It's a spiritual thing.
Heat.
Heat.
Hey.
Hey. Hey.
It's a spiritual
I don't want to give you anything. I don't want to give you up. Fight for love. Fight for love.
Heat. Heat.
I don't want to give you anything. I don't want to give you up. Fight for love. Fight for love. I don't want to give you up. Fight for love. Fight for love.
You know what's important is what we bring when we bring in the joy and the music when we work together.
Hey.
Hey. Hey.
It doesn't matter. You got to bring that light and understand that music in whatever form will help us, will heal us. So that stuff that I said on that house, it's true. It's true. It's no sh
Super fun.
Got to Baby heart.
Got to be with you. Thank God with me. Thank God with me. Bring me down.
It's been a while since I saw your face and nobody. So when I knew that you could see me too, I wanted you all for myself.
It's been a while since I saw your face and nobody else. So when I knew that you could see me too, I wanted you all for myself.
I don't want to focus on
10 9 8 7 6 5 4 3 2 1 zero.
AI capability is advancing faster than our collective ability to shape its direction. How can I run a company on an annual or a quarterly planning cycle when the whole world is changing every month?
The race is no longer defined by breakthroughs alone. It is defined by the decisions that follow. Surely then we need sufficient means of control before it is all too late.
How fast do we scale before capability outpaces our ability to govern it?
GPU consumption is going through the roof and even GPUs we sold four or five years ago now are rising in price faster than good wine.
How do we build safety and trust without slowing the innovation that drives progress? The future in my view is still to be written, but these next few years are going to be very critical as to which way that will go and how we collectively want that to look like.
How do nations secure sovereignty over technologies on which they increasingly depend without fragmenting the global ecosystem that made them possible? For Europe, the answers will determine its place in the AI era. We want to make the most of artificial intelligence to improve our societies, our quality of life, and our productivity.
This is the crossroads. Not whether AI advances, but how fast, how safely, and on whose terms.
A sufficiently dangerous bit of nefarious activity could have massive consequences, even if it's just a one-time kind of slip up. From the infrastructure and capital powering AI to the machines entering the physical world. From quantum computing to scientific discovery. From creative industries, what's a life worth? To national security, the choices made now will shape the balance of power for years to come.
Whoever wins with AI wins.
Welcome to Bloomberg Tech London. Please welcome to the stage Tom Giles, senior executive editor for global technology at Bloomberg, and Tom McKenzie, anchor at Bloomberg television.
Good morning everyone and welcome to Bloomberg Tech London. I'm Tom Giles, senior executive editor for global technology at Bloomberg.
I'm Tom McKenzie, anchor of Bloomberg Tech Europe and also co-anchor of the opening trade. Welcome everyone to Tobacco Dub. Really excited to have you here.
It's great to be here, Tom. Last night we kicked things off with a terrific dinner for C-suite executives. We started off with Max Järberg, CEO of Isomorphic Labs, and he walked us through how AI is transforming drug discovery. We also had BrainCo's Nix Hu, and she explored the next frontier of the brain-computer interface, explaining how BrainCo is looking to build a digital version of a GLP-1. And both of those stories, if you missed it last night, are available on bloomberg.com and the Bloomberg terminal. And we wrapped up the evening with a conversation with BBVA, the head of AI transformation and the head of OpenAI's forward-deployed engineers, to talk in this moment of acute focus on return on investment, on how to use frontier models to really transform a global bank. And today we're going to broaden the lens and we're going to think about, of course, talk about AI, of course, but also defense tech and European sovereignty as well.
Technology, as you know, is moving extraordinarily quickly and nowhere is that more apparent than in the world of AI. But at today's conference, we're concerned not just with what the next technology will do, but also what happens when these systems move at scale into businesses, into economies, and into societies. Questions like who controls them? Who builds them? Where does value get created and what does all of this mean for Europe?
So over the course of the day, we're going to be speaking and hearing from people building some of the most consequential technologies, some of the most consequential businesses, with a focus on things like, of course, voice AI. We're going to be hearing from Mati shortly from ElevenLabs. Defense technology as well, crucial of course for Europe and sovereignty, and everything ranging between that including the world of entertainment as well, how AI is changing that.
But before we do that, I want to take a moment to say thank you to our sponsors. First, our presenting sponsor, Salesforce. Our supporting sponsors, EY, IDA Ireland, and Schneider Electric.
Okay. Well, we have a packed program. Tom, we're going to kindly see you off stage. Tom's going to be back with some fantastic conversations. So Tom Giles is going to briefly leave and be back shortly with Palantir. But before we get there, let's welcome up our first guest, our first speaker. It's Mati Staniszewski, who is the co-founder, the CEO of ElevenLabs. It's a company now worth about 22 billion US with about $600 million ARR. Really one of the most significant European startups with a global reach. So please welcome to the stage Mati Staniszewski.
Hey, great to see you, Mati. Great to see you.
Thank you, Mati. Thank you for doing this.
Fantastic to kick things off with you. I just kind of ran through some of the numbers. You can give us an update if things have changed. $600 million ARR, valued at 22 billion. You're shipping a lot of product and you are scaling globally, you and the team at ElevenLabs. It has been quite the year. How's it been for you?
It's, I think it's a huge pleasure to be building in a current time and age. I think for all of us, and first of all, good morning as well. It's the biggest shift that has happened since the internet, where AI is changing how our business operate, how we operate, and having a pleasure of building at that time and leading some of that change is just an incredible privilege. So this year has been very quick. We grew the team from 300 people to now over a thousand people across the global parts, from the forward-deployed engineering to scaling our research team to the engineering side, while as you said, growing the revenues, opening offices around the world. We really believe in the global approach but having the hyper-local presence to be able to work side by side with all of our partners and happily still leading both as a company combining research and product, still leading on the research side. Building recently, we released our new generation of voice models that happily not only, you know, what I say is of course biased, but independent benchmarks will validate that these are the leading models across text-to-speech and transcription models, and then building a great product with conversational agents that you can deploy across customer experience, across sales, across operations to help businesses communicate with their audience in a much better way.
But that's important because there's a bit of a misnomer that Europe doesn't have really advanced models and isn't building really advanced models. You are a proof point that that is not correct. You are building those advanced models, and consumers being a big part for you, but increasingly the enterprise as well. And you've recently come out with new agents or new models, I should say, voice-to-voice models, that's relatively new because you're doing everything from podcasts to voice agents to increasingly you're looking at music as well. I know that the platform is significant and is growing, but talk to us about the most recent models, how much traction you've seen there, why you decided to go in that direction, and what the potential growth is for that.
