Can Mistral Challenge Big Tech by Doing AI Differently?

Author: Bloomberg Television · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Can Mistral Challenge Big Tech by Doing AI Differently? · Channel: Bloomberg Television · 10:58 · Watch on YouTube

Topics: AI governance and regulation, Markets and economy, Local and open-source models

Summary

Bloomberg examines Mistral’s effort to challenge U.S. and Chinese AI companies through open-weight models and sovereign AI services rather than competing solely on scale. Co-founder Arthur Mensch argues that open models give enterprises more control over infrastructure, sensitive data, costs, and business continuity. The report also highlights Europe’s constraints, including fragmented capital markets and difficulty scaling startups, alongside uneven AI adoption across the region. Mistral says it can serve customers who need control over important workloads, even when they use hyperscalers and closed models for more generic tasks.

Key points

AI investment has surged, while U.S. hyperscalers and Chinese firms dominate the race for advanced models.
Mistral promotes open-weight AI as an alternative to centralized, closed-model providers.
Mensch says open models let businesses choose their infrastructure, keep sensitive data close, and reduce costs.
Mistral says it has developed models with less capital and compute than some U.S. competitors, and is expanding its capabilities.
Station F director Roxanne Varza says Europe has a strong early-stage ecosystem but faces challenges in growth funding, exits, and scaling across fragmented markets.
Jacob Kirkegaard says Europe’s incomplete economic integration hinders both AI model production and widespread adoption.
Kirkegaard distinguishes between producing AI models and adopting them, while noting that relying on foreign models may raise national-security concerns.
Mistral has shifted toward AI services and open models rather than trying to remain a frontier-model lab, according to Kirkegaard.
Mistral positions sovereign AI as especially useful for high-stakes enterprise work, while customers may still use closed models for generic tasks.

Chapters

Start	Chapter	Summary
00:00	AI’s investment race and Europe’s underdog	The report outlines the scale of AI investment and introduces Mistral as a European challenger seeking a different path.
02:04	Mistral’s case for open-weight models	Arthur Mensch explains how open models can offer enterprises infrastructure flexibility, data control, and lower costs.
04:09	Scaling Mistral and Europe’s startup hurdles	Mistral discusses its compute and capital constraints as European ecosystem leaders describe the difficulty of scaling companies.
06:47	Europe’s AI adoption and strategic shift	Jacob Kirkegaard discusses uneven European adoption, the importance of AI users, and Mistral’s move away from frontier-model competition.
10:28	Competing through sovereign AI	Mensch says Mistral serves customers seeking control and sovereignty for important workloads, alongside their use of hyperscalers.

Quotes

The way in which we should look at the world is rather an opposition between the closed model providers — 02:04

a lot of customers of Mistral are actually coming to us because we come and deploy our models close to their data — 03:08

There is actually only upside. — 04:09

The European capital markets are still very fragmented. — 06:15

We bring sovereign AI to our customers — 10:58

Concepts and entities

Mistral, Arthur Mensch, Mistral Large 4, open source models, open weight, hyperscalers, Station F, Roxanne Varza, Jacob Kirkegaard, Anthropic, OpenAI

Full transcript (original language)

00:00 Westin: This is a story about rooting for an underdog. The artificial intelligence race until now has been dominated by a handful of U.S. tech giants. -I believe that the future, if we get this right, can be much more like a new renaissance, than a new industrial revolution. -Whoever wins AI wins. I think that’s very important. -It’s productive. It’s profitable. It’s generating enormous profits. Westin: This small trio of executives and the technology they created have become household names in much of the world.

00:31 And they’re committing unheard of amounts of capital. When ChatGPT hit consumers’ phones in November 2022, investment in AI was $112 billion per year. By the end of 2025, that number increased more than 200 percent to $345 billion. And this year it’s on track to blow past even those vast sums. -When we look back to 2025, where really this borrowing from hyperscalers had really only begun.

01:02 Through the course of 2026, we’re probably looking at around $300 billion in the investment grade market borrowings from hyperscalers themselves. -Next year, we’re expecting another 15 percent more in issuance coming from some of the hyperscalers. Westin: The U.S. hyperscalers are largely up against Chinese firms backed by the resources of their entire government as they pursue lower cost alternatives that require less compute. Lost in this titanic struggle thus far has been Europe, which hasn’t seen the levels of private

01:33 or government capital committed to frontier AI models. But France does have a dog in the race, an underdog to be sure, but one that seeks to redefine the competition in a way that it believes can put it at the front of the pack. -The way in which we should look at the world is rather an opposition between the closed model providers that are promoting a centralized version of artificial intelligence and trying to export it to the entire world and an open version of artificial intelligence where you have multiple actors

02:04 that are contributing to an open ecosystem and then helping enterprises use that and customize their systems on top of an open infrastructure. Westin: Arthur Mensch is a co-founder of Mistral, the leading AI contender for Europe, and he believes it has a built-in advantage over the closed source closed weight U.S. hyperscaler approach. -We actually are the leader in that domain. We are now the best open source model providers for sovereignty deployment. That gives us a special status that I think goes way beyond

02:35 the fact that we’re European. Westin: So on the distinction between open source open weight and closed, what are the advantages and what are the limitations of open source open weight the way Mistral is? -So the advantages for the enterprises are threefold. The first is that you have business continuity guarantees because you’re decoupling the IP layer, which is the model from the infrastructure layer. So you can actually bring the model into the infrastructure you want. You can use multiple providers and you can make sure that whatever happens to the model providing company, you’re still going to be up and running as a business.

