Who Wins if AI Models Commoditize? — With Mistral CEO Arthur Mensch

I'm happy to be here. And thank you for hosting us.

Exactly. We're headquartered in Paris but we have around 40% of our workforce which is actually in the US and a lot of our activity is actually here, so that's why I'm spending a lot of time as well and that's why we are here in New York.

I would say that inherently this is a technology that is going to get commoditized. The reason for that is that it's actually not hard to build. You have around 10 labs in the world that know how to build that technology, that get access to similar data, that follow the same recipes and algorithms which are very short. Actually, the knowledge you need to train a model is fairly short. So because it's short, it actually circulates. So there's no IP differentiation gap that you can create, so it's very hard to actually leapfrog and to be way ahead of the competition because there's some diffusion of knowledge that is just making everybody do the same things. And so the question there is therefore where is the value accruing? And what kind of business model should you pursue to actually make sure that in the end you're turning profitable? And then the challenge that we see with some of our competitors is that they're investing billions or hundreds of billions into creating assets that are depreciating fairly fast because those are commodities. And so for us, it has always been at Mistral, it has always been a question, and one of the biggest questions of the industry, is that you need to invest enough to actually bring value to enterprises but you also need to invest reasonably so that you can build unit economics that makes sense in a world where the creation of a model, which is capital intensive, is actually just bringing you assets that are just in a commodity competition.

Well, I think the issue is, first of all, AGI is a very simple concept. So probably too simple for enterprises. There's no such thing as one system that is going to be solving all of the problems of the world.

It's never going to exist. I mean, you have a wealth of problems. Just like you don't have any human that is able to solve every task on the world, you of course need to have some amount of specialization to actually solve problems. And so we're back from magical thinking to system thinking. We need to figure out what is the data that is going to be used to make the model better at a specific task, what is the flywheel that we need to set so that we accrue more signal from humans interacting with the system so that eventually the application becomes better and better. And so in real life, enterprises are just complex systems and you can't solve that with a single abstraction which is AGI. And so AGI, to a large extent, is what we were not able to achieve and which is basically the north star of 'I'm just going to make the system better over time.' But because, as you said, it's hard to explain to investors that the technology you're building is never going to be matched by your competitors, then there's of course a shift in the narrative. Companies are not building a north star single system that is going to be solving all problems, but that we'll need to go into the weeds of enterprises and solving their actual problems. And I think at Mistral, we've been ahead of time in thinking about this. That's kind of what set our story. Our story has been to assume that eventually AI will be more decentralized, that more customization would be needed because we were running into the limits of the amount of data we could accrue and the limits of scaling laws. And because of that, we created the company on that premise, on the fact that we'll bring more customization ability to enterprises.

Yes. I think if you look at it from a system perspective, you have two components and we'll always have these two components. The first components are like static definitions of how the workflow should be and how a system should behave, and those static definitions are set by humans that are defining how the system should behave. And so this corresponds to the manual information that you're using to define the system. And then there's a dynamic component where you're connecting a model to tools and you're giving instruction to the model and the model can go and call the tools itself. And so it can decide on the graph of execution that it's going to follow. And so that part is dynamic. And there's a static part where you're setting up guardrails or you're deciding you have a tree of decision sometimes. And I think it's a bit utopian and unrealistic to think that you can solve everything with a dynamic system without guidance from humans. And what has happened in the industry the last three years is that effectively the dynamic part has grown because models can think for longer, because they can call multiple tools, because they can code. But the static part remains extremely important. And even if the dynamic part grows, then the static part allows you to create systems that are even better and more interesting and you can solve problems that you were not able to solve before. So the combination of these static systems, which you can call orchestration if you want, and the dynamic systems that you can call agents, is going to stay super important because the two things are moving up together so that we can tackle problems that are more and more complex.