So as a quick state of the business, we are 600 million in ARR, the most recent public numbers, and majority of that today is the enterprise adoption across our work. As a company, we are building the interaction layer of AI. How you interact with AI is one of the key enablers of being able to benefit with AI, AI to being able to serve the customers around the world, and that's the problem we want to fix on the research side. What we believe in is we started five years ago and we started with the first voice model that could turn text into speech and sound human. So finally you had the right intonation, the right emotions, and over the years we expanded to all different types of models. And today, as many many people will see in the industry, you'll have different variations of all the ways you can use speech models. You'll have the component models, you'll have speech-to-speech models, you have duplex models, and then of course voice is just part of it. The intelligence, the context, and how you bring it in conversation is even more essential. It's not only intelligence, it's also the emotional intelligence and how you react in those conversations. So as a company, we effectively build an interaction engine that helps you decide which models to use at which points of the conversation so they can lead to the right outcomes. You don't need a Ferrari for every task. So if it's a basic conversation, just an FAQ query or asking about the product, potentially one of the simpler, quicker, more efficient models is sufficient. But if you authenticate the user and you refund a query after they go through the account, here potentially a more reliable and good model. And as we go through the flow of the conversation with the user, it's not only the models. At some point, you need to know when to bring the human into the conversation too. And that's the second big thing we believe, which is it's going to be a lot of the AI agents, but knowing when to bring the human into the picture and how to continue the contextual layer is important. And across that, that's the biggest part that's driving a lot of the revenue: how you combine the models together with the knowledge, with the deployment, with the safety and guardrails to build conversational agents. And maybe speaking about some of the use cases on how some of the companies here are leading with deployment. Recently we worked with Trainline. I hope there are a few users here of Trainline too, but of course if you are booking a ticket or if you're looking to rebook it or there's a cancellation happens, there's a high flow of traffic happening, and then both through chat and through voice you can engage with Trainline to rebook your ticket very quickly, leading to a few over five times quicker resolution time than previously was possible for those same tickets. And then all the way to Revolut who uses that for powering a lot of the in-app support that customers can have by opening the app and being able to engage in a more natural and quick way, and now looking to engage across different parts of their business too, not only support but how they can proactively offer customers additional information about the products and services and potentially lead to that connected customer journey. So to maybe summarize your question, the biggest things we are seeing are driving it: one, the Turing test for conversational experiences, combining a lot of the intelligence, emotional intelligence is being solved. Two, that combination of AI and human working together is clearly working well to contain and resolve customer queries better, and that connected customer journey is kind of the last key parts that we are seeing across the space.
So because you've talked about wanting to solve the Turing test for conversational AI, do you feel like you've solved that, or if you haven't, what are the key kind of technical barriers towards doing that?
It's a great, it's an interesting one because I think in customer support I think it's solved. I think in that setting which is very controlled, you know roughly what are the expected conversational patterns. I think it's great, you have the right knowledge, you have the right experience, it resolves query in the right way. As you shift to say sales conversation or dynamic conversation that we are having, then it gets a little bit harder. You need to be able to seamlessly of course interrupt, know when to pause, that's already expected. You need to account for some of the parts of if I'm speaking and if you just nod or add a noise, it doesn't interrupt automatically and it continues. And knowing all those patterns is essential. And I think here in certain of those use cases as we think about sales, that's a little bit still harder. But I think in the next six to twelve months, I would believe it's a big part of our research work as a research lab. I think in the next six to twelve months we'll see conversational AI Turing test be passed across all types of different use cases, whether that's the sales operations, some of the marketing potentially, all the way to fan engagement as well across some of the creative use cases that you mentioned as well, where we work closely with talent directly and with the creative industry on how they can engage their audience in a new way in partnership, which is of course a very different part of experience as well.
You mentioned safety earlier and I know you and the team think deeply about the guardrails you have to put in place and about how to avoid the misuse of your models. Can you talk, are you having to put in additional guardrails as the models advance, whether your own in-house models or other models that you're building on top of, and can you give us examples of how people have tried to misuse your models?
Of course, I think it's essential as a company leading in the space across that wider interaction patterns, take it as a responsibility that we need to define how people will interface with the technology around us. And there's a set of very core tenets to that. The first one is as you think about increasing amounts of content, increasing amount of conversations happening with AI, you always need to know you're interacting with AI and you need to be able to encode and watermark that the AI content is being generated, and we do that for all of our conversations so we can trace back exactly how this content was generated, when it was generated, who generated it, so we can take action. Then in general we have additional systems to catch, moderate on the text and voice level on is it going across in a different way. And then more broadly, and that's probably a pattern that all of us are seeing, it's not only the agents working on behalf of our businesses, it's also increasing threat of personal agents. I'm sure many of you here already have an agent, and then as we project the year ahead, we'll have increasing amounts of those agents too that might interact and interface with the business too. And in that third part, we are looking at ways to detect when you are interacting with an agent, whether it's on the text or voice level, building and collaborating with the others on the protocol level, and then deploying a lot of those models publicly. So every business, every government has ability to bring that so citizens and people can know automatically that they are interacting. On your misuse question, we had one example of working with a charity here in the UK which used it against misuse in a slightly interesting way, where you had human spammers calling, and I'm sure some of you also received some calls back in the day from human spammers, and you can detect those calls based on IP of where the caller is coming from, not reliably, but certain IPs are already flagged. You know that the spammer is calling in. So what they created with us was a voice agent that would, when the IP is detected of a caller, the voice agent is being served to that scammer, and then the mission of a voice agent is to waste their time, and then they recorded everything and they published that so everybody could watch those conversations. And well, in certain ways it was not exactly what was meant to do, but we had the longest conversation was over 50 minutes of a voice agent calling and wasting the time. So that was a great example of how you can use the technology to fight against it. But on a more serious note, of course, voice is such a core part of identity. So as we build the ecosystem, it's not only the safety part of how you deploy it, it's also safety and guardrails of how the voice agents interact, knowing that the customer interacts with the agent and having ability to have your voice saved in that realm. We have a voice marketplace where...
Do you think deep fakes become more of a problem or less of an issue as the guardrails are put in place, not necessarily just for ElevenLabs but across the ecosystem, the voice AI ecosystem?
So there's like two phases. I think now we're in the transition phase. In three to five years' time, it will be an expectation that you interact with AI. People will know that pattern, and then I think it will flip. So it will almost be the opposite where today you assume you're interacting with a human unless it's proven to be AI. I think it will be the opposite, which will be you're assuming you interact with AI unless it's proven to be human. That'll be encoding the coding on devices to make sure that you have that human verification step. Then there will be a second layer which is you are interacting with a verified AI conversation. So your personal agent will be known to be your personal agent and not a spoofing, and then by default everything else will be AI. So that I think will happen in the next three to five years. In the transition phase, it will depend on the maturity of the given public and commercial sector. Some places we, I think in UK we can lead in a lot of the applied AI parts of the stack. We happily signed a memorandum of understanding with the UK government and building that framework for a lot of the safety across the wider set of models as well as providing additional accessibility for a lot of the government services through voice and conversation in an easier way and enabling people and upskilling people with that technology too. But one of those core parts is like making it easier for people to know when to expect AI conversations. So in certain parts of the world, I think we will harness that and prevent the misuse from wider open source models which will not have those safeguards. And then I'm looking how we can collaborate with the longer tail of countries that are maybe not as advanced in some of the developments.
Can I bring it back to the business model? I think your revenue run rate was about $6 million in 2022. It's now $600 million. You've been asked this question before, but things have changed. The commoditization of the business and how you think about the risks around that. There's a company that caused a bit of a storm in the US, I think SF, Fish AI, and they do open source, open weight, very cheap voice AI solutions. Are they a risk? Is that a risk? How are you thinking about ensuring that you continue to have that moat and it doesn't get commoditized?
So for us it always was an approach of combining three things together. So building great models, frontier models, happily the best models today. I have a brilliant co-founder and we have some of the best researchers in the world working on voice and audio. Then two, building products that are optimized for specific problems. So really going deeply in what does it take to solve, let's take the customer support case. You know, one thing is models, but as it's moving from a building model to deploying that in production is a different beast. And here you need to be able to bring your knowledge base, you need to be able to configure the right integrations both internally whether that's with the CRM or with the telephony systems. You need the right testing monitoring suite to be able to work for that. And here we invested a lot of time to build a great framework for helping companies build conversational agents. So that product layer is essential. And then the third one is a combination of ecosystem and deployment. How can you now have the wider set of ecosystem where you don't have to start the work from scratch? You have a trusted set of systems that already exist. In our case, if you are starting your work with agents, you already benefit from thousands of simulation tests and guardrails that have been created before. So it's a lot easier to start from that. Similarly with voices, we have thousands of different personas and voices that people can use depending on types of conversation and match that with the deployment. So people that can work, I know later we'll hear more about forward-deployed engineers as well. I used to be a forward-deployed engineer at Palantir too, and back in the day you always need to explain what it means, but now of course everybody has a lot of that. But that deployment side and what they represent of being able to work side by side with the businesses is essential. So combining that part of research, product, ecosystem and deployment from our perspective is essential. So as models get closer in capabilities, we still know that those two other layers, product and the ecosystem, remain essential. Still we hopefully will continue leading with models. So it's a combination of those that would be amazing, but ultimately we believe in you need all ingredients to make something truly successful.
