Mistral AI

What companies are leading in innovation in Europe in 2026?

Europe's Most Innovative Companies

These 50 companies exemplify the innovation, progress, and potential of AI.

AI Innovators 2025-Mistral

Fortune 50 AI Innovators

Sector

Country

Strongest dimension

Years on list

Previous rank

Get insights into Mistral AI including business profile, stock information, news, and rankings from Fortune.

Mistral AI Company Profile, Stock Price, News, Rankings

legacy

mistral ai

Companies

Mistral AI | Fortune

Mistral AI Company Profile, Stock Price, News, Rankings | Fortune

AI models are global. The data, rules, and power behind them are not. As countries in Europe, the Middle East, and Asia push for a more sovereign approach, a new generation of companies is forcing a deeper question about the chips, data centers, data, and models that make up the modern AI stack: Who gets to build AI, and on whose terms?

Dr. Daniela Braga, Founder and CEO, Defined.ai 
Timothée Lacroix, Co-Founder and Chief Technology Officer, Mistral 
Jeremy Kahn, AI Editor, Fortune 

Whose AI Is It Anyway

Conferences

So sovereign AI is is not, it's a full stack problem. It's not just a problem with models and where they're made. It's a problem with data. It's also a problem with data centers and with the hardware that runs in those data centers. Uh, Timothy, let's start with you. Uh, Mistral is known for its AI models, really. I think that's, that's probably what most people know you for. But recently you've started to kind of move into up, up and down the stack into your own data center infrastructure, um. But I guess those chips, you know, in those data centers are still, they're not European made chips, um, so I want to first of all explain, uh, have you explain a little bit what the strategy is here from Mstraw, and then maybe you can talk a little bit about, you know, are there limits to how sovereign any provider can be if you're not in the US or China? Sure, um, so thank you for having me. Uh, we started the company around models, uh, because we were researchers and models was, uh, what we knew how to, what to build. Um, we built these models as open source, uh, because the strategy from the start was to be open and to give control around what we build. We wanted to quickly apply those models, uh, to, uh, give them impact in enterprise, but just chucking the model weights, uh, around the, the world didn't really provide that impact. And so we started building uh an entire platform. We started building a service arm. Um, and more recently we, uh, as you said, expanded in the domain of, uh, data centers where we have, uh, 50 megawatts of compute coming online this summer. Um, to me it's really about giving our customers, uh, the choice, uh, on all layers of the stack for where they want to run what part of their workflow. So typically if you want your, um. Agents dealing with private data to be on-prem, that might be very good. If there are more agentic workflows that require large inference capabilities, uh, you might want to have something that's hosted in your country for strategic autonomy. And so, as you said, chips are still very much on video in our case, um, and today there is no equivalent in Europe. I'm sure there are many alternatives that are going to be built and we'll help them out if that's the case, but today it's really about giving control where we can. Interesting. And are you, are you discussing a European chip with anyone? Are you thinking about doing a chip? So we are, I think it's a bit early for us to be doing chips, and I think the transformer architecture is well known enough that if someone designs a great chip for transformers, we'll be users, but we don't need something that's too specific for ourselves. There are European companies that are. Up and coming and in that case we like to talk to them and help them out, but it's always a longer game. The company's been around for 3 years. I think that's about the time it takes to deliver. Absolutely, Danielle, I want to bring you in here. How does data play into the debate about AI sovereignty? You know, most of these large language models, they've been trained from data scraped from the entire internet, but where does more curated data sets come come into play? Absolutely. Well, uh, we are the 3rd pillar. There's the computers, the models, and there's the data, and we're doing this for 11 years. So data continues to be the differentiator in models, especially in a world where everybody has a lot of data centers, everybody meaning the largest AI labs, and, but they don't, and everybody has Access to publicly available data, which actually causes a whole other problem in terms of copyrights and bias and diversity of representation. Uh, but everybody has the same access. So how do you differentiate your models is on closed source, uh, copyright clear data that you can trace the provenance, you can trace, you can, you can, uh, uh, fill the model gaps with it, uh, and this data in our market, we have the largest closed source marketplace, risk free, keep saying that, that's a big difference here, uh, from Scraping the web, we don't scrape the web. Everything is a rev share with our partners. Uh, we also produce our own data as well. And, uh, and that's really the that even at the, not only at the pre-training level, but at the fine tuning level, uh, you will need both, uh, both axes. Interesting, and we we've seen data labeling companies become really important in part for reinforcement learning, um, you know, getting very specific data on, you know, is this a good answer or not for this, you know, professional use case for Mistral, are you using that kind of data? Are you finding that you're, you're having to source your own data for, for reinforcement learning? Yeah, definitely it's becoming the data. That we use for training models is becoming more and more complex