Well, yes. I think on the consumer side, because AI is starting to be, well, is becoming the way you access information, you basically have an ads business to be built. And that's pretty clearly going to be built. It's not the focus of our company. And then if you look at the enterprise side, we're basically replatforming all enterprise software. So enterprise is about having the right... In enterprises, you have people, you have data, and then you have processes. Historically, there was a fragmentation of the tools to run multiple processes, multiple data systems, multiple systems of record. And there was a fragmentation in teams that were not able to access all information at the same time. And essentially, what AI allows you to do in an enterprise is to start with a unified data, or even you can start with fragmented data sources because the AI is able to navigate them. Then you put an AI on top that is building the right amount of intelligence, understanding what's going on in the enterprise, and then the AI system is able to somewhat generate the interfaces that is useful for every human to actually work. And so that part, that replatforming of the entire enterprise software stack, is the one thing where a lot of value can be created in the enterprise: owning the context engine, the system that is constantly running that is looking at what's happening and figuring out, creating documentation for what's happening, owning the front end as well that are more and more getting generated on demand. So let's say I'm a lawyer, I want to fix one very specific review to make. I just bring my document and then the system actually evolves in showing me the right widgets and showing me the right information I need. So the generative interfaces on top of a context engine that is constantly updating its representation of what's happening in the enterprise, on top of systems of record that are essentially going to be just pure databases. You don't need everything that was sitting on top before. This is where this is going, and that replatforming is going to take a decade because it takes a while to get enterprises to adopt these things, but there's just immense value to be created because suddenly you can reorganize your company around the fact that for many of the processes where you had a lot of people, you can actually run those very much faster. That's on one side, efficiency. And the other thing, which is the most... So that's one of the business modalities in the enterprise. The second one in the enterprise is about working with enterprises to help them take their really proprietary data, the assets being produced by their machines if it's in the manufacturing industry for instance, and turning that into intelligence that nobody else can reproduce. And so making models specifically good at a certain kind of physics when we're working with a company doing planes for instance, or when we're working with ASML making models that are specifically good at operating their machines. That's huge value because suddenly you're not building efficiency within the company but you're effectively unlocking technological progress that was locked by the absence of AI. So that unlock that the new systems are providing, that's immense amount of growth. It's actually harder to measure because the first one is shorter term. You can look at what the company will look like in five years because you've reduced certain parts of the company, you've reoriented other people to be creating growth. That you can create models of that. On the technological side, I think it's a little harder because we know there are things like nuclear fusion or sharper engraving of semiconductors for instance. These are things where we are starting to run into physical constraints and artificial intelligence can actually help to lift those physical constraints. And so the acceleration of technological progress is, I think, where most of the value creation will be. It will take a little bit of time and it will be less measurable and less predictable than the efficiency gains that AI is going to produce. But the two things are as important.

Yeah, that's kind of the fork that we see in the industry. And our view there has been to be on the second one, to really...

Which brings customization, but it also brings decentralization in that if you assume that the entire economy is going to run on AI systems, well, enterprises will just want to make sure that nobody can turn off their systems. So the same way if you have a factory you connect it to the grid, you want to make sure that nobody's going to turn off the grid because they don't like you. If AI effectively becomes a commodity, which is what's happening, and if you treat intelligence as electricity, then you just want to make sure that your access to intelligence cannot be throttled. And so that's also one of the things that open source technology can bring.

Yeah, you run it on your own terms. You create the redundancy you need. You can serve with higher quality of service. You can make sure that whatever the geopolitical situation may be, you can still run the systems if you want. And so that's really on the IT side. So if I'm a CIO, I really look at open source as a way to create leverage and independence. But on more on the scientific side, it's also the only way in which you can create systems that are effectively using the folklore knowledge of your employees. That's the knowledge that you've recruited for decades. The only way to turn it into an asset that nobody gets access to is to create your own models based on those open source models. And so that's... but it's hard. It's hard to actually build those, right? And so that's where you need the right tools. You need the right expertise. And that's the complement business model to building open source models.

They will say that, but then they will put some guardrails on top of it. So you're basically trusting that their engineers are going to give you enough access to the depth of the system. And can you trust that for eternity? I'm not sure. So the issue there is as much a question of control as a question of customization, right? A vendor is going to try to lock you in. So if you get access and if you build on top of open source models like our open source models or anyone's, you're basically less locked into the vendor. And this is a technology which is so important that you don't want to be locked into a single vendor. So that's also the opportunity we bring.

Yes, this is true. I don't think it's the first time that we observe it in history. It's a new technology. It's a new platform. And so the knowledge on how to use it is actually still pretty scarce. So there aren't that many people that can build systems that are performing at scale, that can run at scale reliably, that can actually solve an actual issue. And so when working with enterprises, you always need to have some services on top because of the complexity of implementation, even with fairly well understood technology like databases. But for artificial intelligence, it's even more necessary in that it requires to transform businesses. So you need to also help in thinking how the team should perform around the system itself. And it does require to customize things. So you need data scientists that know how to leverage data and turn it into intelligence, and today this is still a pretty scarce resource. I would say I do expect the part of the software in those deployments to increase. So the amount of the way customization occurs today with fine-tuning, reinforcement learning, these kind of things, this is going to be abstracted away from the enterprise buyer because it's too complex and they actually should just worry about having adaptive systems that are learning from experience and from deployment with people instead of thinking about 'should I use fine-tuning or should I use reinforcement learning' to actually put that knowledge into my models. And the work that we are doing is to try and abstract away from lower level routines that data scientists understand to higher level systems that business owners can actually use. And so it's going to occur and we're working on it. But the service part is still going to be quite important. And today the combination of the two things is the fastest way to value if you're an enterprise. So we've been combining the two.