And you are, as I mentioned at the top, you are a global business built of course out of Europe. The US presumably is a significant market for you.
It is. I have, it's the midterms. It's the elections today. So a lot of US, most of our US colleagues will be focused on and US views will be focused on what's happening there, of course politically.
I just want to, given the market importance of the US, any concerns that if we get democratic control of the House and the Senate that there'll be, that this will be an administration or at least a government or at least there will be some push back, further push back to the AI story and the AI buildout in the US? Are you concerned at all about additional friction, political friction?
So as you said, we started here in London, we have our HQ here in London with 200 people across our business. On the US side, you know, it's less of a concern given a lot of our commercial clients are realizing how much value that a lot of those deployments can produce and that hasn't changed. If anything, that has accelerated. We need a lot of people still to just be able to serve the customers that we are working with. This is probably the wrong setting to say it, but we are hiring, so if anybody is keen to join us, we would love help. But the US elections of course, you know, you have those two parts, the US elections of course, Anthropic's one from a couple of weeks back, and there's this I think a super interesting statistic of the customers that they have that spend over $100,000 or more, a lot of them, 50% of them are Fortune 500 companies that spent that they have, and it ranges from that all the way to eight and nine figures contracts. That total spend represents on average less than 0.1% of the expenses those Fortune 500 companies have spent in the last quarter. 0.1% of adoption. The Fortune 500s, it's closer to 1% but still below 1%. If we use that as a proxy of the adoption of how much value there could be derived, is just being started and we see exactly the same work. It's, you know, we spoke about Revolut and Trainline, but whether it's the companies here like Admiral, internationally like Stripe, Deliveroo, DoorDash, all of those companies are just at the beginning of being able to serve the customers better and delight them in a way that we would love to work together.
You mentioned Anthropic, so I have to quickly get this one in. The Anthropic IPO expected of course, and they filed the S-1. Do you see that as an important litmus test for AI companies, how that is received in the public markets?
100%. It's going to be very interesting to watch. They have a good combination of some of big bets and of course some of their own challenges, but also are proving that you can build a frontier intelligence and combine it with a set of great applications and a set of good things we will hopefully follow and one day get to the same scale.
Yes. Yes. Well, that's led me on to, well, we've run out of time. So maybe I've run out of time to ask you the question you always get asked, which the IPO, but that time frame around 2028, that's the speculation. That's a fair time frame, is it, to think about an ElevenLabs IPO?
It's, you know, like for us the goal is to build incredible impact. IPO of course is not in itself, it's a step in the journey. Would love to IPO, build a public company, help build that step in a longer journey. We'll prepare whatever we can to be ready and run the company as we would run a public company in the next year. 2028 and having that foundation set is a good goal.
Fantastic. Mati, really appreciate your time and kicking things off for us. Mati Staniszewski, of course, co-founder and CEO of ElevenLabs. Thank you.
Thank you.
Thank you. Thank you.
Please welcome to the stage Louis Mosley, executive vice president for the United Kingdom and Europe at Palantir for a conversation with Bloomberg's Tom Giles.
Thanks for being with us today. We're here on an auspicious day for many reasons. Just last night, Palantir reported quarterly earnings and reached a significant milestone. We were discussing backstage, and that is business revenue, excuse me, yeah, business enterprise revenue surpassing government revenue in the US for the first time. What drove that?
Yeah, it's a huge milestone because I think those of you who've been following Palantir for a few years will assume that we are a government business and our largest customer is the US government. Well, we had our quarterly earnings last night and for the first time our revenue from US corporates overtook our revenue from US government, and so I think it's a very significant milestone and it's an indicator of the direction of travel for the company.
Well, I'd like you to say more about that, particularly on a day when, you know, in a few hours polls open in the US, we'll have the midterm elections. There is widespread anticipation that Democrats will retake the House, possibly the Senate, and that would have ramifications for defense spending. What is your, what kinds of scenarios are you thinking through regarding the potential outcomes? And more to the point, potential changes in defense spending that for many years have been advantageous for Palantir.
Palantir has been around for 20 years. So, we've worked with the whole range of Democrat, Republican administrations, Democrat, Republican houses of Congress. And I don't see that changing. I think, you know, the US's commitment to its defense establishment, to its responsibilities around the world remains very strong. Palantir is a key part of that and I can't see that changing.
As fast as your growth is, there is today in the market, realizing that markets can be fickle, but at the same time they are looking at your projected growth rate, looking at that being, you know, I think 15%, translating a 15% increase, a slowdown from 20%. Fast growth companies like Palantir, nobody wants to see a projection for revenue growth that is going to be slower than the last quarter. Why are you being more conservative?
Well, I think it's hard to be conservative when, you know, we yesterday we announced growth rates in our US commercial business of 148%. I think that's unprecedented in public markets for a software company. We're doing that at the same time as, you know, adjusted net margins of 60 plus percent. So our rule of 40 is way, way up there. You know, if you look at that quadrant, there is no other software company on public markets that is anywhere close to that. So it's an unprecedented situation and I think I have some sympathy for the analysts. You know, it's going to be very hard to value Palantir because their spreadsheets, their models don't allow for something that's happening that fast on that kind of scale.
Globally, as you look at where revenue comes from for Palantir, US obviously very important player, where UK has come down from about 20% in 2021 to half of that last year. Rest of the world also came down 2024, 2025. There's a lot of emphasis, and we're talking about it today, on tech sovereignty. As you see regions like Europe emphasize we need to build it ourselves, we need to build it in our own backyard, we're less inclined to be reliant on the US or a US-based company. What impact is the emphasis on tech sovereignty having on your business?
It's generating a lot of questions. A lot of my life is spent talking to customers about how Palantir can guarantee them their sovereignty, how they can have control over their tech stack, how they can have control over Palantir, and most crucially how they can control the intelligence layer in their stack. So this is the model there, the large language models. But actually I think that the numbers that you mentioned, that divergence between the size of our US business and our business around the rest of the world, is not about concerns over sovereignty. It's, I'm afraid, something that actually does keep me up at night, which is you're just not seeing the same rate of AI adoption in Europe specifically compared to the US. So that bifurcation in Palantir's revenue growth is actually about how fast the US is growing rather than Europe stalling or stopping or falling. It's the relative growth rate that's important. And I think you could think of Palantir's revenue growth as in some ways the purest proxy in public markets for the rate of AI adoption in an economy. What we're seeing in the US, that 148% year-on-year revenue growth in corporates, is blue-chip American companies adopting AI and deriving value from it.
Go to the rest of the world, you're not seeing that growth rate. So, what does that tell you? Blue chips on this side of the Atlantic are not yet adopting AI. And that worries me because we're starting to see extraordinary results from those customers, those big American blue chips. And if their peer competitors on the other side of the Atlantic don't start to catch up and enter that race, fast forward four or five years, there's going to be a huge divergence and they're no longer going to be competitive.
