very theoretical and, uh, it's certainly not the case, uh, today. But indeed, uh, it's part of the conversation. I'd say what is also part of the conversation is not this, is it a threat to, uh, to deploy, uh, third-party models that I don't fully understand. The... What is part of the conversation is, am I going to be able to get updates in six months from now? Uh, because what's going on is models have gotten so strong, uh, that they're able to do cyberattacks, that they're... You can use that for pretty much, uh, advanced purposes, let's say. And so the consequence is that they are more and more being treated as defense systems. It's being treated de facto by the US as a defense system, because at some point in May, they started to ban them from outside of the US. If you look at what China is doing today, they're also looking at ways of restricting exports beyond a certain threshold. The reason for that is that effectively this is a dual-use technology. And so as any dual-use technology, security of supply matters even more. You, you, um... And so the, the companies we work with realize that, uh, they need security of supply, not only on the infrastructure side, but also on the model IP side.

Harry Stebbings 11:17 Can I ask you, with that in mind, do you think your OpenAI's and your Anthropic's withholding their best models is good or bad for society? And is that just an inherent feature that will be prevalent within this industry?

Arthur Mensch 11:29 Is it good or bad for society? I don't know. Uh, what is clear, uh, is that there is immense leverage to, uh, centralize the AI providers in a world where there's no providers that is allowing decentralized deployment, that is allowing sovereign deployment. So what we are doing is extremely important for companies. If, if we end up in a world where all companies need to rely on two or three providers that are centralized AI providers, that give them enormous leverage, they will use the data, uh, that is being produced by the systems to make better systems and to go after the business of their customers. 'Cause at the end, this is, this is the end game. This is pre-pretty clearly what is being priced in today. Um, so the approach we're taking is to make sure that this world, in which you have an oligopoly Of AI providers that are leveraging all data being produced by enterprises to make better systems that can then intermediate those enterprises, is not the world we live in. Um, and we're very aligned in that with all of our customers that, of course, don't want to live in such a world. So the-- I would say the fact that there are internet models that are being used to do, uh, certain research, for instance, is an early sign that, uh, co-- of what can happen in a world where you have massive concentration of power in artificial intelligence. Luckily, uh, it's not that complex to build models. It's about deploying capital. It's about having the right team that is training the right thing. Uh, and so we're-- we've been able to do it. Uh, the open-weight frontier is catching up, uh, with the closed frontier. In many domains, it's actually above. And so that means that I don't think we'll end up in a world where if you wanna do mathematics, you need to work at OpenAI. Uh, that's, uh, uh... Luckily, the-there is enough people in the world that don't want this to occur.

Harry Stebbings 13:11 You've said the word energy a couple of times, and y- one of the core underlying tenets of being successful is having the energy and the infrastructure to power European AI. We have a very different energy system to the US, and we have a very different energy system to China. You know, in China, they'll move three million people to put some form of energy infrastructure, and in the UK we won't move a, a birdhouse. Uh, [chuckles] it's quite concerning, despite the joke. How do you think about, ruminate on the energy supply challenge?

Arthur Mensch 13:43 Well, I think the first thing to realize is that, uh, energy is around 10% of the bill in the total cost of ownership of, uh, compute. So we're not talking about the place where we can make most of the optimization. Uh, so the cost of energy, at the end of the day, is not the one thing that matters in the economic of a project. Uh, what matter is the cost of the chips, and what matter is the cost of capital. Uh, this is dominating the, the, the bill of material. So it's a problem. It will become more and more of a problem as we deploy more compute in Europe, uh, because we'll run into supply constraints. Again, it's not evenly distributed, so if you look at France, for instance, there are excessive energy, so you can build capacity. Uh, if you look at Nordics, it's also the case. In the UK, it's a little less the case. Germany, it's a little less the case. But there is a realization that AI is going to drive, uh, energy consumption forward. I would say in the next two or three years, the true, uh, the true, uh, fight, the true challenge is to actually make sure that for Mistral, make sure that we, uh, build as fast as possible to use the exist-- excess of energy that is already there. Uh, and so that is what we are at, at work doing, raising capital, building more, building the team to build all of this capacity. There is excess of energy. Uh, it's not the bottleneck today. It will become the bottleneck in '29, 2030. Uh, but right now the question is: How can we build on existing energy? Now, we need to think ahead of time, uh, in starting in '29, 2030, as multiple gigawatt of capacity are being deployed every year on, uh, on, on AI. We do need to ramp up the production. Um, what, what we are hopeful for is that in France, for instance, there is an enormous, uh, EPR program where, uh, nuclear energy is going to be produced mu-much more. You have large, uh, efforts on the solar side as well. So I would say it's going to be a problem. It's mostly a coordination problem because at the end, if you have demand coming from AI on energy, the business case for building new nuclear plants is pretty well made. Uh, it's-

Harry Stebbings 15:53 It totally, it totally is, and I agree with you.

