Mistral CEO：OpenAI 与 Anthropic 为何将进军你的业务？| 中国模型真的安全吗？

Arthur, man, I’m so happy about this. I’ve interviewed some of our favorite colleagues in the U.S. You’re our last hope in Europe. No pressure, but thanks so much for joining me today, dude.

1. Why Europe Needs Its Own Frontier AI Companies

Thank you, Harry, for inviting me today.

2. Why Mistral Avoided the Consumer AI Market

Arthur, I’ve done some research, and I’m expressing quite biased opinions. But it seems I’m not the only one who thinks this way, because you said the other day that leaving AI to the U.S. and China would cost Europe what’s left of its place in the world.

I’ve said that, and I truly believe this is such an important technology that it needs to be spread around the world. We’re talking about a technology comparable to electricity, a technology that is increasingly subject to export controls. AI will become a major driver of economic growth, and if AI isn’t produced in Europe, it will be difficult for Europe to achieve the growth it needs.

To what extent does Mistral’s success depend on excessive U.S. influence?

Yes, obviously, when we have a Trump administration, we see things like Fable being banned, and that raises a lot of questions and the need for sovereignty.

In a world without excessive U.S. influence, does this threaten sovereign suppliers?

3. Why Sovereign AI Is an Economic Necessity, Not Just Politics

I would say it’s a structural problem. It’s not really related to any particular way of functioning or demonstrating power. This has more to do with the nature of the technology itself. If you’re starting as a business or as a state to use AI in all your core processes, that means you want to make sure the money you spend on those processes is actually accumulating quite close to you. If you don’t, it means you have a trade balance problem.

This means you have huge flows of money moving from one place in the world—it could be Europe—to another place in the world, which could be the United States. And therefore, there’s a macroeconomic problem: the impossibility of creating artificial intelligence and conducting research and development on our own territory. This isn’t about saying that Europe is lagging behind, or about using soft power, for example. It’s structural.

I won’t say that some people don’t get a little emotional about it, but when we launched Mistral in 2023, we started with the belief that it was a sovereign technology. As a company or a state, wherever you are—whether in the U.S., Europe, or Asia—you need to own technology. You need to be able to configure systems and run them close to your data. You need to control your spending and work in true partnerships with companies that understand how to build artificial intelligence systems.

4. What “Sovereign AI” Actually Means for Enterprises

Sovereignty predates the 2025 elections. What we said in 2023 turned out to be true. It got quite emotional when we started talking about Greenland, but it’s mostly an economic issue. This isn’t exactly a political problem. It’s just too important a technology.

When I hear you say that, I think of Lena Fireworks, who came on the show and said something very similar: companies need their own models, trained very close to their own data. How do you feel when you hear me say that? Is this an evolution of the company that allows companies to own their own models and work with their own data?

5. Why Critical Industries Are Reluctant to Use Chinese Models

That’s the category we define, and the reason why companies work with us: we offer them sovereign AI technologies. Sovereign AI technology requires 3 things. You need to have a protected model source, so you need a company that can train the models itself. We have this model factory. Mistral, the big four, actually leads in open modeling. In decoding agents, it leads in cyber defense and cybersecurity.

That aspect is about being able to continually improve the models and tell our customers that in 2 years they’ll have better models because we’re constantly training them and spending billions on it. This is extremely important for anything sovereign. This is what Mistral was created for, and it’s the main area where we invest money.

You also need control over your data. Generally, you don’t want that data going back to AI vendors. So we made sure at a technology level that our agent platform and model-tuning platform actually reside on our clients’ infrastructure. This is extremely important in many cases. When we work with manufacturing companies, they don’t want their IP transmitted back to the subprocessor.

In general, that’s not enough, because AI consumes a lot of energy. When you go to large-scale technologies, you need a lot of GPUs, which are difficult for an enterprise to acquire. You also need token-generation services and, essentially, token factories. You need a combination of data stored in your infrastructure and tokens produced in a managed infrastructure at a lower cost.

So we, as a company, do 3 things. We train the models, deploy the platform and data plane at our customers or in their cloud, and then we provide the compute. That’s why we build our own infrastructure. Those are the 3 ingredients needed to serve sovereign AI workloads.