And yet, you do see an emphasis on homegrown alternatives here, even former Palantir executives going and founding their own defense tech companies. You have competition from Helsing and you have, you know, across the region. In addition to that, you have large language models that are seen as potential competitors to the work that you do, will teach you how to analyze your data, you don't need to rely on Palantir. There is a rise in competition and that is also surely having an impact.
I mean, welcome the competition, and I think if I just start with the large language model providers, the frontier labs, they're creating a market for us. They're teaching the world how powerful AI is. But you can't capture the value of that technology without diffusion. And Palantir does the diffusion. You go to Anthropic's S-1, the document they published as their prospectus before going public. The big case study in there is the use of Anthropic's model Claude by AIG to automate underwriting and various bits of claims processing and insurance. How was that done? Through Palantir. So they are making a market for us. And I think the actual direction of travel, and you know, we're seeing this in the US. Again, I don't think Europe has caught up to this at all, is a growing resistance in corporate America to being dependent on an emerging duopoly in the intelligence layer. That's to say Anthropic or OpenAI, that they are handing over their data. They're giving these model companies access to the inner workings of their business as they diffuse the model. The model is then learning the tacit knowledge, the process knowledge that makes these businesses what they are. And it's incredibly expensive. Where the frontier now is, take an open weights model like Nvidia's, train it on your own data with the traces of your own decisions, and it turns out it is better performing than the generic frontier labs' latest model on a specific task. So what that tells you is the future is for these big blue chip companies, and indeed maybe for governments, is much more task-specific AI which takes open weights and trains it with, you know, in the jargon, the out-of-distribution data. So that's to say the data that only that organization has. It's its own understanding of its processes. Take the open weights, adjust, fine-tune, and you have a better model than what you can get from the frontier labs at the frontier. And of course the cost at that point is essentially the cost of compute. So there's a huge difference in how much this is costing you, not just on performance. So you're getting better performance at vastly lower cost. And I think that's actually the future.
Technology aside, Palantir has a perception problem. It is associated with the US government that has demonstrated, has wavered on support and reliability as an ally to Europe. That's public. That's not just a question of, that's not just subjective. Palantir is criticized for uses by ICE, uses by Israel there. What does Palantir need to do to address that side of the equation? You can talk about your technology and how sophisticated it is and how much better it may or may not be than the frontier models, but there's still the perception problem. What do you do about that?
I think that the perception problem, as you describe it, is a consequence of Palantir being a company that believes certain things and is proud to believe those things and isn't ashamed to say those things. We from our inception, we built the company to defend the West. We believe in the West. We believe in liberal democracy, the rule of law, the values that make up the West broadly defined. I'm not talking about a geographic construct here. I'm talking about a values-based one. Alongside that, we therefore will support the democratically elected leaders of those countries. I may personally disagree with them, but we think it's our duty to help them deliver the mandate that they have been given at the ballot box. It means we support Israel and we're proud to say that and so on. And I think there are many people unfortunately who disagree with us. And when I say unfortunately, I mean they don't just disagree with us. They would rather we didn't exist if they disagree with us. We believe in free speech. We believe in living in a world where people can have divergent views and debate these things, and therefore we stay in business and we make our case and you can agree to disagree with us, but that's I think the source of the controversy.
We're in London, and I want to talk specifically about your relationship with UK government and your contract that's at stake, right, with the National Health Service, 330 million pounds. Recently the UK Science, Innovation and Technology Committee put out a report about digital services in the public sector recommending breaking the contract. As the chair of the committee put it, Palantir represents an unacceptable point of weakness in our digital infrastructure. Where are you in terms of your discussions, relationship with the prime minister, and what's the state of where you are with this important contract?
Well, we have a very important business in the UK. The NHS contract that you mentioned is probably the most prominent of those contracts. It's a very important one to roll out our software to every hospital in England. Help them to integrate their data, help them to deploy AI. We've been doing it for roughly two and a half years. It's a seven-year contract that was structured as a three plus two plus two. And we're
coming up to therefore the end of that first three-year period. And the government has a decision to make about whether it wants to extend for the following two years of the contract. We don't know what the government is going to decide. We're waiting to hear. You can imagine what I hope they're going to decide. I think we have a very strong case for why that would be the right decision to make. There is an organization called NISTA which evaluates large government infrastructure projects.
That is one of the only green-rated technology programs that government has. Full stop. Out of 217 government programs that are evaluated, I think only 27 or something are green-rated. Ours is one of those. So it's highly performing. It's doing what it was intended to do. In fact, it's exceeding many of those objectives. But we will wait and find out. It's up to the politicians to decide.
But surely it's not just a matter of, like, you know, twiddling your thumbs and waiting for the government to make a decision. Are there concessions that Palantir needs to make? For example, is there a universe in which Palantir might, like, separate out its UK operations into a separate entity?
Well, a lot of that is actually, if you're referring in the sense of the sovereign sense, it's because actually the original business we had here was a defense one. And as you may know, the requirements on any foreign company providing essential services into defense in Britain, as is in fact the case in really every country including the US, there are all kinds of requirements that are made on those companies to make sure that they onshore key aspects of that capability and that the customer, in this case the Ministry of Defense, has operational control over the capability that they're providing. So a lot of that is in place.
I don't think the government should have any concerns about operational sovereignty over Palantir, and I very much hope that, you know, this is what's in the interests of patients in this country. The software program for the NHS has helped 117,000 patients get operations in the hospitals that are using it over the period that might not otherwise have happened. So it's a highly performing contract.
So there's no need for, like, a separate domicile, for example.
Well, we have a UK entity that, as I said, is subject to all kinds of controls. The NHS has very specific prescriptions around data residency, for example. So none of the NHS data that they process in Palantir software ever leaves UK soil. So many of these things are in place.
Okay. The US is using Maven Smart System, Palantir's Maven Smart System, in its war with Iran. Early in the war, there was a US strike that hit a school, killing more than 120 children. Your colleague Mike Gallagher said that had Palantir been in charge of the underlying data set, we could have avoided that tragedy.
And yet, we've reported, according to people familiar with the matter, that since then, US CENTCOM has made key changes to the way it targets, its targeting process, including a lot of upgrades to Maven. So could you help us understand how it is that Maven, excuse me, that, yeah, Maven Smart System is part of what the US, part of the underlying data that the US is using in Iran, and yet shouldn't be responsible for that tragic strike that killed children?
So, you'll appreciate obviously as a British national, I don't have a US clearance and so I will not obviously have the details of this dreadful case. But I can assure you that Palantir is not responsible for the tragedy and Palantir is not responsible for the underlying data quality that then eventually will feed into Maven Smart Systems and elsewhere where decisions are taken.
So it is that distinction between the data source and then the software that you're using to process the data, and if there are errors or inaccuracies in the data that feed it, it is the classic garbage in, garbage out problem.
So from your vantage point though, what should change about the ways that the US is relying on Maven Smart System to avoid issues like this in the future and also to avoid the concern or the perception that we should hold Palantir responsible?
Well, these are always very tricky cases because they exist in a classified realm and so none of us have really the full details and all I can give you is that abstract distinction and assure you that Palantir was not responsible for that tragedy.
Your CEO Alex Karp was, as I understand it, the first Western CEO to go to Kyiv after Russia's invasion in 2022. Palantir's been a major supplier to the Ukrainian war effort. I want to talk about kind of the perception of the war in Ukraine from here, from local. There's a shift underway in certain parts of the US and a growing sense for some in some circles that a Russian victory is inevitable.