Arthur Mensch 15:55 Yeah.

Harry Stebbings 15:56 I am a big believer in Adam Smith's invisible hand, but the lead time is very real.

Arthur Mensch 16:01 Yeah.

Harry Stebbings 16:01 And so that is just-- Uh, the 29, 2030 i-is a challenge.

Arthur Mensch 16:05 Yeah. Nuclear plants, uh, take longer than that to build. Uh, solar plants can go faster. Uh, and say I do believe in that invisible hand. The states have to play a role, uh, in facilitating, uh, the build-out, but also in planning ahead of time. They, they have a long, uh... They have a longer-term vision than en-enterprises in general. And so they-- We are spending some time, I'm spending some time explaining that, uh, the energy demand is going to grow, uh, and that it needs to be addressed somehow. It needs to be part of the business plan of, uh, of any energy supplier.

Harry Stebbings 16:42 I just wanna take it in turn. You said about the cost of chips being important. Can I ask you, what do you know now about compute buying, forecasting demand of compute chips that you didn't know?

Arthur Mensch 16:55 It's quite a process. Uh, so we've learned to run that process to match supply and demand. Uh, it's a complex industrial process. It has nothing to do with software. Uh, you need to understand the latency in the supply chain. Uh, you need to understand that the supply chain comes from multi... I mean, supply chain is fairly complex 'cause you have GPUs, but you also have CPUs, you have storage, uh, you have electrical parts. Uh, you need to make sure that your own suppliers have no problem with their own suppliers. So that aspect is something we actually learned, but it's a, it's a fairly complex one where you need to manage resources. Uh, you need to have multiple forecasts. You need to be able to forecast your demand for multiple buckets. So enterprises, which are our main ICP, but also AI natives that can, uh, take larger piece of GPUs but want them for a shorter period of time. Uh, you have, uh, companies like Microsoft, uh, that are off-taking capacity from us that we can, uh, use to finance more capacity. So at the end, uh, the problem of building a cloud is about matching supply and demand. It's about managing the supply chain ahead of time, giving visibility to your suppliers, and it's about matching all of this with the finances. Uh, you need, uh, project financing. You need people that are strong at understanding credit. Uh, and so it's, uh, it's actually quite interesting, uh, in, uh, to, to build such a business. What we've learned is that, um- At the end, if you, if you're able to front capacity ahead of demand, you get something that is pretty easy to sell. Uh, and it's going to be the case for the next few years. And so we're very grateful that we anticipated the fact that we would run out of supply fairly quickly, that our customers would have supply chain problems for their own GPUs, because now that we have those GPUs, we can serve their needs.

Harry Stebbings 18:47 Totally get you there. That, that was one, which is cost of chips. Cost of capital was another one. If you were resource unconstrained completely, what would you do today that you're not doing?

Arthur Mensch 18:59 We would build much more capacity. I think that's a, a big bottleneck in, uh, in build- in building capacity is, of course, capital. And we see the demand, and we know that our customers are, uh, running out of compute. Uh, they-- As we take some of their use cases and we help them scale it, suddenly they realize they don't need eight GPUs, but they need thousands GPUs, and that's for one use case. And so the-- we see that demand ahead of our customers, so we have pretty high conviction. That's why we put equity at work to build capacity. And so, uh, what we are working on, what we have always been working on, is to make sure we have more capacity under our management, even by, by building ahead of demand. Uh, and today, if we just look at what Europe needs, it's pretty clear that there is not enough compute. Uh, you may argue that there is maybe a little too much compute being built in the US, but what is clear is that there is a lot of slack, uh, to build capacity in Europe in a way where you are 100% confident that it's going to be allocated. So that's, uh, that's what we are working on.

Harry Stebbings 20:03 But where you said, "Hey, there's a role to be played by the state as well," what would you like to see the state, be it France, be it the UK, do that they're not doing today?

