Do you have enough compute now?
We don't have enough compute now, but we have plenty of compute coming online in 2027, 2028.
Arthur, it is wonderful to be talking to you. I've followed Mistral for a long time, and I appreciate you being an early DB Live guest.
Yeah, thank you very much for hosting me, Dre.
So, you know, a lot of Americans still know Mistral as the French open-source startup, but you guys have really evolved. You're, you know, building frontier-level models, infrastructure, enterprise business. How would you describe the company today?
So we are building sovereign AI, as defined as powerful AI systems that we can bring to enterprises so that it gets deployed very close to their most critical data, their most critical processes. It brings them return because typically enterprises get return from artificial intelligence when it gets deployed close to what matters for them. And what that took us to do to build those sovereign capabilities that are important in the US, outside of the US, that are important for states but also for large enterprises, is that we're building three things. So the first is we're doing research, that's how we started. So we're building frontier-level models, we're pushing the open-source capabilities in general, and we're basically one of the best labs doing so, and pushing capabilities in cyber. Pretty happy with the performance of the upcoming open-source release. In that respect, we have started in 2024, and we're pretty well down that road to build an enterprise business where we are deploying our models or technology or agents and adapting them to the most critical workloads of our customers.
We can do it in a way where the technology is portable. So we bring it to the infrastructure of our customers or we host it on ours. And what we've built, because of some supply constraints that we were running into for customers in particular in Europe, is that we've built infrastructure. So we serve GPUs, we serve tokens on infrastructure that we physically own.
So we're fairly vertically integrated in that respect, in that we operate from metal, so deployment of servers. We typically rent data centers, and so we kind of stop there. We're not a real estate company, but we deploy servers and we go all the way to building the business applications for our customers.
Because in AI, that's what makes sense to do. You bring harnesses, you bring models that already know how to do a lot of things, and very quickly you can build business applications that have true return for customers. So that's what we do, and that participates in addressing those sovereign AI needs that our customers have.
Right. I'm sure you're aware that this has kind of been the year of open source in the US. You've seen just massive amounts of adoption first with developers and increasingly among enterprises. But there's a lot of talk mostly about the Chinese models, about the Kimis of the world, the DeepSeeks, the Quens, less about Mistral these days. What do you think Americans are missing? Should they be relying on Western open source models?
Well, first of all, there's been a lot of new talks and new traction on the model-specific aspect of Mistral because of Luch, whose open-source version is actually coming online next week. And so we're quite excited by the adoption rate that we see across the world, including in the US. In particular, the cyber capabilities of that model are state-of-the-art against the Chinese models as well. And so that is very attractive because if you're doing cyber defense as a company, you basically want to rely on open-source models because you absolutely don't want this thing to be turned off at the wrong moment. So that's creating a lot of usage on our side.
What's happening is that enterprises realize that AI is going to be about business continuity. As they replace some of their processes with AI systems, they want to make sure that they can change their providers, they can change the infrastructure it runs on, they can negotiate prices. And so that means that using open-source technology sitting on that is portable across multiple infrastructures is a great asset for them. They're also looking at control over other IP as something super important because when you have critical IP on industry and engineering, for instance, there is no way in which you bring it to companies that may want to intermediate you. So they want to build their own AI systems, and the best way to build their own AI system is to use and modify open-source technology. And then finally, cost is starting to be an issue in that all these things generate a lot of tokens, and so the sensitivity to cost means that as an enterprise decision maker, there is no reason why you should go and pay for an IP margin where you can go and access open-source models that are devoid of IP margin. And so just from a cost arbitrage perspective, it's bringing a lot of companies to use open-source models. Now, they've used a lot of Chinese open-source models because for a while they were the best. What we are seeing now is that now that we're the best in multiple domains, we're getting a lot of traction there. We're also pretty convinced that from a security point of view and a safety point of view, it's much better for enterprises to rely on Western open-source models, and we're very happy to be leading that front.
Right. And that message is certainly getting through, it feels like, especially over the last few months. But in the US, there's a lot of different companies that have sprung up to actually serve open-source models, as you've been talking about. You guys are sort of moving lower down the stack. You're serving them, you're renting, you know, capacity. Why the decision to do that? And, you know, how much have you had to take on from a capital point of view to be able to serve it also? And how has that changed the business model?
The reason why we started to invest there is that we realized how supply-constrained the world would become on compute. And so if you're in a situation where you're supply-constrained, you don't want to have too much intermediaries in between the one that owns the critical supply, which is in that case chip providers and Nvidia in particular, and the user. And so for us it made a lot of sense to start owning those GPUs and operating them so that we could secure supply and so that we could secure supply for our own customers. So in the end, it's a very simple trade decision that there's not enough supply. So you want to be as close as possible to your suppliers, and that means that it's much better to buy than to rent. That's how we decided. And we also realized that in Europe there was not enough investments. We could not find the right cloud partners. They were not strong enough, not able to operate GPUs at the size that we needed. And we had that competence internally because when you're training large models, you actually need to know how to run GPUs pretty efficiently. So it made a lot of sense for us to start growing our own fleet. And we're very happy to have done that today because now we can offer fully independent token generation. We can train our models on compute that we own, and we can scale the business much faster than we could have if we had relied on intermediaries.
Do you have enough compute now?