And in fact that's something that has been espoused by a very major US executive, Elon Musk, who's also doing business, also providing Starlink to Ukraine. I want to get a sense from you about kind of how you, the propriety of an executive like that taking a public stance and even making decisions like shutting off, restricting Starlink, that has a real impact on the war effort in Ukraine.
Well, I think first thing I would just emphasize is the point you made earlier around the perception in Europe that the US is maybe an unreliable ally, that its support for Europe, for Ukraine, has been wavering. We, Palantir, have been supporting Ukraine from the very beginning of the war and our support has been steadfast.
Notwithstanding, as you say, in the US it is a much more complicated position to take than in Europe where most people are very supportive of Ukraine. In America, as you know, it's a much more mixed picture. And, you know, we might be controversial in Europe for supporting Israel, but actually in the US typically we're controversial for supporting Ukraine.
But my point earlier, we believe things, we stand behind them. And we also take the view that when we partner with a country or an organization, it is for the leadership of that country and that organization to decide how and where they use the software. It wouldn't be right for unelected technology executives to be making those kinds of, if you like, micro decisions about whether their technology can be used in this or that military operation.
If once you're in that relationship with a customer, you do have to trust the customer. And if there is that values alignment, you believe in the same things and you support them, then you've got to stand by them. And that has been the principle on which we have always approached these things.
In the last few seconds, I want to talk about NATO's adoption of Maven Smart System. Talk about what that says about the possibilities for Palantir contracts with other individual NATO member governments.
So we, yes as you said, we now are the operating system at the, if you like, the supranational NATO level. So this is not the individual countries within NATO, but it's NATO itself, and that, if you like, is the binding layer that brings together the alliance. And if the alliance therefore were to fight tonight, it would be with Palantir.
But of course, what we need is each of the member states to get to a point where their own domestic or sovereign capability is at least as good as that. It doesn't have to be Palantir, but it needs to be as good as Palantir. And so we're now seeing the rapid adoption of our software across Europe by ministries of defense. Obviously, the US and the UK were the two countries that went first in the alliance, but we're now starting to see new customers emerge.
That's individual apart from the NATO. Exactly. How many have you been able to achieve thus far?
So far, Sweden and the Netherlands are both customers, and there's a number more in the pipeline. For example, we signed a letter of intent recently with the Polish government, and there are others I could mention, but you could think of these as the sovereign versions of Palantir for each of these countries that are completely distinct from any other version of Palantir that they own and operate themselves independently.
Which they can then configure to their own needs, but will allow them to interoperate and collaborate seamlessly with their allies in the event that there were a war.
So, we should expect to see more of those individual.
All right. We'll end it on that note. Louie, thank you for being here. Appreciate it.
My pleasure. Thank you. [applause] Thank you so much.
Please welcome to the stage Suzanne Ashman, managing partner at Sovereign AI Fund, and Johan Buckfist, Chief Financial Officer at Mistral, for a conversation with Bloomberg's Mark Bergen.
Thank you. [music] [applause] [music]
Good morning. Sovereign AI, one of the favorite subjects here in London and across Europe. Yan, I'll start with you. Your company, your CEO, Arthur Mensch, a couple weeks ago put out a brief manifesto of sorts, 'build or surrender' and the idea concept. And I think the salient fact was that European companies pay more than 260 billion euros every year to American cloud and software providers.
And that was underscoring Mistral's argument that that needs to change. Tell me about the decision process for that post, what it means for you and your position as CFO.
Got it. First of all, I'm very glad that he came out with it. I mean, time is of an essence right now. We're going through this transformational shift and it's probably bigger and happening faster than any other transformational shift in the past. I think from a European perspective in the industrial revolution, we were pretty on top of. In internet one, I think we were maybe not as successful in leveraging that in full.
I mean, 80% of the spending we do on software today goes outside of Europe, for example. And I think now we have the opportunity and the potential to leverage this next kind of transitional phase, and we are trailing behind, but that doesn't mean the race is lost yet, but we do need to take some pretty drastic and quick decisions to catch up.
How far behind? Like, is there a timestamp? How many months feel like on the frontier?
I mean, to be fair, Mistral is a three-year-old company. I think Anthropic is around five and OpenAI is 11 or so years old. So in that aspect, I think we're actually tracking pretty well to some of those metrics. But what we do need is we need more compute to train our models.
And ultimately, the main difference between the models we are releasing, maybe around three months behind the real frontier models, is the amount of compute that we have available. We talked about, we trained our last model, the Large 4, on about 3,800 GPUs on the Blackwell GB200s, and our friends in the industry are training on many multiples of that.
So, in that sense, I think we've done really, really well coming out with a model which is a frontier model in the open source community and especially strong on the compute and on the cyber capabilities.
Well, we'll come back to that compute and chip question. Suzanne, what do you think? Does the UK need a build or surrender mantra? What does that look like here?
Yeah, I think in terms of the UK and Europe in general, we feel that it's important to be at the frontier in certain areas of the AI stack, and I use the word frontier broadly. So not just, you know, general purpose large language models, but also like at the frontier of chips or at the frontier of drug discovery.
And at the Sovereign AI Fund, we tend to think of tech sovereignty not about like pure independence, because no country, not even the US, is purely independent. They are as reliant on ASML as we all are. But it's about having interdependence so that other nations are as reliant on a UK or European tech as we currently are on them.
You mentioned chips. You've done a lot of chip investments, some of which you probably can't talk about or won't talk about. And tell us why chips? It seemed like that was sort of leading, and the AI ministers talked about that as the strategy going forward.
You know, we think about different layers of the stack where Europe and the UK in particular can compete, and chips felt like a really important one for us. And in a large part, it's because we have a talent base here, the talent base coming out of ARM, out of Graphcore.
And you know, while as a fund, as an investor, it can be easy to point to themes, areas of a stack that we really like to invest in, like everyone else, ultimately you can only invest, and companies will only be built into global champions if they have fantastic founders with like relevant technical expertise.
And we look across the UK ecosystem and think, you know, chips is one of those categories where there's a talent density that we think could be used to build global champions.
This is mostly around inference, right? Like that's the bet here because what, the training, the chips for training is basically a like lost game at this point.
Yeah, I think the UK doesn't have a competitive player on the training side, and I think the infrastructure and particularly energy costs make that incredibly problematic anyway. But in the next race, whether that's inference and beyond, like it is plausible that we have a better set of conditions to be able to compete.
Yeah. Johan, when you're looking at Mistral, I mean, I don't, maybe you can tell me how much of your compute budget goes towards chips right now.
We spend a lot of our money on building GPUs or buying GPUs and kind of build out our infrastructure primarily for our own use cases of training models, but also to serve our customers.
How closely are you looking at some of these sort of, we can't call them rivals, but like maybe potential challengers to Nvidia?
I mean, it's common sense to assess all the players out there to make sure that we always take the right decisions for the company. Nvidia is a very close ally to us and also an investor, as we have disclosed previously as well. So we have a very good collaboration with them and also with the Nemotron kind of collaboration now where we're building the next Mistral Large 5 on their compute.
And we do that in collaboration with Nvidia. Something we look forward to.
Do you think there's a day in which you'll use a UK or European made design chip for inference at some point?
I think Europe needs one.
For sure. The hardware stack is equally important to have in a sovereign perspective. So very much in favor that you continue to invest in that.
And just on the chip side, we're at a 620 million pound fund. That's relatively large in terms of...
New number as of last week, right?
New number as of last week, up from the 500. You know, relatively small in terms of like budgets in this category, and we can invest, but actually probably our biggest needle mover is being a customer.