Arthur Mensch 20:15 I mean, we've been quite vocal about the kind of ways in which they can, uh, intervene in a way which is, uh, market neutral, uh, and that just would facilitate investment. Uh, the first thing is, of course, to facilitate permitting and to, uh, make sure that there are dedicated sites where things are accelerated, uh, where, uh, all stakeholders with a regional or state level are aligned. And I would say France, uh, Sweden, uh, are making great job at that, for instance. So that's mostly on facilitation of aligning stakeholders, 'cause these are physical assets. Uh, they are built, uh, within regional communities, and it makes... it's super important to make sure that everybody is benefiting from it. And so the, the, the help of states happens there. What they can do at the financial level, uh, and we've said that, uh, publicly, is that state guarantees on leasing data centers, for instance, on reserving space for ten or 15 years, are super useful for companies like us that have relatively small balance sheet and that knows, uh, how to build and who to build it for. Uh, the-- So that's a role they can play by giving some guarantees, uh, putting some guarantee on the table to build further and ahead of time. Then, of course, public offtake. Uh, we see, uh, public sector is, is an, an important customers for us from a lot of different states in Europe and outside. Uh, and again, we bring them value. They bring value to the citizens. It creates offtake, uh, for companies like us, and that allows us to build further. And if you build further and faster, you build cheaper, uh, in this day of supply, supply constraint.

Harry Stebbings 21:58 Can I ask, you said you sell to states that, and kinda going back to that element of kinda sovereignty, how, how far down does sovereignty need to go? And what I mean by that is, like, we obviously have Cohere in Canada. Does, like, every country need their own model? Is it blocks that need their own-- H-how do we think about, like, the granularity of sovereignty?

Arthur Mensch 22:17 I think the granularity that is allowed is to... You need a fairly large market to build, uh, relevant models. Uh, and today, Mistral is able to build relevant models because we operate not at national level. Uh, we're a French company, but we serve Europe, we serve the US, we serve, uh, Asia. And what we need is alternatives, but there will not be a single sovereign AI provider and model trainer at every state level. This is not going to happen.

Harry Stebbings 22:43 So what percent of your customers are European?

Arthur Mensch 22:46 If we look at, uh, our revenue base, it's around 60%, uh, and the rest comes from the US and comes from APAC.

Harry Stebbings 22:53 And they use you to do post-training harnesses for their own models?

Arthur Mensch 22:58 Yeah, European ones, they use us for multiple thing, for multiple reasons, but we help them build applications, uh, on, uh, either things that are related to the efficiency of their businesses or related to accelerating the R&D. We work with manufacturing companies, with financial services. We work with defense companies, and we have pretty strong verticalization approach to all these three, uh, domains, uh, and the fourth being public sector. The way we work with manufacturing companies is, in general, we are, we are, we are with their R&D teams, and we help them build better products. So we help ASML build better machines, operate better machines. We help Airbus build better planes. And we bring reasoning models, but we also bring the physics models, the models that have a quantitative understanding of the world. In financial services, we build models and harnesses that understand markets. We also bring models that understand quantitative time series to help do prediction. So we have, uh, we come with a set of models that allows to quickly build applications that bring value to our customers. Uh, so that's how we serve them. Um, and I guess that, back to your point on sovereignty, uh, the-- all of these companies have a sovereign AI strategy. They need deployment of the technology in a way where they have business continuity guarantees, affordability constraints, et cetera.

Harry Stebbings 24:19 We see, uh, OpenAI, Anthropic, DeepSeek move into application layer and also move into chip layer with Halapeno, with Anthropic's rumor, DeepSeek are doing their own. Will we see a Mistral chip? How do you think about the importance of verticalization and owning every layer?

Arthur Mensch 24:36 No, it's not planned. We have gone down the stack, uh, in-- or I would say upstream, if you look at it from an industrial perspective. We have gone from training models to serving the models we train on infrastructure that we own. Uh, so that's, that's already a pretty significant move. Uh, that's a move that is now... that took around 18 months to, to set up, and that is now pretty successful because we can serve the GPUs, uh, we can serve tokens, uh, we have, um, tokens that are purely Europe-generated, in that it escapes any form of US extraterritorial law. It matters a lot for our customers. So that took us some effort. We are doubling down on this effort to scale that further. Uh, and I think that's going to keep us busy for a while. Uh, so the-- we are pretty vertically integrated in that we, we have, uh, cloud services infrastructure, and we go all the way to building the business applications for our customers. You mentioned the application layer. In AI, you bring the platform, and you build custom applications for customers. That's how it works. That is what AI allows you to do. So-

Harry Stebbings 25:42 And so this is, this is through an FDE army that you're-

Arthur Mensch 25:45 FDE.

