The next guest is someone who I got to meet a few years ago at the Paris AI Summit and someone I've really wanted to interview on stage for a long time because I think he brings a really fresh and interesting perspective to the AI conversation. Please help me welcome CEO Arthur Mensch.
Set that down. I want to record all the insights that I'm going to get for you using AI. Although as I said, we shouldn't be talking about when we're using AI. We should just be doing it. So the last time we had a chance to really talk in public was at that AI summit. How have things been? You know, Mistral obviously for those who don't know kind of you know one of the leading non-US or Chinese AI companies, one of the leading AI companies period I think you would say. How has Mistral grown in the last couple years?
Compared to February last year, we did 20 times. That's been quite a high growth and so it's been a pretty busy year in terms of business, in terms of growing the team as well, in terms of technology as we've really pushed the capabilities of the open source models that we've released but also built the platform on top to deliver value to the enterprise which is really the business where we are investing. And so it's been a busy year. It's going to be a busier year even in 2026. Everything that we've said from the beginning of Mistral which is that customization was important, which is that control and strategic autonomy on the topic of AI was important for multiple reasons, turned out to be true and to be true faster than what we even expected. So that's been quite interesting.
Well, I want to talk about finance and some other topics later on, but I want to dive in on open source because I think when some people talk about open versus closed, it's more of a religious debate than a business debate. But my guess is that that isn't how your customers think. Do you feel like they're coming to you because you're open source or they're coming to you and the fact that you're open source lets them do what they want to do?
I mean they come to us because we know how to get them to the outcomes they need and usually that entails customizing the systems and that entails customizing all the way to the models. And if you have a policy where you assume that the models are not where the value occurs but that the models are essentially free, but that it matters to actually train them and it matters to be the one who certified that these have been trained on good data etc., then it makes a lot of sense for us to open source and it makes a lot of sense for our customers to rely on open source systems. And so that ability to customize the systems deeply and that ability to deploy it wherever they want it to be deployed. So whether on prem, on private cloud, and on Mistral Compute which is our infrastructure endeavor, or whether on the edge which is actually a very fast growing business for us, edge and robotics, well open source matters because you basically want some infrastructure technology that you're going to build upon and all infrastructure technology and software have historically been open source. And so what we've seen from a competitive perspective in the last year is that we took the leadership in releasing open source models in the West, meaning Europe and the US. We've been mostly head-to-head with Chinese companies that are releasing extremely strong models and effectively that competition has led us to be around two three months away from the closed source model. So it's a very interesting dynamic where what we've essentially been saying from the very beginning which is that the model layer is a free layer is turning out to be true and that what matters is customization.
Given that, why use your own models? Why train your own models? Why not let someone else do the hard work if you're not going to get the value from that?
Well, we get the value from knowing how to train models because when we work with customers to do deep verticalization in specialist domains, when we connect ourselves to the proprietary modalities whether images or time series or whether text based data which is fully different from what we can observe in our training set, our ability to train at scale is extremely critical. So that's a key aspect. So owning the training stack is super important. Now the other thing which is important and that relates to the competition that we see in between us and I would say the Chinese models is that a model is essentially compression of knowledge and you can put anything in it. So in particular when you start using it for coding, that's something you want to make sure that they are not going to introduce some flaws or some security problems. And if you don't work with a trusted partner that is the one that is producing the open source models, then it's a little more difficult because you're essentially since you don't have access to the training data, you don't exactly know how the things have been produced, you need to trust the one that has produced it. And that's where it is important for our customers that we're able to train our models.
And I want to talk about the enterprise because I assume that's still most of your business is enterprises.
This is 95% of our business.
From your work with different businesses of all sizes, where is AI delivering the most value today for the companies that are succeeding, that are moving from pilots to production? Obviously we know about coding, we know about customer support and a few of the areas. Where else are the companies finding success and what helps them succeed?
So beyond those two things which are pretty well established, I would say on the customer care side of things it's still a little early. I would say enterprises will have a tendency to say that they're adopting at fast etc. because that but if you go internally there's a little more friction than that. But where we see success is when companies go with a certain function to fully automate and to fully reorganize around AI processes and AI management. And that's especially true in supply chain. In supply chain you need to deal with multiple actors, multiple legacy software and you need to run long and durable processes. And so for those if you put the right effort and if you combine AI generated code with LLMs that can interact with dynamic input, you can actually get to 80% improvement in dispatching containers, you can get to large improvements in ensuring that like in managing your bills of materials. So that's a big area related to we work a lot with manufacturing companies. So that comes up quite often and it's very core to their business. It's core to their operations. So yeah, supply chain management, very very big. And then what is exciting but again quite early but we see a lot of signs that this is going to accelerate things is the fact that you know like research companies that do a lot of research and development have engineering problems, their biggest source of acceleration is reducing their development cycles and that typically takes more than software engineering because you need to know the physics of the systems you're building, you need to embed AI into the systems themselves. And when we work with ASML for instance which is a very engineering heavy company but not a very software engineering heavy company, we are bringing them leverage on things that are mechanical engineering, physical engineering and that is where again this is early but that's where we see a lot of opportunities. So AI accelerating technological progress and what's interesting there is that you do need to go into customization of the models themselves because you're lacking the vertical expertise and the data to we are lacking it to do it ourselves and so we do it with partnerships.
That makes sense. The other thing, you know, a lot of companies have said they're getting stuck, they're having trouble getting the ROI they thought they could as they move from pilots to production. When you hear that, when you see that, where is it that companies that are getting stuck, what tends to make them get stuck?
I think the issue is they don't think and that's what we try and solve in the first place when you go into those enterprises. The issue is they don't think about building AI applications as something that have a life cycle where you build something, it's not working, it's only working 80% of the time, then you need to iron out the details. You're building a workflow automation, the problem is not the 80% happy path automation that will work. It's the 20% edge cases that will go and require human to intervene. And so that part is actually where it's hard because you need to have an iterative process. You need to have a data scientist mindset where you look at evaluation which again means that you need to observe everything and then you figure out what kind of data do you need to make sure that the next time you see this edge case this is not going back to a human. So that approach which also entails bringing beta deployments that are not fully accurate to the people. And so that also means that you need the right human and machine interactions. That's where things get stuck because the prototypes look good. You've seen it. You've shown it to the comex, they're excited but then it doesn't work. And you need to go to the 20% that remains and that's where you need to iterate on the systems. You need to connect to more data. You need to eventually to actually modify the weights.