And so take the chip space. There was a 180 million pound advanced market commitment to chip companies, and that essentially says we, government, will purchase your chips if you meet these technical milestones. And I think historically in the past, and Mistral's a good case study of this, like in the UK and Europe we have been quite squeamish about picking winners in a way that they haven't been in the US.
You know, you take SpaceX and they have really been the focus of a lot of government contracts and support, and in Europe we haven't tended to do that. And I think in Germany they've shown in the defense sector whether it's Helsing or Stark, in France now with Mistral, that if governments are serious about having winners and global champions from their country, you have to put like real capital and support behind them.
And I think on the chip side, the 180 million, we don't expect it to be split between 100 companies. We really want it focused on the handful of UK-based chip companies that we think could be the global champions.
Yeah, I think, I mean, I've heard the phrase that Macron is probably Mistral's best salesman. I don't know if you agree with that, but France has picked a champion, and to the extent that that's certainly helped you, I don't know, do you think that's a hindrance sometimes? Do you think you worry that maybe you kind of get pigeonholed as a French AI company or even just a European AI company as opposed to a global one?
I mean, not at all. I think what France has done really well is to really understand the importance of Europe having an AI champion, and they've been very helpful in spreading that message. I do think that there are more European nations now kind of really picking up that narrative themselves as well and understand the importance of getting this right.
And that it's a kind of an existential threat to Europe not to get it right, both on the technology side, but also on the economic side as well as the security side, to be honest. And you do need to have your own capabilities in this space, and Mistral is the leading European champion in that, so it kind of makes sense that you're more of a European AI champion than a French AI champion.
I should have said at the onset, apologies. We are taking audience questions here for the first panel, and so please feel free to submit them. We have a really fascinating one which I think that we haven't really addressed China yet. But obviously, like a lot of you, recently released a large open-weight model, and it's kind of positioned it as the West and Europe needs an alternative to China.
You're obviously racing against a lot of companies, like we can just, you know, DeepSeek is obviously one if not many, competing these models that are coming out at a rapid clip. Is there a trade-off between keeping up with the Chinese on open source and safety?
No, I don't think so. We take safety very, very serious, and the last model that we released with the cyber capabilities was very potent. So we did make sure that the right parties get kind of access to it, and we also limited some of that capabilities in our public release as well. And I think safety is one of our, if not the most important aspect here when we release new models.
Yeah. Suzanne, what do you think about that? And then also the second part of that is, you know, are there concerns in UK government and UK companies about using Chinese open source models, and you think there should be?
I think the sort of safety has become a sort of more problematic term, and I think if Europe overindexes on trust, probably, and the focus on trust in these models, and when I look...
You mean Brussels, basically.
Yeah, yeah, and also the UK with their focus on their AI Security Institute. I mean, someone's written a question like, is 100 million enough to compete? Fortunately, it's not 100, it's 650, but still, like, it's relatively small in terms of the global capital base. But the AI Security Institute in the UK is like the best funded of its equivalents globally. So I think...
we in Europe take trust and security probably relatively more seriously, and I think that's really helpful when you start to operate in regulated spaces where when we are looking at do we go for a Chinese open source model or a European one, you know, cost and performance is part of that delta, but also like trust and security.
And so people say, oh, will you use a worse European or UK-based product? And the answer is no, because trust, security, sovereignty should be taken into account as like part of the equation on which providers you go for.
But you think there's a risk of overindexing on that?
I don't, because I think we're fortunate that players like Mistral are so close to the frontier of open weights models that for the absolute vast majority of even like complex government use cases, it's a very effective solution.
Go back to the build or surrender, the building the frontier models, that is a costly enterprise, and we can look at those figures from some peak, from Anthropic's S-1, I assume you read them when we all did. What was your take with your CFO hat on?
I mean, with the CFO hat on, I think it's an absolutely remarkable and so impressive of what they've been able to build in that short period of time. Of course, as anyone else, you kind of react on the liabilities on the balance sheet that they have, and that also put a lot of requirements on their continuous growth, but ultimately you can really see the value of this kind of product in the market by just looking at their revenue growth.
Yeah. Well, one thing that we flagged is the customer concentration. I think it was close to 80% of their in the first nine months coming from five customers. Is that a concern at Mistral?
It's not a concern from us because we sell differently. We sell bilaterally to our customers, so more of a sales-led motion like Palantir. At the same time, we are a frontier lab, so we're somewhere in between Anthropic and the Palantir model.
But there are trade-offs, right? Like the Palantir model is much more cost efficient, I would assume, than the building at the frontier, building, training these models.
Yes, for training models it comes with a cost, but it also comes with a lot of benefits because you can customize that model for specific use cases at the customer, which especially in an open source environment allow the customers to be in control over the kind of sovereign cloud that the deploy is on, and they also can control the data, where the data is shared, how the kind of models are being trained on their data and their kind of use cases and the customer data as well.
So I think it's clear to us at least that customers are favoring more and more the open source path because it gives more control to the actual companies.
Is that customers, can you say how much of your customers right now are outside of France?
Oh, the absolute majority is outside of France. Yeah.
As of since I joined six months ago.
Okay. For sure. And is there a goal to get that number at a higher rate?
I think the goal is to be able to offer pretty much any large enterprise and institutions as well, governments around the globe, our services because it makes a lot of sense.
Then it will always be easier to sell our services in Europe and maybe in general outside of the US and China because they have their own champions, but in general we see strong traction and growth everywhere.
Suzanne, I'm going to pick on one of your investments in Ineffable Intelligence, which is, I don't know if you like the term Neolab, but David Silver, researcher, and what I think it was, what, the billion dollar seed round, right? Like I assume he's probably out there raising again, like that sort of building those next frontier beyond large language models.
How do you see that with your VC hat on? Like is this a winner-take-all market? Like the expectation here that Ineffable is going to be like the next Anthropic, or are you kind of placing as many bets as possible?
Yeah, I think, you know, Dave Silver is one of the leading global researchers. So when you see that round and he's raising a billion at a multi-billion dollar valuation, as an investor, I still think that given the strong weight you put on a founder, that is a bet that you are happy to take, and we are delighted that he's continued to build in the UK.
I think the DeepMind ecosystem has meant that he is one of a number of researchers that previously worked at DeepMind who are spinning out to launch new...
Okay. How many have you invested in? Do you know?
Okay. A handful of companies founded by DeepMind alumni.
I think it is, I mean it is also public that we've invested in Isomorphic, like Demis's new drug discovery business.
Would you invest in his next business if he launched one?
Is he planning to launch one?
Not that I have heard of. [laughter]
So, Neolabs in general, of course, we would love to have more like frontier type companies in the UK, and Dave Silver and others we hope will be the people that build them.
Just to follow up on a sort of UK point on the Anthropic S-1. One of my concerns for the UK ecosystem and for the European ecosystem in general is that on the retail investor side in the UK, a tiny proportion of households have exposure to public stocks, right? And that is fundamentally different in the US where like the majority of households have exposure. Now every kid's going to have a SpaceX share apparently.
And so the challenge for the UK and for Europe is like with this wave of blockbuster IPOs and the wealth creation that happens around them, is the UK effectively like sitting that out because as a country people largely have their savings in cash and historic products like cash ISAs, etc.
And we do need to find ways to, I think, make individuals and consumers in this country like feel the benefits of AI and feel the benefits of the wealth creation.
What's a method? I mean, that Anthropic is doing the sort of 5% public benefit corporation? I think that's just in the US though, but is there a similar model that you think that France or, like you're from Sweden, that other European countries could take?
Yeah, I think there is. But I think that sentiment around AI is like mixed, and I think the larger companies in the space need to find ways to like bring the public with them and to more clearly communicate like the benefits of AI.
