Why Mistral Just Raised €3B to Build Sovereign AI | Lightwork

0:01Mistral, a three-year-old company in Paris, just raised a$3 billion Series D. It's now the largest equity round of any European tech company has ever been done. Antoine, you were there for the seed. James, you've been involved on the growth side. And Lightseed's been there in every round since. So I figured it'd be good to get the both of you on just to talk about what happened here and where you see the company going. I guess to start, I'd love to hear a little bit more about both of your involvement. Maybe Antoine with you first and then James. But just first, how you got connected with the Mistral team and just level set a little bit of how that has changed over the last three years too.

0:36Yeah, some of the Mistral lore. So, I mean, look, we've been spending with Lightspeed quite a bit of time looking at tax modalities across the world. So we looked across India, Israel, Europe and the US at a bunch of different bets being made. And we looked at this with the teams in all those different geographies. And when we got the opportunity to meet with the Mistral team, it intersected with a bunch of themes that we were paying a lot of attention to. It feels weird to say this now, given how much the world has changed. But back then, the universe of founders that had had experience training foundation models at scale was very small.

1:13And the founding team of Mistral was part of a select group. The focus they had on open weight strategies was one that we thought was very differentiated at the time as well. And the background of Guillaume specifically on the Lama side of things alongside with Timothée, given the really good results they'd had back in those days were a sign and a testament to the caliber of the engineering team. And the compliment with Arthur coming in with his Google expertise was something that we're paying a lot of attention to. And I guess in true venture fashion, we kind of got to know the team through a mix of different channels.

1:46I was attending a research conference on using AI for formal math proofing, which back then wasn't really a big topic. And we got connected through a founder to Arthur at the time when he was thinking on his side of things. So that was kind of the genesis on our end. Oh my gosh, math conference. And James, how did you get involved? Yeah, happy. First of all, great, great. Thanks so much to have us on. I've been at Lightspeed, you know, for the last eight years. I'm a part on the group team backing software and just AI companies at a scale. And so I've been involved with just a number of these companies that we've backed, such as Grafana, Glean, Wiz, you know, Stride and Databricks, and just seeing up close how some of those organizations have been built, how they aggregate capital to fund it, where really others stall.

2:30Like Antoine said, the company has kind of been evolved from a frontier lab into a vertically integrated AI company with open models, an enterprise platform, and a forward deployed engineer strategy inside their customers to really drive adoption and ROI, and now kind of doing its own compute. And similar to Antoine, I was also born and raised in Paris for my first 12 years. So it's also a good excuse to come back to my roots. Yeah, I do want to definitely get into this really large Series D announcement and more on the raise. But before that, maybe Antoine, since you've sort of been part of the journey since the start, what has changed the most since you first met them?

3:07And especially, you know, even between like the last valuation between 11 billion to 21 billion, what has changed between all of those rounds? Well, I mean, I think I could give you an answer around how much the company has changed. I think there's kind of two layers. One piece, I think, is just seeing the founding team grow. I mean, when we made the bet on the team, it was very much, you know, a research team that hadn't had much company building experience. And, you know, it's already one thing to be an engineer and grow into a CEO and a leadership team. It's another to go from research to engineering to products and then to commercialization.

3:45And the team has had to take on all of this in a very rapid, rapid pace. I think also has changed quite a bit. And, you know, if we're referring to the last round relative to this latest announcement, the geopolitics of things have accelerated a lot of the underlying trends, you know, the topics of owning your AI, the vertical integration, a realization that AI diffusion has to come through the handholding and the, you know, helping those large organizations think through what are the core use cases that they can work on automating and using AI for. And those are trends that, you know, even I'd say 18 months back weren't necessarily the most debated trends, but over the last six months have been pretty undeniable.

4:29And so it sounds like there's obviously a lot of emphasis on sovereign AI and, you know, with the series, the announcement, James, you being on the board, like what is, is that part of the impetus for the raise? Where does a lot of this money go to? Yeah, no, it's a great question. And, you know, doubling down on Antoine's comments, I think the company really is at an inflection point right now and wanted to capitalize on its momentum. You know, what does that mean? Scaling compute capacity, building out its infrastructure, accelerating commercial growth and its product offerings, and also expanding in terms of its international footprint.

