Thank you so much. I know we're standing in your way of lunch, so we'll make it fun and interesting and short. Really appreciate you taking the time here. We're just chatting a little bit about you're hiring over here in Silicon Valley. Tell us what you're looking for, who you're chatting with here.
So at Mistral we have a fairly big team actually in Palo Alto historically. That's where we put our US headquarters. So we have a lot of researchers, also have a pretty big go-to-market team, and have plenty of customers using our compute, in particular working more and more with cyber vendors as well to deploy some of the models we've been doing. We are very proud of the open, of the cyber capabilities of our latest open source release, and so that's a big part of the conversations we're having at the moment.
One thing you've said about OpenAI and Anthropic and the frontier labs taking data from their customers, using it to create competing products, and that you guys are different and you're not doing that. What do you say to that argument that they deny that they're doing that?
Well, they may deny that they're doing that. There is some track record of them doing that a little bit. The true thing is when we work with some of our customers in the manufacturing world, in the financial services world, is that they are not going to trust us when we tell them that we're not going to train on their data. We prove that we cannot see their data, and we do it in two ways. We do it either in deploying our technology on their premises, so in a way where structurally the data is not flowing back to us, or we do it with more confidential guarantees that we give on our own cloud offering. So those two things mean that they don't need to trust us because we have basically a zero-trust agreement with them. And I think structurally you need as a customer to ask for these kind of things because even if it's not on purpose, you can see your vendor actually be the target of a cyber attack, and then that's how you actually lose IP, and that's how you may have a lot of problems. So the guarantees we give to our customers means that structurally their IP and the processes they want us to help run are actually going to run even if we are to disappear, and they know that the data they have is in no way going back all the way to our systems.
But isn't it in Anthropic and OpenAI's best interest to not mess with their customers? I mean, they want to keep their customers, keep them happy.
At the end, what you see is the business of artificial intelligence is either you behave as a third-party provider and you provide an horizontal technology to your customers and you help them build custom systems that are leveraging their assets, or you go vertical. You go vertical after vertical and you try to actually capture as much value as possible on the value chain. So I think the two strategies are valid. Our strategy is to provide the horizontal open-weight platform so that our customers can actually build their own system. So the manufacturing customers we have, they are using our technology to build planes that are flying better when it's Airbus. They are using our technology to build better lithography machines when it comes to ASML, to build better cars with BMW. And we are doing that in a way where we bring all of the tools for running industrial agent, industrial engineering loops, so agents that are designing new objects, that are running physical simulations and everything you need to build better products. But we're doing that in a way where structurally our customers know that their data is not flowing back to us. And so they know that we're not going to try and intermediate them. They know that we're not going to use the know-how or the data that we are helping them process into building a competitive product.
But you guys do hide some of your weights, right?
We have some proprietary assets, in particular the specialized models that we have. Again, we can deploy them on the premises of our customers. So we're not exposing them for free, but we can help our customers deploy it where they need it to be. So everything that we have is portable by design because when it comes to enabling the running critical processes, the business continuity requirements of our customers means that they want portability and they want to be able to move from one platform to another in terms of their infrastructure. And so the commitment we have is to provide them with the weights wherever they need the weights. Then we have two categories of models: some of them that are open source and for free, and some of them that are more proprietary and usually are the specialized ones.
So you're sort of making that your, you know, differentiating factor, but you're not as big as Anthropic or OpenAI. You're not as cheap as the Chinese AI labs. You know, is there a need for Mistral? Like, where do you see yourself in that competitive set?
The category we're defining is what we call sovereign AI. So technology that is portable, technology where you have business continuity guarantees, technology where you have affordability guarantees. We're actually quite cheap. The models that we serve are quite cheap. We serve our models, but we also serve any kind of third-party open source models. And then the guarantee we give is again a guarantee on IP, a guarantee of building meaningful use cases on the most critical processes of our customers. So the way we work is very different from some of our counterparts in that we go much deeper. We go to the critical processes of our customers. We go very close to the data they care the most about, and that sets a lot of constraints, and we're pretty uniquely positioned to address those constraints. So we define that category, and typically our customers would work with public AI in certain domains, but on the domains they really care about, which are R&D and manufacturing for instance, they will work with us because they want to have a partner that understands the business and can build custom applications. They want to have a partner that can give them guarantees that whatever happens it's still going to be running. That gives them guarantees that the data is not flowing back to us.
Are you finding customers who particularly come to you because they don't want to align themselves with the US or China? They like the fact that you're in Europe.
Yeah, there's a second factor for our growth. So again, the first one is really differentiation on the technological level, which means that this has been a global play on certain verticals. We have the right go-to-market motion. We actually understand the businesses of our customers quite well. So we know how to set up agentic engineering loops for instance. And that's super important for anyone that has a physical footprint in the world. But there is an amount of geopolitical fragmentation, and structurally the AI supply chain is fragmented. You have multiple, it's fragmented and coupled. But when it comes to the services, what our customers are realizing is that if they're in Europe, it's very possible that they won't get access to some of the US models. So they need a European supplier. It's very possible that they will not get access to the compute they need. So they actually need compute that is sitting in Europe and operated by European players. It's the case in Asia as well. Everyone is hedging their dependencies and actually sourcing some of the technology from multiple parts of the world. And we are very strongly hedged against dependencies that may be excessive to US providers in Europe, to US providers in Asia. When we're operating in the US, it's absolutely not about hedging. It's about the technological differentiation.