And I think the industry as a whole like really misses that Jamie Dimon figure that, you know, finance has in the US, someone who can communicate really clearly as the expert. And I think if that was a European AI voice, that would be incredibly helpful.
Would Mistral do something similar, like a public benefit corporation?
I don't know. It might be on the table.
What you mentioned earlier talking about Europe sort of missing out on the internet, you were at Spotify, with arguably one of the biggest winners in Europe on the internet, and you were there when they went public. What lessons, what are the things that you took away from that, and then what are you avoiding as you're sort of thinking about Mistral's future?
Well, first of all, we were pretty alone in Europe back then. It felt at least that we were very alone. So most of our peers was from the US. Now that landscape is much different. We went public in a more complicated way at the time at least because we had never done before, which was a direct listing.
The reason why we wanted to do it was many. First of all, we didn't need to raise capital. We were well capitalized already. And we wanted, or Daniel especially also wanted to go public in a way that made sense for him and for the company, and going public in the direct listing was actually a more democratic way of doing it. You are able to speak to the entire market. We had a capital markets day before we went public.
And we allowed anyone to buy shares at the opening instead of directing those kind of shares to specific investors, and there were no lockups and anything like that as well. So the transparency and the access to anyone was much greater, and after that of course we had Slack and Palantir and Asana and many others that kind of came afterwards.
The audience is going to ask the question I was going to ask anyway: when's the right time for Mistral to go public?
Good question. We don't have any plans to go public at the moment. Of course, we have plans to go public at some point, but any details of those plans has not been concluded yet.
Sure. Where is the right place for Mistral to go public?
I think that's also a question for later, but wherever it makes sense for the company, its share base and our customer base.
Yeah, we talked about chips at the onset in the sovereign. We didn't talk about the other kind of big part of the stack, the cloud and hyperscalers. Do you think about, I mean there's not a, tell me if I'm wrong there, obviously there's some of these neoclouds which you haven't invested in any neoclouds. Is that because, yeah, why not? Have you, I'm sure you looked at them.
I think arguably, I mean there's sort of one UK-based neocloud in Nscale, and you know, our goal as a fund is to invest inception stage through sort of Series B, and by the time we were founded in April, you know, Nscale was way beyond our...
Would you invest in them if you were...
Yeah, I think we, I mean, compute is clearly an incredibly important part of the jigsaw, and I think 15% of global compute is in Europe, and clearly that is not enough. So yes, I mean, it's a category we'd be interested.
Yeah. And Mistral Compute is, I mean, you're kind of going, you're building that sort of neocloud yourself, right? Like, tell me, I know we have limited time, but just quickly, like, how much is that something that you still see Mistral doing in the next five years?
We're going to continue to build it, and I think we said 1 gigawatt by 2030 or something, and that still holds true. The reason why we're building it out is actually more strategic than maybe for neoclouds, because we need that compute ourselves, right? It's a way for us to lower our cost for our training, and also we need it to serve our customer base.
So it's actually something that's strategically important for us to continue to build out, and also we see it's extremely important for Europe to have AI infrastructure that is run and has a full stack sovereignty of Europe.
There'll be more Mistral data centers in Europe in the next few years?
Oh, that I can say definitely. Yes.
Okay, great. With that, thank you both. Thanks so much. Appreciate it.
Thank you. [applause]
Please welcome to the stage Fabian Hedén, co-founder and chief technology officer at Lovable, for a conversation with Bloomberg's Shireen Ghafari.
Fabian, thanks so much for being here with us today.
Thanks for having me.
Yeah, excited for this conversation. So, for those who don't know, you are the co-founder and CTO of Lovable, one of the most popular vibe coding apps out on the market today. You started Lovable as an open-source project in 2023, in the wake of ChatGPT, as an open-source project called GPT Engineer. Can you tell us a little bit about the origin story, how you were inspired to start Lovable, and how did you decide to make it a standalone coding app?
Yeah. So first off, I have to give the credit for the open source project to my co-founder Anton. He's the one that kind of put it out there. But I'll give the quick origin here then. So me and my co-founder Anton, we used to work together at a YC startup in Stockholm. And then I quit after 10 months and wanted to do something more ambitious, I suppose.
And one day Anton came knocking on the door and he asked me, have you heard about this GPT Engineer, which is what you're referring to? And I told him, like, no, should I? And from there we got to talking about what you can actually use with this technology, because GPT Engineer back then, which was kind of the precursor to what Lovable is today, it was, I think, one of the first tools to show that you can build software end to end with AI or large language models.
And from there we thought, okay, if this technology is going to continue evolving, which it has, can you actually do something a bit more ambitious than just build these tools for developers, which is what most people were doing at the time? Like autocomplete within an editor for developers, for example, was a very popular use case.
But we thought, no, okay, let's make something for the 99%, for most of the people in the world that have not spent years learning how to code, and we think actually that's where you can get the most value out of this.
So you now, the last reported public figure is that you have some 600 million in annualized revenue. What is the biggest driver of that growth? What is the business model of Lovable?
Well, the biggest driver of our growth has always been very product-centric. So we call it the wow factor, where someone comes to the product and they experience for the first time within, you know, 40 seconds or a few minutes, a first generation of actually building something with Lovable.
So that can be like a simple website, a landing page, it can be more complex kind of SaaS platforms or internal toolings for your company, you can build apps and all of these different things. And the first time you do this, a lot of people experience this wow factor. And what happens at that point is that they go and tell others, and that is really the basis for our growth.
Is there a certain kind of user or company that you are seeing really take up Lovable's apps?
Well, it's very split. So if you look at segments, we have everywhere from kind of solo founders, which I think it's this new movement of solo builders, solo founders who are creating things where they before wouldn't have had the resources to create what they can do today.
So that's a whole new kind of market that has opened up, and then up to kind of the SMBs, to mid-market, up to the enterprise, we have two-thirds of the Fortune 500 as well on our platform.
And out of those two-thirds of the Fortune 500 companies using it, is the average user sort of a very motivated person within the org who is getting their own Lovable subscription, expensing it, or are these corporate, you know, top-down programs?
Yeah, so it's a mix. Our main motion is really a bottoms-up motion where users in these organizations are finding us organically again from this kind of word-of-mouth motion that we have, and they're often people who have a problem and they're kind of distributed out in the organization.
And, you know, the usual way that we would be able to build the software that they need is to wait for engineering to serve them, which, you know, either takes a long time or it doesn't happen at all because it's hard for them to prioritize all of these needs. And so with Lovable, these people are empowered to build something themselves.
So in a lot of ways, I see similarities with our founders on the platform who are very empowered to create new businesses. We see a lot of people within these enterprises empowered to really do their work end to end.
So when OpenAI, Anthropic, Google, you know, are all putting out AI products that can generate apps and all trying to get not just these, you know, corporate accounts but everyday people and users making their own coding products and coding apps with AI, what is Lovable's moat?
Yeah, so since we started, there's always been competitors, and there's been competitors on kind of the startup side. Some launched earlier than us and were kind of the underdog, kind of just trying to execute and out-execute these people, and now we've gotten to a place where I think we've out-executed quite well.
And so I think that's, you know, that's always a core reason to success, execution rather than kind of thinking about moats. As we've done this, our users have really taken us down a path where we're supporting the full lifecycle of what they're doing.
So you can think of if you come to Lovable, you start building something, and after you build that, you want to run it, you want to host it, and after that you want to get some outcomes of what you've built, right? So if you're building a business on Lovable, for example, you want to grow that business probably, and we really focus across this lifecycle.