5:05And so even with this raise, actually the strategy remains the same. Frontier research, you know, deep enterprise verticalization and a full stack AI cloud on own infrastructure, you know, from token generation and deployment observability to the runtime for agents and the interfaces where humans actually work with them. And so this raise, you know, positions Mistral globally as the open-way platform, you know, so that customers can have customizable, small, efficient, multilingual models that deliver both in terms of performance and cost efficiency. And so this round also brings in Samsung and also EQT onto the cap table and also deepens the partnership with ASML amongst other strategics.

5:54And so enterprises are going to need sovereign infrastructure to deploy AI systems. And that's exactly where Mistral sits and what this capital is going to enable them to accelerate. Going a bit deeper into that, that open way aspect, I think for your average person who hears a lot about the news, a lot of the open-weight noise has come from China. And you all have ended up backing up one of the most important open-weight labs in the world, especially in the West. I think to make it tangible for folks, why does it matter that someone in the West is doing this? And why does sovereign AI matter for that matter?

6:27Maybe I'll take it off to you, Antoine. Yeah, so I think there's different ways to approach the question. If I start on the sovereign AI side of things, I'd say that sovereign AI has gone through different cycles of definitions. I'd say that if we were defining sovereign AI two and a half years ago, it was maybe through the lens of regulation and the EU AI Act. It was through the AI safety summits and what sovereignty meant through the lens of countries. I'd say that today sovereign AI has taken more of a broad definition in the sense that yes, countries and specific blocs in the world are thinking about their own national interests.

7:04But the understanding of sovereign AI through the lens of as a company, owning my data, owning the compute and being able to really, you know, control what we're doing with the model is something that matters a lot. And it ties itself naturally into this open weight strategy, which is really about controllability. And I think what people care about and the reason for which it matters in the West is really having models that are able to compete close to the frontier, if not at the frontier, while giving customers and the enterprise more broadly the ability to control their models. It's interesting you framing this because, you know, obviously a couple of years ago, a lot of people were pretty focused on just like bigger models or capabilities, but now sounds like control is the bigger question here.

7:49Are you seeing that with Mistral's customers? How has that translating into the Mistral story as well? Because I think another thing that I felt like was very clear is that Mistral is not trying to be, you know, European chat GBT. What do you feel like they're trying to fill in this ecosystem? You want to take it, James? Sure. Yeah. I mean, look, I think, you know, if we were to take a step back, Antoine kind of mentioned it, but, you know, open weight models are inspectable, portable, and deployable on your own infrastructure. And if you take a step back, and what we've seen in the last kind of six to nine months, mission critical workflows and customers are demanding these open weight models.

8:29And so there's been a lot of buzz around open weights reaching the frontier. And a lot of it has come from China, which is great for the field, but a real problem for anyone running AI inside a regulated environment, whether it's a bank, a government, public sector companies. And so you don't really have no sovereign buyer in Europe or the Gulf or even Asia is putting critical workloads on weights that they can't trust the provenance of. And so you're in a bit of a no man's land and that's the biggest tailwind for Mistral. The best open models are becoming increasingly Chinese and then the best closed models are American.

9:09And there's an enormous set of buyers that can use neither. And so Mistral right now is one of the only Frontier Lab that sits at an intersection of the two and can capitalize on it. Just to add to this, I think one of the interesting elements with the strategy that Mistral has articulated over the last few weeks around thinking about serving other models beyond their own models and thinking about third-party models is the ability to serve such models with the constraints and oversight they're able to provide with their own models. and this flexibility to think about, you know, providing open-weight models that are closer to the frontier, if not at the frontier, enables enterprises to think about it in terms of, you know, cost optimization, efficiency, but also in terms of the frontier.

9:53And I think it's not something that's mutually exclusive between enterprises for their really toughest problems that are still maybe a bit fuzzy and undefined, thinking about using frontier models, while for the defined workflows, thinking about how to define agentic workflows with models that are maybe open-weights, if not at least certainly more efficient. Because when you're running those workflows at scale and you're thinking about this AI diffusion across hundreds of thousands of employees, the bill starts just growing very rapidly. And over the last few months, we've been seeing this in the general conversations.

10:23Conversations around routing, around the optimizations of models are certainly things we've been starting to hear much more of. Yeah, it sounds like they are certainly kind of going across a stack of models, infrastructure products all at once. I want to get to that. But I think I even saw in the announcement that they are wanting to build a gigawatt of European compute by 2030. So I'm curious to hear more about that build out, James. Yeah, no, for sure. It's such an important question. And as we've seen with every other lab, staying competitive and differentiated requires a compute strategy. And so Mistral's compute strategy really has three legs.