6. Are Chinese Open-Weight Models Really Safe?

Can I ask about one thing? You said the model is protected. I’ve had the best experts in the world on the show, Arthur. I’ve been very lucky, but they tell me different things. Some say, “Chinese open-source models—you download them and run them locally. It’s perfectly safe.” And I’m like, “Wow, that’s great.” Then the next week someone says, “We don’t know if there are prompt injections you don’t know about, or backdoors. Absolutely not. You can’t be sure they’re safe.” Help me here, Arthur.

At Mistral, one of our responsibilities is to do inference for open models. We launch our own, but we also launch others. Why do we launch others? Because customers want choice, and when you have the infrastructure, you want to provide as much choice as possible to saturate it. That’s why we started serving more and more open models. In different areas, you may want to choose a different model. You may want smaller ones or larger ones.

The more choice you have, the better you can utilize the infrastructure, and the lower your costs can be. In a sense, this is a commodity market, so you want the cost to be as low as possible. That’s why we serve these models. But what you say is true: models are a compression of knowledge. If you don’t control the data supply chain, if you don’t know what goes into the models, you can’t guarantee anything about the outcome.

When we serve our own models, we can provide our customers with more guarantees than when we serve third-party models, precisely because, theoretically, you can verify that in a certain environment, with a certain context, the model behaves in a certain way. There’s no way to rule this out. There’s no way to measure this. We can track it online; that’s what we do. But you can’t rule out that it’s being used as an attack vector.

I think it’s very theoretical, because I don’t think that’s the case today. But on the other hand, it’s part of the threat model. That’s why our customers ask us to provide them with all the open-source models available. We help them do this in a safe way, so they can control what happens. But in many cases, they really want to use our models because we can tell them that the data supply chain, everything we’ve trained them on, is something we’ve already tested.

What percentage of token traffic, approximately, comes from your own models compared with other open models?

I would say about 80% comes from our own models.

Wow. Okay, that’s much more than I thought. As for the other 20%, everyone tells me, “Oh my God, American CIOs and IT directors will never use Chinese open-source models.” And in fact, this fear really exists. You have some great clients. Are they afraid to use Chinese open-source models?

It depends on how important their industry is to governments in general. If you work with defense, for example—we work with defense customers in Europe—they will never use Chinese technology. If you work in the financial sector, again, because it’s critical infrastructure, they’re very reluctant to do that. If you work in retail or in certain areas of manufacturing, it depends on where you are; the reluctance varies.

I think there’s a common understanding that open-source models, if deployed on a sovereign infrastructure like ours, present a threat only in the sense that they could go rogue and exhibit strange behavior in certain areas because they could be attacked through the supply chain. There’s a common understanding that this is very theoretical, and that’s definitely not the case today.

But it’s part of the conversation: “Is deploying third-party models that I don’t fully understand a threat?” Part of the conversation is, “Will I be able to get updates in 6 months?” What’s happening is that the models have become so powerful that they’re capable of carrying out cyberattacks and can be used for quite advanced purposes.

And so the consequence is that they are increasingly seen as defense systems. In fact, the US views them as a defense system because, at some point in May, it began banning their importation outside the US. If you look at what China is doing today, it is also looking for ways to restrict exports above a certain threshold. The reason is that this is effectively a dual-use technology, and with any dual-use technology, security of supply is even more important.

The companies we work with understand that they need security of supply not only on the infrastructure side, but also on the intellectual property side of the models.

7. Why OpenAI and Anthropic’s Closed Models Create Too Much Centralized Power

Can I ask you, given that, do you think OpenAI and Anthropic hiding their best models is good or bad for society? Is this just an inherent trait that will be common in this industry? Is this good or bad for society?

I don’t know. What is clear is that there are huge forces toward centralizing AI providers in a world where there are no providers that allow centralized deployment, i.e. sovereign deployment. What we do is extremely important for companies. If we end up in a world where all companies have to rely on two or three centralized AI providers that give them enormous leverage, those providers will use the data generated by the systems to build better systems and meet their customers’ needs. After all, that is the ultimate goal.

This is quite clearly reflected in the price today. Our approach is to make sure that we don’t live in a world where there is an oligopoly of AI vendors using all the data generated by enterprises to build better systems and act as intermediaries for those enterprises. Our clients very much agree; they certainly don’t want to live in such a world.