Harry Stebbings 25:45 Yeah?

Arthur Mensch 25:46 Yeah. It's FDE, but building on top of common primitives. So, uh, workflow management, uh, agent, uh, agent management, uh, management of sandboxes, observability of the deployment of different business applications. So at the end, we are building cloud services that are adapted to run and build with AI agents.

Harry Stebbings 26:06 How do you think about doing it through Mistral FDEs versus partnerships economy, which we've traditionally had?

Arthur Mensch 26:11 It has been super helpful to have our FDEs front and center at the beginning of our journey, uh, because it allowed us to understand exactly how, where enterprises were struggling, uh, how they started with individual productivity. It was not enough. Chatbots were not delivering any kind of value creation for them. And so we went and spent some time with the managers, the business unit, uh, owners, and understood the kind of core processes they needed to automate with artificial intelligence. Uh, we did it in a way where we built business applications, custom, uh, applications that are manag-managing state, uh, applications where agents are, are operating the process end to end. And it was super helpful for us to understand the primitives to build. So we built durable process execution, for instance. We understood how to orchestrate agents in a way that was very effective. So having those, these FDEs, uh, our FDE, uh, sit with customers directly has been super, uh, useful. Now, what we have done since last year is to start propagating the tools to some of our partners, uh, like Accenture, like TCS. So system integrators, uh, that are great at scaling things, uh, and that once you give them the right tools, once you give them the right training, uh, they can help scale because they have just more people. Um, but having people actually sitting hands on keyboard, uh, from Mistral with our customers is super useful for our customers, uh, and it's in high demand. Uh, it's super useful for us. It's not something we can scale indefinitely, uh, because we can't have tens of thousands of FDEs working for us. But it's something that now we are scaling through partners, and the two motions coexist. So we are-- in many accounts, we have FDEs that are working directly with customers. They are operating, uh, the projects. They're making sure that the delivery is occurring, and we work hands in hands with system integrators to do, uh, to do the work, to then do the maintenance and these kind of things.

Harry Stebbings 28:12 Can I ask you, what line of revenue do you not have today that you think will be most significant in five years' time? When I was listening to you there, I was thinking, "I bet you wouldn't have necessarily predicted these different motions when you started." [chuckles] When we first met, Arthur, where, I don't know if you remember, but you were sitting on a park bench just chatting to me about life. [laughs]

Arthur Mensch 28:31 I was, yeah. You did not invest. You should have. Uh, but-

Harry Stebbings 28:34 Dude, I fucking regret it so much. [laughs] I, I, so much. I, I, I saw Jean-Charles yesterday, and every time I see Jean-Charles, we talk about you. Um, and, um, yeah, it's, it's a brilliant reminder of just, you know, how much money I lost or missed out on. [laughs]

Arthur Mensch 28:50 Sorry about that. No, what line of revenue... I mean, we have these two lines of revenue. We have one which, where we deploy the platform that is truly licensed revenue, uh, for building business applications, and then we have the consumption revenue coming from the infrastructure business that we built. And the two things go hands in hands because we are able to, um... Our customers, they, they come in general for the licensed motion, but then they realize that, uh, they don't have enough GPUs, uh, and then we are trying to be helpful to them, uh, by providing them with either GPUs or, or token services. So the two things are very well combined together, um, going to-

Harry Stebbings 29:25 Do you think you will-- do you think you'll have an inference business, a, a big inference business in, in three to five years' time? Do you think you'll have a sandbox business? What-