today. It's about not just data, it's also about entire environments. So it's not only the question, but it's also the full environment where that question might get answered and so it becomes really specific. We do work typically where we get data from different companies to simulate questions on real actual questions. Um, we, yeah, there are lots of softwares that that can be simulated as well. Yeah, and Danielle, are you seeing the type of data that's valuable change? It has become more specific because the generic models are performing already decently well. So now you're going more into specialization, you need PhDs in all the 20 knowledge domains or plus that you need, ah, you need real data from enterprises production. Uh, you need, um, medical data, uh, physical AI data, but in the, in the sense that you're gonna apply to, uh, in specific devices and so it's becoming, it's becoming really more. Specific Yeah, interesting. I'm gonna come to the audience for questions in a minute, but, um, so please think of your questions. Uh, for, uh, Timothy from Estral, to what extent are you really leaning into sort of the idea that you aren't American and how is that, has that become a selling point for Estral? Uh, I think it's not really about not being American. I think the goal is to be a global company. The one, thing that we do use is that we are seen as more neutral coming from Europe and so. Providing control and being from Europe, I think is a good view for folks that want to be, uh, yeah, strategically autonomous. Right. Um, let's go to the audience for questions for our guests and I don't see any right now, I'm gonna, oh, there's someone there. There we go. Please state your name. OK. Hi, I'm Ali Warson. I'm a partner at an early stage venture capital firm called Up Partners. How do you decide the quality of the data inputs? And I know that there are data brokers, um, and there are, and, and large foundation models are paying hundreds of millions of dollars to these companies. How do you choose which companies to engage with and how do you decide? Yes, this is, this is the quality that we're looking for. David Smith, are you using data brokers? So yes, Guillaume is the person that would best answer this, but I think the short answer is data is a bit of an annoying thing to work on for scientists and engineers, but it's the most important thing, and that's how we've built Mitral to be efficient. It was by focusing on data. The, the answer for most of the quality checking is either go through the data, sample it, see if it's, if it looks good, but then the longer answer is you also have to try and train the models and see what impact it has on the downstream tasks. Danielle. Yeah, well, it comes with uh data consistency, relevancy in the marketplace. I mean, if there's, if there's no point if there's a lot of it already, it's a lot of the Uh, that competing factor, uh, data in our AI world deprecates value over time. So what is hot today has a two year window in the of, of self, uh, um. Interest and then it starts to decay. It doesn't mean that someone later on will want to do that again and but the value comes can come down 10x in a year. So if you, it's a, it's a constant and of course, it's the, it's metadata is the key as well. Uh, also, of course, the legality, we spend a lot of cycles with our partners just to make sure that everything is Uh, bulletproof from, uh, from copyrights and GDPR and privacy, right, other questions. Oh this is one in the back there. Hi, Brad Foreman, the Journalist Group advertising and Crux Apex. Um, I'm just wondering in terms of the data that you're selling, so how easy is it to use that as training data, because I think you're, you're talking a little bit about that, um, and I'd just love to hear more about how is everything training data on there. Yeah, well, it's everything. There's 2, there's two sides in AI modeling building. There's pre-training data, especially in the world that we live in today, retraining is off the shelf, at scale, ready to go, licensable data, well, licensable for the license of business that we operate. And there's post-training. In the post-training, you have custom data collections, You can maybe slice and dice pieces of the pre-training universe, and you have the labeling aspects of it and of course, more of the, what we call the interaction of model evaluation, people in the loop interacting with models live or not live, but that's another, it's more the human intense side of it. Uh, yeah, you, you need, it depends on where you are on the stack. Most of our clients are in both sides of the stack. They, they start with the pre-training and but they are moving to the next application, the next product, and they need another boost of pre-training, whether the new thing now is the point of view, egocentric type of that if you are building a model that is reflecting our cognition from a vision standpoint. So, It's, it's all the, it's all the sides, and they go hand in hand, right? Uh, Timothy, I, I want to ask you, uh, for Mistral, you know, you're, you're, you're competing out there against models from anthropic, models from OpenAI. These are companies that are growing very fast, have, you know, huge, uh, huge amounts of venture capital behind them. Uh, how do you, how does Masral really look at competing against those kind of really big players? Yeah, so you've mentioned that we're going very fast. We started 3 years ago. We are now from, uh, we have offices from Palo Alto to Singapore. We're over 1000 employees. So I would say we've done quite a good job ourselves in terms of speed, and we have banked on a slightly different approach which keeps us closer to. Um, the harder, deeper problems of the enterprise that we, uh, collaborate with because by allowing our entire stack to be deployed on-prem, uh, or wherever the customer wants, it's, it really means there is