I would say that the fact that there are internal models being used to conduct certain research, for example, is the earliest sign of what can happen in a world where there is a huge concentration of power in artificial intelligence. Fortunately, creating models is not that difficult. It’s about capital allocation and having the right team teaching the right things. And now we’ve been able to do it.

Open-weight models are catching up with closed models. In many areas, they actually exceed them. That means I don’t think we’re going to end up in a world where, if you want to do math, you have to work at OpenAI. Fortunately, there are enough people in the world who don’t want that to happen.

You’ve mentioned energy several times. One of the key principles for success is having the energy and infrastructure to power European AI. We have very different energy systems from the US, and very different energy systems from China. In China, they’re going to relocate 3 million people to create energy infrastructure, but in the UK we’re not moving a birdhouse. That’s pretty disturbing, despite the joke. What do you think about the energy supply problem?

8. Why Energy Is Not the Biggest AI Bottleneck Today

Well, I think the first thing to realize is that energy is about 10% of the total cost of owning a computing system. So it’s not where you can make the biggest optimization. The cost of energy is ultimately not the only factor that matters in the economics of a project. What matters is the cost of chips and the cost of capital. This dominates the cost of materials.

It will become an increasingly big problem as we deploy more computing systems in Europe, because we will face supply constraints. Again, this is unevenly distributed. If you look at France, for example, there is a surplus of energy, so you can increase capacity. The same is true in the Scandinavian countries. In the UK, this is a little less relevant. In Germany, it’s a little less relevant.

But there is a realization that artificial intelligence will stimulate energy consumption. I would say that over the next 2 to 3 years, the real struggle, the real challenge, is to ensure that Mistral is built as quickly as possible, using the surplus energy that is already there. That’s what we’re working on: raising capital, building more, and building a team to create all of these capabilities.

There is excess energy. This is not a bottleneck today. It will become a bottleneck in 2029 or 2030. But now the question is: how can we use the energy that already exists? We need to think ahead. Starting in 2029 or 2030, as AI capacity of several gigawatts is deployed each year, we need to increase production. We hope that, for example, France’s huge EPR program will produce much more nuclear energy. Great efforts are also being made in solar energy.

I would say it’s going to be a problem mostly of coordination, because ultimately, if you have demand for energy from artificial intelligence, the business case for building new nuclear power plants is pretty well-founded.

Absolutely, and I agree with you. I deeply believe in Adam Smith’s invisible hand, but the deadlines are quite real. So 2029 or 2030 is a challenge. Nuclear power plants take longer to build. Solar power plants might be faster. I believe in this invisible hand, but states should play a role in facilitating construction and planning ahead. They have a longer-term vision than businesses in general.

9. What Mistral Learned About Buying and Forecasting Compute

So we spend some time explaining that energy demand is going to grow and that this issue needs to be addressed somehow. This should be part of any energy supplier’s business plan.

I just want to ask you about the importance of chip supply. What do you know now about forecasting demand for computing chips that you didn’t know before?

This is a rather complex process, so we learned how to run it to match supply and demand. It’s a complex industrial process that has nothing to do with software. You need to understand supply chain latency. The supply chain is quite complex because you have GPUs, but you also have CPUs, storage, and electrical components. You need to make sure that your own suppliers don’t have problems with their suppliers.

We studied this aspect, but it’s pretty complex and requires managing resources. You need to have several predictions, and you need to be able to forecast your demand across multiple periods. This applies to enterprises that are major providers of in-house products (ICPs), as well as those that use artificial intelligence and can use a larger number of GPUs but want to use them for a short period of time.

There are companies like Microsoft that are taking capacity away from us that we could use to fund more capacity. Ultimately, the problem with building a cloud is one of supply and demand. It’s about upfront supply chain management, ensuring transparency from your suppliers, and balancing all of this with the finances needed to fund the project. You need people who are good at financing, so it’s actually quite interesting to build a business like this.

10. Why Europe Still Does Not Have Enough Compute

We learned that, ultimately, if you’re able to stay ahead of demand, you end up with something that is fairly easy to sell. And it will be like this for the next few years. We’re very grateful that we anticipated we would run out of stock pretty quickly, and that our customers would have supply chain issues with their own GPUs, because now that we have these GPUs, we can meet their needs. We are completely suitable.