Arthur Mensch 29:33 Uh, yeah, if you look at this granularity, I would say the GPU side, like everything sitting on GPU, is going to continue to grow because the demand is enormous. You have what we call ancillary services, uh, like running sandboxes, running-- scaling up and down sandboxes very easily when you're running, uh, reinforcement learning algorithms, for instance, or when you're deploying many agents on the same tasks. Uh, that is a growing business for us, uh, because it comes from the demand of our customers. And in turn, that brings us to buy more and more CPUs. Uh, so I do expect what we call the runtime, uh, side of things, so, uh, where agents execute tasks, call tools, uh, uh, compile code, and run the code to grow. Uh, it's pretty clear that, uh, when you are in synchronous task, where you are only using GPUs to, uh, answer conversations, you didn't have that. When, when you start to have managed agents that are running on the background, uh, doing a lot of tasks that are CPU-intensive, you need more CPUs, you need more sandboxes, and you need, uh, more storage as well because they produce a lot of data. Uh- So those two things, I would say CPU-related tasks, data-related tasks are going to be growing line of businesses for us

Harry Stebbings 30:47 A big part of the business is selling to, to governments. Um, what does no one know about selling to governments that everyone should know?

Arthur Mensch 30:55 And the first thing is that, uh, there's no such thing as selling to governments in general. Uh, all governments are very, are very different. Some governments are small enough so that it looks more or less like an enterprise sales motion. Uh, it's the case of, uh, governments where everybody in ministries know each other, for instance. Uh, other governments in larger countries are slightly more complex. So again, the complexity of finding where to create value, of understanding what the customer needs grow with the size of the country, I should say.

Harry Stebbings 31:27 Can you help me on the inference market? It's something that I just totally didn't know enough about. Like, what I know now that I wish I'd known 12 months ago was the, it's just infinite demand for inference. It's just the depth is astonishing. What happens with the inference market?

Arthur Mensch 31:45 I mean, the inference market behaves very much like a, a commodity market, which is pretty high margin today. Uh, in that if you have supply, if you have the right quality of service for certain kind of models, you can sell them on demand, uh, because you have a lot of demand coming, uh, directly for cu- from customers that are non-committed. And then for customers that are sufficiently mature, uh, they need to have guaranteed quality of services. They need to have service level agreements, and there they need to have reserved physical capacity, and that's, uh, something that we can provide because we have the right forecasting models and we have the right supply. So you have these two things. You have one which is on demand, um, which very much behave as, as a commodity market, where pricing matters a lot. Uh, and then you have, let's say, a premium, uh, side of the market, where the customers actually want to have access to-- want guarantees on latency, want guarantees on the speed of generation for a certain number of, of, of requests. And the more they can be precise about the shape of their demand, the better our pricing can be, uh, the better we can shape our own infrastructure to make sure that they, uh, have the right quality of service for the right price.

Harry Stebbings 33:00 Can I ask you, we have so many different neo-lab providers. There've been 102 different neo-lab providers created. Every week there's another DeepMind spin out and another. There's so many. What will separate neo-labs that thrive and those that don't?

Arthur Mensch 33:16 Well, the first thing, uh, I mean, we started as a neo-lab ourselves. It wasn't called that way in 2023, but, uh, uh, effectively, you know, when you hire-- when you start with, um, when you start a company, you have a bunch of silver bullets to hire the best, and we used all of our silver bullets, uh, to hire researchers. And so effectively, we started as a research company. I think what has allowed us to scale and to now operate in a much wider way, uh, still with this number one pillar, which is to deliver models, uh, and that's why Star Large 4 has been very successful for us. Um, and I think it's very exciting a-as well in the, the new releases that are coming 'cause we're making pretty good progress. What has been successful is that we focused early on on a market that would scale. Uh, so we went for the enterprise market, which comes with a certain number of complexity, uh, because of course, uh, enterprises customers are, uh, demanding, as they should be. Uh, and we, we have adapted our product to meet those demands. But the fact that we were able to both operate at the research level, at the highest level, and, uh, able to constantly recruit and invest further in making stronger and stronger models, and the fact that we built a customer-centric motion where we understood the problems of our customers, have allowed us to scale. And that's something that every neo-lab needs to think about when they start building

Arthur Mensch 34:39 their business.

Harry Stebbings 34:39 But how many people are there actually in Europe who truly have the ability to create these companies and have, have been it and seen it like you have? Like, I don't know, maybe a Yann LeCun has done. Is it 20? Is it 100?