nothing that's uh out of bounds for what we can deal with since AI is really. About aggregating an entire company's knowledge, I think having no limits to what we can deal with has been a real accelerator for us. Got it. And those other companies I mentioned, Anthropic, OpenAI, as you know, are anticipating IPOs very soon. Are we going to see a Mistral IPO? So Arthur has been public about the fact that this was the goal of a company, and I think it's a, it's a nice goal to have. I don't have timelines for you today. OK, fair enough. And then we also had, we had Boris Cherny from Ananthropic on stage earlier. A lot of you saw that. He talked about the fact that he's sort of managing on a daily basis like 80 or sometimes several 100 even 1000 AI agents. How are you working, Timothy, within Mistral? Yeah, so I would say we see. A lot of variability between uh in our developers into how many agents they manage. I think the big topic for us today is uh on the science side, really automating a lot of this. Um, I think the, the name that's been out there is uh recursive self-improvement. I, I think it's a fancy name for really applying AI to reduce the operational side of billing AI, which is a lot of looking at data, running experiments. All of this can be really automated and improved. Um, and then we use all of our platform vibe for all of our operations and running the company internally. So you're sort of, you think you're going to be close to recursive self-improvement as well. Uh, to me it's, I mean, the name is, is a bit grand, but yeah, there is a lot of improvements that we can make into how quickly we build these things. Can we at some point really just drive from afar? Yeah, I think that's, that's the goal. Cool. And bringing back to sort of the sovereignty question, I want to. Ask both of you, what do you sort of see as the role for what are being called AI middle powers, you know, countries that have some of this capability, they have maybe one or two really good model builders, they have some data center capacity, but maybe not as many colossal multi-gigawatt data centers as the US and China, and they maybe have some capacity around hardware, but again maybe not at the same scale. Danielle, what do you see as what role are these countries going to play going forward? You know, I built a company between the United States and Europe. I'm originally Portuguese. I studied in Portugal, did my PhD there. I've been in AI for 26 years. I'm a huge advocate of Europe as um the um. Uh, the grown-up adults in the room of the world, uh, setting up a lot of the values and a lot of the, and, and teaching the, teaching history to the world and, and for us to all learn with past mistakes of humanity. Uh, so I, I do think that, uh, uh, middle powers, as you're calling them, Europe, Middle East. Uh, especially Europe and Middle East, it's, it's really crucial that there's, uh, there's competition in the, in the world scene, um, and, and that, uh, because of all this geopolitical, uh, instability, uh, these countries have to rely on their own sovereign models, their own, uh, sovereign data, uh, for because we don't know right now what's Timothy, what, what's your view on this? Yeah, I guess in terms of data center and compute power, I think Europe is one of the strongest grids out there, so it's been, uh, it's been really nice not having to rely on on generators, but otherwise, yeah, it's, it is going to be part of the fabric of society and so as such, it's something that everyone needs to understand and partly control. Interesting. Other questions from the audience. Oh, we got one right here, Alexi. Uh, Alexey Oreskovic, Fortune, uh, curious what you think with all the talk right now of, uh, the US, uh, taking a stake in some AI companies and, uh, Quasi, uh, nationalization over there, how do you think that plays into the sovereignty debate and uh Or any justification for it. Um, I, I, I honestly don't know how, uh, what the future looks like in terms of how the power of the AI company grows, how it's concentrated or not. Um, I think it's a debate for the, it, it's something that needs to be in the public debate. Uh, ultimately, because as I mentioned, I believe AI is going to be very important for society and who owns, uh, that power, uh, is very important. At Mitral, we've always been, uh, in under the, we've always followed the strategy of folks need to own that. Um, but yeah, I, I think it's a crucial discussion to have, and you've, you have a partnership with the French government, right, on some of the data center stuff, um, not necessarily. I mean, it's, uh, they like us because we've been, uh, successful, but I, I'd like to think we've been, um, uh, partnering with the entire European, Middle East, Southeast Asia, and US ecosystem as well. Um, yeah, uh, I, I think for Europe, it's more about, uh, making sure there is stable demand that we can foresee and bank on rather than really government subsidies. Interesting. And Danielle, I want, I want to ask you on this question. Um, about data because obviously we're talking about private data, but there's quite a lot of public data out there and there's been talk about, well, if public data is being used to train a model, should there be some benefit back to the public that contributed that, that data in the first place? Absolutely, absolutely. Uh, the problem is both in the United States and Europe, it's, um, it, it's so hard to get alignment between. The agencies here, the countries in, and it's like a, it's a very deep federal division. It's really hard to get that bulk data in one place and with all the access that is that is required. All right, fantastic. That's unfortunately all the time we have today for our panel, but thank you so much for listening. You guys have been fantastic. Thank you so much.

Mistral AI

Mistral AI in Fortune Rankings

Mistral AI in Fortune Rankings

Videos