It was one thing, namely the cost of the chips. The cost of capital was another. If you had completely unlimited resources, what would you do today that you aren’t doing?

We would build much more capacity. I think the main bottleneck in building capacity, of course, is capital. We see the demand, and we know that our customers are running out of computing resources. When we take some of their use cases and help them scale them, they suddenly realize that they don’t need 8 GPUs; they need 1,000 GPUs. And that’s for one use case.

We see that this demand is outpacing the demand of our customers. We have a pretty high conviction, so we bring equity into capacity building. What we’re working on, what we’ve always worked on, is providing more managed capacity, even by anticipating demand. Today, if we just look at what Europe needs, it’s quite obvious that there isn’t enough computing power. You could argue that the US is perhaps building a little too much computing power, but it’s clear that Europe has plenty of resources to build capacity in a way that you can be 100% sure will be distributed. That’s what we’re working on.

11. What European Governments Need to Do to Accelerate AI Infrastructure

But you said there’s also a role for the state to play. What would you like a state, whether it’s France or Great Britain, to do that it isn’t doing today?

We’ve been pretty vocal about ways they can intervene in a market-neutral way that would simply encourage investment. The first, of course, is to facilitate the issuance of permits and ensure the availability of specially designated places where everything happens quickly, and where all stakeholders, both at the regional and state levels, coordinate their efforts.

I would say that France and Sweden, for example, are doing a great job with this. So it’s mostly about facilitating stakeholder alignment, because these are physical assets. They are built within regional communities, so it’s extremely important to make sure that everyone benefits from them. That’s where state aid comes in.

What they can do financially, and we’ve said this publicly, is provide government guarantees on data center leases—for example, reserving space for 10 or 15 years. That is extremely beneficial for companies like us that have a relatively small balance sheet and know how to build, and for whom...

That’s the role governments can play: providing some guarantees and ensuring further construction happens on time. Then, of course, there’s government procurement. The public sector is an important client for us in many different countries in Europe and beyond. We bring value to them, and they bring value to citizens. It creates sales for companies like us and allows us to build further. If you build further and faster, you build more cheaply in these times of limited supply.

Can I ask you, going back to this element of sovereignty, about something you’ve said to governments? How deep should sovereignty go? By that I mean, we obviously have a cohesive Canada. Does each country need its own model? Does this require blocs? How should we think about the granularity of sovereignty?

12. Does Every Country Need Its Own Sovereign AI Model?

I think the right level of granularity is a market large enough to build relevant models. Today, Mistral can create appropriate models because we don’t work at the national level. We’re a French company, but we serve Europe, the US, and Asia. We need alternatives, but there won’t be a single sovereign AI provider and model trainer at the state level. That won’t happen.

What percentage of your customers are European?

If we look at our revenue base, it’s about 60%. The rest comes from the US and the Asia-Pacific region.

And are Europeans using you to create post-training models for their models?

Europeans use us for many things and for different reasons. We help them build applications related to the efficiency of their business or the acceleration of their research and development. We work with manufacturing companies, financial services, and defense companies, and we have a fairly strong verticalization approach in all three areas. The fourth is the public sector.

The way we work with manufacturing companies is generally this: we work with their research and development teams and help them create better products. We help ASML build better machines and run them better. We help Airbus build better planes. We provide reasoning models, but also physical models—models that have a quantitative understanding of the world.

In financial services, we build models that understand markets. We also offer models that understand quantitative time series to help make predictions. We have a set of models that allow us to quickly build applications that bring value to our customers. That’s how we serve them.

Going back to your question about sovereignty, all of these companies have a sovereignty strategy. They need technology deployed in a way that ensures business continuity and availability, among other things.

13. Why Mistral Is Going Full Stack—but Not Building Its Own Chips

We see OpenAI, Anthropic, and DeepSeek moving to the application level and also to the silicon chip level. According to rumors, Anthropic and DeepSeek are making their own chips. Will we see a Mistral chip? How do you assess the importance of verticalization and mastery of each level?

No, that’s not planned. We’ve moved further up the stack—or, I would say, higher up if you look at it from an industrial perspective. We’ve moved from training models to serving the models we train on infrastructure we own. That’s already a pretty important step.