Arthur Mensch 34:54 I mean, I'm not a VC investor, so I would... You probably know it better than I do. Uh, but there's a, a lot of talent in Europe generally. You have a lot of entrepreneurial talent. Uh, you have a lot of, uh, researcher talent, and you have a pretty good combination of those two, so that's why you see a lot of neo-labs, uh, in Europe. Now, the, the hard thing is it's not only about being talented, it's not only about being able to sell, uh, and having something that you can successfully create a business with. It's also today about being able to access compute. Uh, I think a big bottleneck today for many neo-labs that I know is that, uh, accessing GPUs, uh, to actually do interesting research in certain verticals, uh, is hard because there's an e-enormous supply crunch. And so that's why in the supply we have, we are reserving some capacity for certain neo-labs because we believe that, that they can grow and that we can grow with them. Uh, but effectively, if you're a founder today and you wanna do something which is very research centric, the biggest problem you have is to get access to enough compute, uh, with a decent price.

Harry Stebbings 36:01 Can I ask you, you've mentioned the word researchers there. How crazy is the researcher hiring market today?

Arthur Mensch 36:09 Well, it's pretty competitive, [chuckles] I should say. Uh, again, if you have the right mission, uh, if you have the right compensation structure, uh, if you're doing something that matters for the world, uh, people will join us, uh, will join you. Uh, and I would say, uh, we're very grateful, uh, for, uh, the dedication, uh, of all of our employees to making, uh, Mistral successful. Uh, the quality of the Mi- Mistral Large 4 is, has been a success, uh, has a lot of traction today and is leading the pack in open-weight, uh, in, in the open-weight category and coding and in cyber defense. It's only due to the dedication of this team- Uh, in the last three, four months. The resilience as well, because when you run a model, uh, we've, we've posted our tech report, uh, recently. Oftentimes it doesn't work the first time, and so you need to be able to course correct, uh, in a way where, okay, you're going to spend two more months to actually make it work, and it's going to be painful. Uh, we're expecting to release, but we'll need to release a little later. It's going to be painful for everyone, uh, but everyone needs to, uh, be all hands on deck, uh, and do the right thing. I mean, I'm very grateful, uh, for the resilience of the team, uh, given the, the scarcity of

Arthur Mensch 37:26 resources that we have.

Harry Stebbings 37:27 I just had Satya from Microsoft on the show-

Arthur Mensch 37:28 Yeah

Harry Stebbings 37:28 ... and he talked about Mustafa's statement before, that we'll see $100 billion training runs in the next few years. Do you think that's a little bit hyperbolic, or is that actually well within reach?

Arthur Mensch 37:40 You do have data world before, I should say, uh, in that, uh, at some point, if you want to make your model very strong at certain domains like chip design, for instance, or, uh, if you want to make it very strong at operating industrial engineering workflows like we do, uh, you do need to have the right environments to train them. And this part, uh, takes some skill and taste. It's not something that you can scale as easily as just stacking compute. Uh, so I would say the, the bottleneck, uh, that we're all running into today is effectively to be able to pr- to source enough, uh, quality data. It's compute. We need more compute. We are deploying more compute. Uh, getting to 100 billion of, uh, training runs, I think, is not for tomorrow, clearly.

Harry Stebbings 38:24 If everyone's so short on compute, why, why is Nabuurs 40% down? Why is Kohui 40% down? Why, why is the AI trade cracking in public markets if it's so clear that everyone is so compute starved?

Arthur Mensch 38:40 It's a little unclear to me, uh, how, uh, markets are reacting today. What we see is that, uh, we're fully able to finance, uh, our, the, our own compute given the quality of the collateral that we have, uh, in term of, uh, of, um, of revenue commitment and corporate commitments. We, as a company, operate as a vertically integrated, uh, company. So we build compute, uh, but for higher level purposes, for serving the enterprises, uh, for building business applications. So we are anchored on the value creation of our customers, which means that at the financing level, we haven't had to go and leverage enormously the company on debt financing. Uh, and so that gives us a lot of confidence in what's going to happen. But effectively, as in any industrial build-out, uh, it's tempting to get a lot of leverage, but at some point, leverage break. And so that's, uh, why we're very happy to be vertically integrated and to operate from business, uh, application creation all the way to, uh, the infrastructure part.

Harry Stebbings 39:48 What did you not do with the benefit of focus that with hindsight you go, "Mm, I wonder if we should have done that"?

Arthur Mensch 39:56 We did not go deep into the consumer market, but, uh, I think we're very grateful that we did not do it. We did not go into chip design. I think that's, uh-

Harry Stebbings 40:06 Well then, pause. So what, like, why, why do you feel grateful you didn't go into the consumer market?