11:03One, training its own frontier models. And then the second one is providing compute to their own customers, because securing capacity for them is a strategic offering that helps with adoption, customization, and privacy. And then the third one is providing compute to hyperscalers. You know, we saw with the recent Microsoft deals, where Microsoft's actually going to leverage Mistral's expanded Europe-based GPU infrastructure to support the delivery of its own cloud and AI services. Now, you know, if you take a step back, why do it all? And I think that Europe has no homegrown hyperscaler. And so Ms.

11:43Froll's deepest customers in financial services, defense, manufacturing, and even the public sectors are exactly the ones that need independence from the US clouds. And so if you don't own the infrastructure, you can't really sell sovereignty. Anton, you've obviously been a key part of our Europe team. How do you feel about this sort of stack that they're building? How does that evolve from the original vision since Seed and would you have expected that they go into this full stack vision? I mean, the world has evolved a lot over the last three years, so I won't claim to have been able to forecast everything that's happened since.

12:21I think that one of the things that was very important to Arthur from the get-go, it maybe wasn't framed through the lens of sovereignty, but this idea of saying, how do we make sure that we're able to provide a third way a capability for companies and countries to think about how they could own their AI. And when it comes to the compute piece, the latest European compute unit proposition that they've had around securing demand to be able to really build out this type of compute is something that we found to be very creative. Because as James was mentioning, the reality is that when looking for offtake agreements and trying to build the types of demand that are required to be able to fund those very ambitious compute projects, there aren't just that many today.

13:04And being able to go to some of the largest enterprise buyers and bringing them in on this European narrative around building this compute that is homegrown and that needs to be owned by Europe is something that we think is very important. And the reality is that this narrative on the compute side isn't just owned by Mistral. And the more compute players are able to provide this in Europe, the better. And we actually think this beyond Europe. It's not just a European topic. People in the MENA region, in the APAC region are thinking about this themselves. And so far that Mistral can support them as they think about their own developments, they'll be doing it.

13:36Yeah, that makes a ton of sense. I guess that maybe ties to my final question for you both as we come up on time here. As you think about, you know, the Series D over the next 12 months, next couple of years even, what is exciting you most about broadly the space, but also Mistral in particular? I'll kick it off to you, James, first. Yeah, I think a few things come to mind. I mean, first, you know, Arthur and the team here have been one of the strongest with regards to just overall product velocity. So I think just the product doing more of the work, you know, Mistral built this on embedded engineers.

14:11And the next phase is also going to be customers really building and adopting the platform even more. Second, and we've talked about it for the last few minutes, but it's really having the compute coming online and having the majority of their customers running on it. And then lastly, Antoine just touched upon it, but it's also so important, which is just geographic expansion. The company is and has been doubling down and solidifying its European leadership. But the sovereign playbook actually is going global. And we're seeing some of that pull, you know, in the last kind of six to nine months, whether it's in Singapore, you know, Middle East, India, South Korea, you know, Mistral has the operating model for it.

14:51And so I do think more broadly, you know, this is the year or in the next year, the market is going to stop asking who has the biggest model and start asking who controls the intelligence that they actually run. And so Mistral was built for that, you know, really before anyone was asking it. And so we're, at least from our stance at Lightspeed, we're really proud to have participated in every financing rounds, especially the latest one, which is the largest private financing in Europe's history, and to remain one of the largest shareholder on the cap table. I'll give maybe a more aspirational answer on it from the tech side.

15:26I think that the company has been focusing a lot, especially with the MEAI acquisition not too long ago, on physical and critical industries. And I think that if Mistral is able to support some of the leaders across, you know, industrials, defense, robotics, those are the types of topics that I'm personally very excited about. And, you know, I was working in data science before, and one of the topics I've always found fascinating is CFD modeling, which still to this date is not really a solved problem. And if Mistral, along with its teams, can work on helping accelerate this change, I think it's something that will be pretty exciting across the physical world.

16:02I love that. It's such an energizing note to end on it, that there's just still so much on frontier to be done. But thank you so much for taking time out of your day. I know it's a busy day and it's just so helpful to have two people who've been so involved kind of demystified beyond the headline. So appreciate it. And hopefully we'll have you back for more. Thanks, Claire. Sounds great. Thank you.