It took about 18 months to set up, and it’s now quite successful because we can sell GPUs. We can sell tokens. We have tokens that are generated exclusively in Europe, which means they avoid any form of extraterritorial US law. This is very important for our customers, so it required some effort from us. We’re doubling down on these efforts to scale this even further, and I think that’ll keep us busy for a while.

We’re pretty vertically integrated: we have cloud services and infrastructure, and we go all the way to building business applications for our clients.

You mentioned the application level. In AI, you bring the platform and build your own applications for customers. How does that work? Is that possible because of the range of capabilities you have?

Yes. It’s functional, but it’s built on a set of underlying primitives: workflow management, agent management, sandbox management, and overseeing the deployment of various business applications. Ultimately, we’re creating cloud services adapted to working and building with AI agents.

How do you think about doing this through Mistral functionalities compared to the partnership economy that we have traditionally had?

It was extremely helpful to have our CEOs at the helm from the beginning of our journey, as it allowed us to understand exactly what challenges businesses were facing. They started with individual productivity, but that wasn’t enough; it didn’t create value for them. So we spent time with managers and business unit owners and understood which core processes they needed to automate with artificial intelligence.

We did this by building business applications—user applications that manage state, where agents manage the process from start to finish. It was extremely helpful for us to understand the primitives for building these applications. For example, we created reliable process execution and figured out how to orchestrate agents very effectively. Having these full-time directors work directly with clients has been extremely helpful.

Since last year, we’ve started distributing these tools to some of our partners, such as Accenture and TCS—system integrators that are great at scaling. Once you give them the right tools and training, they can help scale because they simply have more people. But having people from Mistral sitting at the keyboard with our customers is extremely beneficial to those customers, and it’s in high demand.

This is extremely useful for us. It’s not something we can scale infinitely, because we can’t have tens of thousands of full-time employees working for us. But we’re scaling it through partners now, and these two movements coexist. We have a lot of clients and full-time employees who work directly with them. They manage projects and deliver the work, and we work closely with system integrators to get the work done and handle the maintenance.

14. What Mistral’s Business Model Could Look Like in Five Years

Can I ask you, what revenue line do you not have today that you think will be the most significant in 5 years? As I listened to you, I thought perhaps you wouldn’t have foreseen these different revenue streams at the beginning. When Arthur and I first met, I don’t know if you remember, but you were sitting on a park bench, just chatting with me about life. I’m like... Yeah, you didn’t invest. You should have.

Dude, I’m so sorry about this. I saw Jean-Charles yesterday, and every time I see Jean-Charles, we talk about you. This is a great reminder of how much money I missed out on. Sorry about that.

Well, we have 2 revenue lines. One is where we deploy a platform, which is really licensed revenue for building business applications. Then we have consumption revenue from the infrastructure business we’ve created.

These two things go hand in hand. Our customers usually apply for a license, but then they realize they don’t have enough GPUs. We try to be helpful by providing them with either GPUs or token services. These two things go together very well.

Do you think you’ll have a big reasoning-inference business in 3 to 5 years? Do you have a business in sandboxing?

If you look at this breakdown, I’d say the GPU side—everything that runs on GPUs—is going to continue to grow because the demand is huge. We have what we call ancillary services, for example, running sandboxes. Scaling sandboxes is very easy when you run reinforcement learning algorithms or deploy many agents for the same tasks.

It’s a growing business for us because it comes from customer demand, and in turn, it leads to the purchase of more and more processors. I expect the runtime environment to grow as well: the part where agents perform tasks, call tools, compile code, and run code.

It’s pretty obvious that when you were doing synchronous tasks, just using GPUs to respond to conversations, you didn’t have that. When you start having managed agents running in the background and doing a lot of CPU-intensive tasks, you need more processors, more sandboxes, and more storage because they’re creating a lot of data. So those two things—processor-related tasks and data-related tasks—are going to grow, as is our business.

A large part of your business is selling to governments. What is something people don’t know about selling to governments that everyone should know?

The first thing is that there’s no such thing as selling to governments in general. All governments are very different. Some governments are so small that it looks more or less like an enterprise-level sales process. That’s the case in governments where everyone in the ministries knows each other, for example.