Arthur Mensch 40:12 I think it's a very packed market, and it's very hard to operate at the highest level with enterprises and also to run what is inherently going to be an ads business. So, uh, that part, that decision to focus on the enterprise segment, also that decision to focus on the four verticals that we have decided to focus on, manufacturing, financial services, defense, public sector, has been very critical to focus the company. 'Cause the problem with artificial intelligence is that it can do everything, uh, and you need a source of focus. Uh, and so the source of focus we've picked is, uh, comes from the market, so we wanna have the models that serve the best agentic engineering workflows, uh, that serve the best, uh, financial decision-making, uh, that serves the best defense purposes and everything related to state sovereignty

Arthur Mensch 41:00 as well.

Harry Stebbings 41:00 With r- with RSI, do you need a source of focus? Is that not-

Arthur Mensch 41:06 I mean, using AI to improve AI models do require to focus as well, because you have limited amount of resources to spend on problems. At the end, the bottleneck, uh, here is indeed compute. Uh, the bottleneck is also that everything that you do that, which is recursively self-improving or just for self-improving, because it means the same thing, requires to have well-defined problems. Uh, if you have a well-defined problem today with artificial intelligence, you can later climb it. In many cases in enterprises, the problems are not well-defined. Uh, and so the true friction to economic growth is not the fact that we cannot solve hard problems that are well-defined. It's that we-- it's hard to solve ambiguous problems where you have ambiguous objectives, where you have humans in the loop, and that part you cannot really solve with self-improvement. You can only solve with models being deployed and interacting with humans, and that's the part that we are, uh, that we are focusing on, I would say. But we, we are investing a lot on making self-improving loops on well-defined problems as well. But that is the part that is constraining the explosion of AI usage, I would say, in the enterprise. The poorly defined problems on which you want to def-- to deploy AI, but you have no way to lay, lay the claim because the definition of the problem is not proper. And so you need to spend some time. That's why you have FDs. That's why we have FDs, to try and better define the problem.

Harry Stebbings 42:32 Arthur, I've gotta be blunt with you. I s- I joked at the beginning that you're like Europe's last hope. Um, there is a... And I don't mean that, like, super dramatically, but there is a very big chasm between you and anyone else in terms of Neo Labs that, that will aspire to be the next Mistral. Do you and the team feel the pressure?

Arthur Mensch 42:51 We do feel some pressure, but as anyone in the field, uh, this is, uh, fast-growing, fast-moving. Uh, you need to, as an entrepreneur in AI, you need to deliver on the model side 'cause that's the one thing that shows, uh, the progress you're making. Then enterprises need to see the value, so you need to have people that are, uh, understanding the true business problem. So it's, it's a fairly different problem to solve. So we do feel some pressure, but I think it's a very healthy pressure. Uh, it's a very important moment in time and in history. Uh, we need to be, uh, meeting the expectations of the general opinion. At the end, we need to show that AI is useful. Uh, the kind of deployment we're making, the kind of noise this is making in the market, uh, this technology needs to appear as a positive technology. Needs-- We need to show the positivity of that technology. Uh, and the only way in order to, to do so is to promote sovereign AI as a category, uh, is to make sure that, uh, that creates economic growth and opportunities for everyone, to make sure that the gains of AI is well distributed. And so, of course, it creates some pressure, uh, because it's not a given. Uh, if, um, if we don't succeed, I don't know, uh, what's going to happen in the world. I think we-- so we do have a role to play. I think the mission we play is important, and I'm, uh... It's pressure, but it's also quite exciting.

Harry Stebbings 44:17 Will AI inflate GDP growth by more than 2% a year?

Arthur Mensch 44:22 Well, that's the, that's the entire question. What kind of a bottleneck do you go and solve to actually inflate the GDP? That's the reason why we were very bullish on industrial engineering. Um, at the end, if you want GDP growth, you need more production. Uh, you don't need more intelligence. Uh, you need more production. In order to have more production, you need better factories. You need objects that are more efficient. You need better planes that are flying for longer with the same amount of oil. Uh, you need better energy production, uh, that are more efficient, less carbon intensive. Uh, you need, uh, better ISML machines. You need, um, better services also in financial services. So in industry and engineering in particular, uh, using AI allows you to do better design of objects. So if you set up the agentic loop to design safer cars, for instance, you can have your car design team go much faster and iterate faster on that product. Uh, that means that you can sell more. That means that you can sell, sell more cheaply. Uh, that means that you're building safer products. In the aerospace industry, uh, if you design a better plane, it's going to be out there for forty years. Uh, if it runs faster with less fuel, uh, then it's going to create GDP growth. So at the end, what truly matters is to enable AI to have a physical effect. Uh, and this effect we find mostly in manufacturing and financial services.