I mean, we are there to help our customers get to value.

We're here to... but to get to value, they need to have great models, and to get to value, they need to have the right tools to train the models. And so the best way to create those tools is effectively to train the best models. So the two things are extremely linked together. We create models that are very easy to customize. We create models with tools that we then export to our customers so that they can use them, and we help our customers train their own models. So you can't go and sell to an enterprise that you're going to help them create very custom systems if you can't show to the world that you're effectively the leader in open source technology. And so the two parts are equally important. The first is enabling the other, and there's effectively a flywheel there because we make our choices when it comes to the model design in a way that is enabling the various customers we have. So one example is that we've put a lot of emphasis on having models that are great at physics because we work with manufacturing companies that run into physical problems. So that's the flywheel that we have set up by having the science team and the business team actually sit together.

Well, if you look at the trends in 2024, I'd say there might have been like a six months gap. If you look at the trend in 2025, I think the gap is more around three months. So I guess it's up to anyone to guess what the gap is going to be next year. But effectively, this gap has been shrinking quite significantly. The reason for that is that basically you have a saturation effect when you pre-train models around 10 to the 26th flops. The reason for that is that there's only that much data you can find to compress when you pre-train models. And so effectively, labs that maybe started a little behind created enough compute capacity to train models at this kind of scale, and efficiency has also increased. And so what it means is that today everybody has access to 10 to the 26th flops facilities over the course of a few months.

That's a measure of compute. And well, that's a measure of compute times time. So you need to... 10 to the 26th flops is something that any lab today can achieve in a couple of months. And because of that, the saturation effect means that open source models have caught up because closed source...

The gap is shrinking. And I think it's going to continue to shrink. And so the question is not whether open source will catch up, but when. And I think we're already there in many respects.

Models that were started ahead kind of run into that wall of pre-training. So what that means is that this is only going to continue shrinking. And if you look at the latest open release we did, which is Death Desk 2, a coding model, it's performing I think around the performance of Anthropic around two or three months ago. So yeah, I think the gap is shrinking. And again, I think the question is probably not posed in the right way because it's also offering two very different distinct value propositions. On one side, this is well managed and you will depend on the provider itself. On the other side, it takes a little more effort because you will need to own it more, learn about how to customize it, use the right tools for doing so, and maintain its deployment if you choose to deploy it on your own facilities. But at the end, this is creating the leverage you need against closed source providers. So the two categories are effectively different, but if you look at the pure performance side, they are definitely converging.

They will get better in more and more specific domains. I think we've really collectively made them very clever and able to reason about long context and able to call multiple tools, etc. But if you want to effectively put them into production in a bank or in a manufacturing company, the models need to learn about all of the knowledge that is contained within the companies themselves. So what it effectively means is that for very precise directions, let's say I want to make my model extremely good at discovering materials or extremely good at designing planes, I will need to go and sweat it a little bit, get the right reward signal, get the right experts, and ask them to make my model specifically good in that very precise direction. So we are definitely not done doing that, because what we are all racing for is the right environment and the right signal provider for specific capabilities. But the broad horizontal reasoning capabilities we're still going to improve them, but nobody is going to improve them in a way that creates a strong gap versus its competitors. So the strong gap is actually in working with vertical experts that know exactly how they design a plane and that actually explain to the model how to do it. You have a wealth of directions that you can take because you can do it in physics, chemistry, pharmaceutical, biology. So to me, the most exciting part of what's going to happen in the next two years is that explosion of very precise directions in which the models are going to get better. For us, the opportunity is to have the right platform for enabling those kinds of verticalization, whether with enterprises or AI startups that are working on very verticalized capabilities. So that's my view of where the field is going. We have been about horizontal intelligence growing and things getting more and more clever, and the next two years is going to be about taking models and making them extremely good at a certain skill set. That's actually more exciting because we're getting to a point where you pick a domain and can just make it superhuman, but we're not going to make it superhuman in every domain at the same time.