Governments in big countries are a little more complicated. So again, the difficulty of finding a place to create value and understanding the customer’s needs increases with the size of the country, I have to say.

15. Why the Inference Market Behaves Like a Commodity Market

Can you help me understand the inference market? It’s something I just didn’t know enough about. What I know now, and wish I’d known 12 months ago, is that the demand for logical inference is simply endless. The depth is simply amazing. What’s happening in the reasoning-inference market?

The market for reasoning inference behaves very much like the commodity market, which today has quite high margins.

In the sense that if you have an offering with the right quality of service for certain types of models, you can sell them on demand. There’s a lot of demand directly from customers who have no commitments. For clients who are mature enough, they need guaranteed quality of service. They need service-level agreements, and they also need reserved physical capacity.

That’s something we can provide because we have the right forecasting models and the right offering. So you have these 2 things: one that works on demand, which is very similar to a commodity market where pricing is of great importance, and then the premium side of the market. Those customers want access to latency guarantees and generation-rate guarantees for a certain number of requests. The more precisely they can describe the shape of their demand, the better we can price it and shape our infrastructure to provide services on the right criteria at the right price.

16. What Separates Winning AI Labs From the Rest

Can I ask about this? We have so many different service providers near the labs. 102 different service providers were established near the laboratories. Every week there is another profound breakdown of the mind, and another. There are so many of them. What will distinguish thriving labs from those that don't?

First of all, we started out as a lab ourselves. In 2023, it wasn’t called that, but when you’re starting a company, you have a bunch of tools to hire the best, and we used all of them to hire researchers. In fact, we started as a research company.

I think that allowed us to scale and work much more broadly, while keeping the main pillar of providing models. That’s why StarLaws has been very successful for us, and it’s exciting to see the progress we’re making in the new releases coming out. Our success came from focusing early on a market that could scale. We chose the corporate market, which has a certain complexity because corporate clients are demanding, as they should be, and we adapted our product to meet those needs.

Being able to work at the highest level of research and continually invest in creating stronger models, while also building a customer-centric motion and understanding our customers’ problems, allowed us to scale. Every lab working nearby should think about that when starting to build its business.

17. Why Compute Access Is the Biggest Bottleneck for New AI Labs

How many people are there actually in Europe who have the ability to build these companies and have experienced it the way you have? I don’t know. Maybe Yann LeCun has done it. Is it 20? Is it 100?

I’m not a venture capitalist, so you probably know this better than I do. But in general, there’s a lot of talent in Europe. You have a lot of entrepreneurial talent, a lot of research talent, and a pretty good combination of the two. That’s why you see a lot of new labs in Europe.

The most difficult thing isn’t just talent or the ability to sell and build a successful business. Today, it also applies to access to compute. A big problem for a lot of new labs I know is that getting access to GPUs to do interesting research in certain verticals is difficult because of a huge supply shortage. We reserve some capacity for certain new labs because we believe they can grow, and we can grow with them. But if you’re a founder today and want to do something very research-oriented, the biggest challenge is getting access to enough computing resources at a decent price.

Can I ask about research? How crazy is the market for hiring researchers today?

I have to say it’s pretty competitive. If you have the right mission and compensation structure, and you’re doing something important for the world, people will join us; people will join you. We’re very grateful for the dedication of all our employees to Mistral’s success.

The latest Mistral release has been successful. It’s very popular today and leading among open-weight models for coding and cybersecurity. That’s only because of the dedication of this team over the last 3–4 months. Resilience matters, too, because when you train a model, as we described in our recently published technical report, it often doesn’t work the first time. You need to course-correct to make it work, and it can take another 2 months to really get there. That hurts.

We’re expecting a release, but we’ll need to release it a little later. It will be painful for everyone, but everyone has to be fully involved and do the right thing. I’m really grateful to the team for its resilience, given the limited resources we have.

18. Are $100BN AI Training Runs Actually Coming?

I just had Satya Nadella from Microsoft on the show, and he told us about Mustafa Suleyman’s previous statement that we’ll see a hundred billion training runs over the next few years. Do you think that’s a bit exaggerated, or is it really quite achievable?