Harry Stebbings 45:52 Huh. Dude, I'm aware of time. Uh, I'm gonna do a quick fire round with you. Okay?

Arthur Mensch 45:57 Please do.

Harry Stebbings 45:58 Okay. What do you believe today that most around you would disagree with?

Arthur Mensch 46:04 I think we strongly believe that this is a fragmenting market, uh, and that sovereign AI is going to be the most important category, uh, of AI deployment in the enterprise. In that when enterprises are going to deploy the technology, business continuity, cost, and control of their data and IP is going to be the one thing that matters. Uh, so I think that's a very strong conviction that we have, uh, that is shared by a lot of enterprises and sh- and shared by a significant part of the market, uh, but of course not shared with some of our competitors.

Harry Stebbings 46:35 What is a extremely wild prediction today that will be very common in five years' time about the way that we live, the way that we work?

Arthur Mensch 46:44 I think the, the way we work as engineers, uh, has already changed for software engineers, is going to completely change for any kind of engineers. So I'm super excited by, uh, that, uh, approach of designing better products, uh, in a way which is fairly autonomous, uh, thanks to artificial intelligence. So it does not concern everyone, everybody, but everyone with a technical background, uh, and not only the software engineers that were the first affected and the mathematicians. Everyone with a technical background will be operating AI agents in a way that will give them enormous leverage. So this in itself is going to m- be a major shift in the, the way the labor market

Arthur Mensch 47:28 is actually working.

Harry Stebbings 47:29 What do you do about going public? 'Cause one day you will want to, okay? But y- y- do you-- you can't go public in Europe.

Arthur Mensch 47:38 You can absolutely go public in Europe. Uh, well-

Harry Stebbings 47:39 Oh, come on

Arthur Mensch 47:40 ... the first thing, yeah, you need to-

Harry Stebbings 47:41 Come on. You're gonna go public in France? No.

Arthur Mensch 47:44 What used to be true ten years ago I think is no longer true. There's a huge appetite from investors from across the world to invest in Europe, both on the private equity side, but also on the public equity side. And so the bottlenecks that used to be there, uh, are no longer there. So we at Mistral, uh, are of course getting prepared on the-- you know, you have a lot of operational constraints when you, when you go and become public. Uh, we're also pushing and, and delivering on the, the commercial milestone that we, we, uh, need to deliver on. But, uh, I don't think this is any more a bottleneck. Uh, and there are widely successful companies, uh, some of them are stakeholders, that are firstly, uh, listed in Europe. So there is no, uh, reason why not to go and, and an IPO in Europe at some

Arthur Mensch 48:31 point.

Harry Stebbings 48:31 Final one for you, dude. What are you most excited about when you look forward to the next five, seven, ten years that not many people think about?

Arthur Mensch 48:41 I'm quite excited by how AI is going to enable foundational research, uh, in domains where we lack understanding of the world. So, uh, in physics, for instance, uh, the fact that we can both run AI simulations at a scale that was nowhere reached, uh, and that's the team we bought, uh, called AIME, uh, is specialist in that, is going to enable new discoveries. And so the fact that we can build, uh, simulators and that we can have agents operating those simulators to understand better the world is going to be an enormous enabler of the understanding of the world. So I come from a, a math and physics background, and so I'm quite excited by the opportunity ahead of us in the next generation of researchers.

Harry Stebbings 49:26 Do you know what, Arthur? I blame you. I blame you when you sat on the bench and you didn't speak at all like you speak now. Like, the Arthur then and the Arthur today are just, like, they paradoxically different with respect. It's amazing to see the leader that you are and you've become. I, I, I honestly, dude, I have so much respect, admiration, gratitude for you and all you've done for Europe. But wow, it's amazing. [laughs]

Arthur Mensch 49:52 Thank you for all the praise. There's still a lot of things we need to deliver on, but, uh, we are quite happy with the progress we've done. Uh, I think the mission is important. Uh, and so if anyone watching us, uh, wants to join us in that journey, uh, please, please, uh, contact me.