03:08 The second thing is really closely linked to the fact that, when you want to have return on investments on your AI deployment, usually those AI deployments are dealing with very sensitive data. And in many cases, this is not the data that you want to share with a centralized AI provider. And what that means is that a lot of customers of Mistral are actually coming to us because we come and deploy our models close to their data in a way where we don’t even see that data. The third aspect, which has grown in importance

03:39 in the last few months is the fact that AI is getting expensive. And so in order to build the business cases of AI deployment, you do need to take into account the cost. And the cost, if you pay the IP layer, if you pay for the models provided by centralized AI providers, you’re actually going to pay much more than if you’re paying for open source models being deployed on your infrastructure, on the infrastructure of an open source model provider. What that means is for an enterprise you are asking me about the downside and the upside.

04:09 There is actually only upside. There is no reason why not to use open source models today, except the fact that it’s a little bit harder because it came a little later. But our job at Mistral has been to make it much simpler for our customers to deploy AI– AI systems on top of open source model. Westin: Mistral recently released its latest open weight model, the Mistral Large 4 or “Le Chonk” that raised 3 billion euros to scale its innovation. -We have been under capital and compute constraints for the last two years, but we managed to produce

04:42 a model extremely strong despite the fact that we were not controlling a lot of compute. This is actually about to change because we’ve made the right investments for doing so. So we’re very excited by the upcoming roadmap because we’re just able to scale the training. We’re able to scale inference capabilities for our own customers. We do it in a way which is… which I think is a little more frugal. We’ve been able to produce– the kind of technology we’ve been able to produce we produced with much less capital than some of our U.S. counterparts.

05:13 Westin: Mistral is taking a different approach to AI, one that requires frugality. But how much of that difference in approach is required because of the limitations inherent in being a European tech company? -Mistral is… I wouldn’t say it’s the exception. It’s probably one of the first that we’ve seen to this scale and size in Europe. Westin: Roxanne Varza is the director of Station F, a European startup incubator. -Our ecosystem is somewhat less mature. We’re catching up. We’re picking up pace. I also think Europe’s fragmentation

05:44 has long been an obstacle. So you’ll have usually a local dominant player in France, in Germany, in the U.K., that will be big in that country but it won’t be at the scale of what you have in the U.S. The gaps, I think, still are mainly in late stage. I think we service early stage very, very well. We’ve kind of filled in all the funding gaps. People are now building. But being able to scale, being able to find growth capital being able to exit, those are still challenges that we have to get over. I think the European capital markets

06:15 are still very fragmented. I mean you have local markets in quite a few countries. And then if we talk about actually acquisitions, acquisition culture is starting to pick up. But I do think a lot of the acquirers still come from abroad. There’s kind of this general feeling that if you want to be acquired, the acquirers will come from the U.S., might start to come from Asia now, but less so in Europe. And so this is really something that we need to correct. Westin: Jacob Kirkegaard is a senior fellow at Bruegel. He agrees that Europe’s lack of a single deep capital market

06:47 makes competing in AI a challenge, but also that it could present an opportunity. -So overall there’s no doubt that the still not fully integrated European economy in that sense is a hindrance both to, I would argue, model creation or the production of AI models, but also to the seamless harmonized or similar even diffusion of AI. There is no doubt that Europe, as it exists today,

07:18 will see significantly faster AI adoption. I believe in mostly northern Europe, Scandinavia, the Baltic countries, the Netherlands, other countries with relatively lax or relatively flexible labor markets, as opposed to others, where it will take longer than that. So in some ways this, you know, non-harmonized, the unharmonized or incompletely harmonized European economy

07:48 does benefit some countries while setting others back. Westin: For Kirkegaard, Mistral’s position in providing AI services rather than frontier AI models may be critical in delivering for Europe the productivity gains many anticipate. -When you’re talking about AI and broad productivity benefits or productivity growth boosts, it’s very important in my opinion, to make the clear distinction between AI producers or model generators creators,

08:21 the sort of Anthropic and OpenAIs of the world, and then the AI users or AI adopters. And this is really the key to understanding AI as a technology is a general purpose technology. From an economic point of view, there is obviously no doubt that who produces the models are not really that important as long as you have access to them. But the problem with that argument, some would obviously immediately say, is national security. Because you could very well, in Europe,

08:55 if you sort of hypothesize a world in which Europe also relied on almost exclusively U.S. and or Chinese open source models, that one day Washington and Beijing would say, “All right, we ban the exports on the supply of these models “to foreign customers.” I think in many ways Mistral is… it’s certainly an outlier in the sense that it is arguably the most advanced, largest,

09:26 sort of pure AI company in Europe. It was, I would say, until recently, close to, if not at the frontier of model development. They have now, as I interpret their recent strategic shift, decided not to compete or try to… or invest to develop closed frontier models anymore. So they are a company

09:56 that really is more in the provision of AI-related services and open source models rather than a, you know, frontier lab, so to speak. They are no longer, I would say, they don’t want to be viewed in the same category as, you know, Anthropic or OpenAI. Westin: And it’s in that category of AI provider that Mensch’s Mistral thinks it has an advantage. The category of open source, open weights,

10:28 that give the enterprises using AI critical digital sovereignty. You just had a new round at 3 billion euros, I think, it was reported. That is a valuation of around 21 billion euros, maybe 24 billion U.S. dollars. Very large company by most standards. But compared to the hyperscalers, it’s orders of magnitude smaller. Can you compete with the really big guys? -We play in a different category. We bring sovereign AI to our customers and our customers are sometimes using hyperscalers for some of their workload.

10:58 They’re using closed models for their very generic workloads, chatbots, etc. But when it comes to deploying it for building new airplanes, for building better machines, for finding more sources of energy, they typically want to have a solution that is… that gives them the control they need, that gives them the sovereignty they need over their deployment.