I think you have a lot more visibility into that than I do. At some point, if you want to make your model very strong in certain areas, like microchip design, or very strong at industrial engineering workflows like the ones we work on, you need the right environment to train it. That part requires skill and taste. It’s not something you can scale as easily as just pooling compute.

I’d say the bottleneck we all face today is actually the ability to produce enough high-quality data. This is a compute problem. We need more compute, and we’re deploying more of it. Achieving one hundred billion training runs is clearly not something that’s happening tomorrow.

19. Why Public Markets Are Punishing AI Infrastructure Companies

If everyone is so short on compute, why is Cerebras down 40%? Why is Cornwave down 40%? Why is the AI trade selling off in public markets when it’s so obvious that everyone lacks computing resources? I’m a little confused about how the markets are reacting today.

We’re fully capable of financing our own compute, given the quality of the collateral we have in terms of revenue commitments and take-or-pay commitments. We operate as a vertically integrated company. We create computing resources for higher-level purposes: serving enterprises and creating business applications. So we’re focused on creating value for our customers.

That means we haven’t had to involve the company too heavily in this financing, and that gives us great confidence in what will happen. But, as with any industrial construction, there’s a temptation to take on a lot of leverage, and at a certain point that leverage can disappear. That’s why we’re very excited to be vertically integrated and to work across the whole stack, from building business applications to infrastructure.

What haven’t you done that, with the benefit of hindsight, you think you maybe should have done?

We didn’t go into the consumer market, but I think we’re very grateful we didn’t. We didn’t get into chip development.

Well, pause there. Why, for example, are you grateful you didn’t enter the consumer market?

I think it’s a very saturated market, and it’s very difficult to work at the highest level with businesses while also managing what is essentially going to be an advertising business. The decision to focus on the corporate segment, and to focus on the 4 verticals we chose—manufacturing, financial services, defense, and the public sector—was very important to the company’s focus.

20. Why Self-Improving AI Still Needs Well-Defined Problems

The problem with artificial intelligence is that it can do everything, and you need a focus. The focus we’ve chosen comes from the market. We want to have models that best serve agentic engineering workflows, financial decision-making, defense purposes, and anything related to state sovereignty.

Regarding recursive self-improvement, do you need a focus? Isn’t that right?

Using AI to improve AI models also requires focus, because you have a limited amount of resources to spend on problems. Ultimately, the bottleneck is indeed computation. The bottleneck is also that anything you do that’s recursively self-improving—or just self-improving, because that means the same thing—requires well-defined problems.

If you have a well-defined problem today, you can solve it with AI. In many cases, problems in enterprises aren’t clearly defined. The real obstacle to economic growth isn’t that we can’t solve complex problems that are clearly defined; it’s that we have a hard time solving ambiguous problems, where you have ambiguous goals and people involved. That part can’t be solved by self-improvement. You can only solve it with models that are deployed and interact with people.

That’s the part we focus on, I would say. But we also invest heavily in creating self-improvement cycles for clearly defined problems. I would say that poorly defined problems are holding back the explosive growth of AI use in the enterprise: you want to deploy AI to solve them, but you can’t because the problem definition is wrong. So you need to spend some time on it. That’s why you have forward-deployed engineers. That’s why we have FDEs, to try to better define the problem.

21. Does Mistral Feel the Pressure of Being Europe’s Best Shot at Frontier AI?

Arthur, I have to be honest with you. I joked at the beginning that you were Europe’s last hope. There’s a very large gap between you and anyone else among the nearby labs that aspire to be the next Mistral.

Are you and the team feeling pressure?

We feel some pressure, but like anyone in this industry, we’re in a fast-growing, fast-changing field. As an AI entrepreneur, you need to work with a model because it’s the only thing that shows your progress. Then businesses need to see the value, so you need people who understand the real business problem. That’s quite a different problem to solve.

We feel some pressure, but I think it’s very healthy pressure. This is a very important moment in time and in history. We need to meet the expectations of public opinion. Ultimately, we need to show that AI is useful. The type of implementation we’re doing and the noise it’s creating in the market mean this technology should look positive. We need to show the positivity of this technology.

The only way to do that is to promote sovereign AI as a category. It’s about making sure that it creates economic growth and opportunity for everyone, and that the benefits of AI are distributed appropriately. Of course, that creates some pressure, because it’s not a given. If we don’t succeed, I don’t know what will happen in the world.

I think we have a role to play. The mission we’re on is important. It’s pressure, but it’s also quite exciting.

22. Can AI Add More Than 2% to GDP Growth?

Will AI increase GDP growth by more than 2% per year?

That’s the whole question. What bottleneck will you solve to actually impact GDP? That’s why we were very optimistic about industrial engineering. Ultimately, if you want GDP growth, you need more production. You don’t need more intelligence; you need more production.

To have more production, you need better factories. You need objects that are more efficient. You need better planes that fly longer with the same amount of oil. You need better energy production that’s more efficient and less carbon-intensive. You need better ASML machines. You need better services in financial services as well.

Particularly in industry and engineering, the use of AI allows for better design of objects. If you set up an agent cycle to design safer cars, for example, your car design team can work much faster and iterate on that product faster. This means you can sell more, sell cheaper, and create a safer product.

In the aerospace industry, if you design the best airplane, it will be on the market for 40 years. If it runs faster with less fuel consumption, it will lead to an increase in GDP. Ultimately, it’s really important to give AI a physical effect. We see this effect mostly in manufacturing and financial services.

23. Why Sovereign AI Could Become the Most Important Enterprise AI Category

Dude, I’m aware of the time. I’ll have a quick chat with you now, okay? This is something new. Good. What do you think most people around you would disagree with today?

I think we strongly believe that there is market fragmentation, and that sovereign AI is going to be the most important category of AI adoption in enterprises. When businesses are looking to implement technology, business continuity, cost, and control over their data and intellectual property will be what matters. That’s a very strong belief that we have.

This is shared by many businesses and a significant portion of the market, but certainly not by some of our competitors.

24. How AI Will Change Every Engineering Profession

What is an extremely wild prediction today that will be very common in 5 years about how we live and work?

I think the way we work as engineers—what has already changed for software engineers—will completely change for any type of engineer. I’m really excited about this approach to developing better products that is pretty autonomous, thanks to artificial intelligence.

This doesn’t apply to everyone, but everyone with a technical background—not just software developers, who were the first to be affected, but mathematicians and everyone with a technical background—will be controlling AI agents in a way that gives them enormous influence. This in itself will be a significant shift in how the labor market actually works.

25. Why Mistral Could Eventually IPO in Europe

What do you think about going public? Because one day you’ll want to, right? But you can’t go public in Europe.

You can absolutely go public in Europe.

I would say, yes, you need to go public in France?

No. What was true 10 years ago, I think, is no longer true. There’s a huge appetite from investors all over the world to invest in Europe, both from private capital and public capital. The barrier that used to exist is no longer there.

At Mistral, we’re certainly preparing for the fact that, as you know, you have a lot of operational constraints when you go public. We’re also looking to hit the commercial milestones that we need to hit, but I don’t think that’s a bottleneck anymore. There are very successful companies, some of which are first-time listings in Europe. There’s no reason why we shouldn’t do an IPO in Europe at some point.

26. Why AI-Driven Scientific Discovery Excites Arthur Most

One last question for you, dude. What are you most concerned about as you look ahead to the next 5, 7, 10 years?

Few people think about this. I’m very excited about how AI will enable fundamental research in areas where we lack understanding of the world. For example, in physics, the fact that we can run AI simulations on a scale that we’ve never achieved anywhere—and this team that we bought, called Emmy, specializes in this—will enable new discoveries.

The fact that we can create simulators, and have agents that run these simulators to better understand the world, is going to be a huge driver for understanding the world. I have a background in mathematics and physics, so I’m really excited about the opportunities that lie ahead for us in the next generation of researchers.

You know what, Arthur? I blame you. I blame you for when you sat on the bench and spoke completely differently than you do now. Arthur then and Arthur today are paradoxically different, with respect. It’s amazing to see the leader you are and the leader you’ve become. I honestly, man, have so much respect, admiration, and gratitude for you and everything you’ve done for Europe. But wow, this is great.

Thanks for all the praise. We still have a lot to do, but we’re quite pleased with the progress we’ve made. I think the mission is important, so if anyone who’s watching us wants to join us on this journey, please contact me.