And so if you ask me about the labs and so on, they focused a lot on the first part, which is building your software, and what comes after is something that we've gone really deep on. And this is about building out services, making sure an agent can work well with these services, such as doing SEO, doing email campaigns, handling payments, and hosting databases.
All of these are things that are really hard to do at a scale, especially when the load on the systems are fundamentally different than what had existed before. So we've had to kind of go really deep on the architecture of these systems to build them in a way where we can support over a million projects being created every week on them.
So Lovable is also different than the OpenAIs or Anthropics of the world, right? Because you are using whatever model that the company thinks is best for the app that your users are trying to build. What is the most popular model today that you're drawing on at Lovable?
Yeah, so we've always used a mix of models whenever you're building something with Lovable. So we have billions of tokens every minute flowing through the system, and as opposed to kind of if you're a model lab, you're a bit locked into your own models, right?
And maybe you can leverage some open source and so on that we're seeing some labs starting to do. But for us and kind of our value prop to our users when it comes to the model layer is that we partner with all the frontier labs so we can get all of them, and I can't disclose exactly how much we use of each one. But when...
you say the most common ones, like the top three that you'd say?
Well, I mean, it's your kind of Anthropic models, it's your OpenAI models, and it's your Gemini models as well. And in addition to that, we also have open source models which we can also tune, right? So we have our own Lovable models based on this, and this really gives us kind of the full spectrum.
Where, you know, you hear people talk about this Pareto frontier, where you're trying to optimize for whatever task you're doing to find whatever has the right cost, latency, and quality. And that looks different, that trade-off looks different for each task. And so depending on what you're doing, we'll serve you a different model.
Are the open source models, as they're getting better and better, just a lot more cost-efficient for you? And are you seeing a shift toward using more open source models than you were, let's say, a year ago?
Yes, definitely. I think open source has been historically, they haven't been competitive with the frontier, but in recent months they've really advanced a lot where it's very competitive on cost and latency and also actually quite competitive on quality.
And so a lot of what we do is able to evaluate these models in real time so that when you give us a query, we can route you to whatever is best. And so if you give us a simple task, we will route you to a smaller model that is faster and cheaper, and that will be reflected in the price that you pay us as well. So, you know, we've had this era of like...
Token maxing, where a lot of enterprises go out there and say, okay, everyone go wild, spend your tokens. And you know what follows from that is they're like, oh, now we're spending a lot of money on tokens. So let's figure out how we can do this efficiently. And that is part of being able to route to different models to do so efficiently. So, Anthropic, Google, OpenAI, you know, other players, they're all really angling to be number one. Who's going to get to AGI or superintelligence or whatever you want to call it first? Do you think this is a core part of your business, is testing these apps or these models head-on against each other? Do you think there's going to be a winner-takes-all and one will kick off a virtuous cycle and win the AGI race or not?
I would hope not. So, I mean, part of the reason why we started Lovable is because we think it's very important to democratize this technology and to really spread access to these models. And so, you know, part of our philosophy in designing our product is that it should be very easy to use without sacrificing any sort of power usage of it. And I think this is something that has historically been basically impossible to do. But now when you have a lens, you can combine power usage with a product that is easy to use. And this gives us the opportunity to distribute this technology to people with varying degrees of technicality. And so for us, that's really core to our mission. And I think being able to build a product like this that anyone can use, and then also being able to use these different models, is of course quite important for us. As someone, you know, we're doing our own thing on the model layer now as well, but we heavily rely on our partners here. And having that ecosystem where there's at least two free players is, I think, very important.
So while there's a lot of opportunity in democratizing access to anyone, even people like me who are not coders, being able to build an app very quickly with Lovable, there are also risks. There is a lot of concern about agents right now and the various incidents we've seen of agents, you know, going rogue, so to speak. And just one example involving Lovable and some other vibe coding platforms was where researchers found some corporate data that was exposed and phishing sites created with those kinds of apps. What went wrong in that situation? How do you stop this kind of situation from happening in the future?
Yeah. So, when you're creating kind of millions of projects every week, right? There will be a lot of bad actors that want to do something bad with the platform, or good actors who might be targeted by hackers, right? And so I think as a platform we have a lot of responsibility in making sure that whatever you create on the platform is safe and secure. And so, you know, we do a lot of things around this, such as subsidizing anything you do to make your app more secure. And so, for example, security scanning is a feature we have to respond to this, where we help you scan your project and find any vulnerabilities, and we give this for free to our users, and we might even force some users to use it. Right? So we really see it as a shared responsibility with the user, where we have to take a lot of proactive actions to ensure that our users are educated and actively thinking about this, even though they don't come in with that background. And so I think it's fundamentally a very hard problem. I read the newspaper this morning here at the hotel, and I see, you know, I'm from Sweden, and what happened here a couple of days ago was that all the supermarkets got hacked, or the two largest ones got hacked, and they set the freezers to be 55 degrees Celsius. And you know, just a few weeks ago we had a vulnerability in PHP that made basically the majority of the sites on the internet vulnerable. And so I think we're in this era where there are offensive capabilities that have risen to an extreme, and that for us, you know, it requires that we put the same defensive capabilities in place. So that's something we do on our platform and also help our users do for their apps that they're building.
You know, we have seen the incredible growth of coding agents, and, you know, not just expert coders, but everyday people being able to use AI to code being one of the key drivers of AI's growth as an industry in general. And with Anthropic right now positioned to be going public very soon, I think we're seeing the real test, right, of this business model in the public markets. Are you watching how this Anthropic IPO goes? And do you think for these fast-growth startups like yourself, are you looking to that as an indicator of whether or not you should be thinking about your own company's potential path to profitability or going public in the years to come?
Yeah. So, I think their business model is quite different from ours, right? They have much heavier capex to do what they're doing. And so, there's some similarities that we're acting in kind of the AI space, but then there's a lot of differences. I think it's great that we see them evolve and go to public markets, where I see it as a validation of the space, really. And for us, we've always been very close to our users, right? When you create something with Lovable, you see it right in front of you, right? The output is kind of part of the interface. And so I think there's always been this question of like, how much value does AI give you? And I think this is an increasing focus on validating that value, even for use cases that are less clear than say what we're doing.
Do you see Lovable as being able to replace certain traditional SaaS software in the long term for companies?
Well, it's happening today already. So, but yes, in the long term even more so. So, you know, we see this across segments again. So I can give you an example here of Rafael, a founder in Brazil who has an AI education company. They have about 50 employees. They're on track to do 20 million ARR this year, and they've built all of their software on Lovable. So all of his employees are kind of just building the tools they need. They don't have SaaS in the traditional sense, right? With a platform like Lovable, you can really build any of these solutions for yourself. And then for larger businesses, we're seeing when you actually need to customize software more to your use case, that's where we're seeing the first use cases being tackled, where, you know, if you have an industry-specific CRM, then in a lot of cases it might be easier to build something that is actually fully customized to you as a company, not to you as an industry.
What percent of users on Lovable are actually creating new businesses essentially with the app?
Yeah, so most of our users come with an intention to build a business or work within an existing business. So that's around 80% of our users.
80%. Wow. And is there an estimate of how much money those Lovable entrepreneurs are making on average?
So we don't track exactly how much money they're making because we don't have access to that data. But it is quite interestingly a shift in kind of where we get our revenue from. I mentioned this kind of build, run, and grow lifecycle, where we're seeing more and more of our revenue come from our users kind of hosting their sites and getting traffic. So for example, the traffic to our users' websites is on an order of magnitude more than our website itself. So we host our website on Lovable as well, and that's about 4% of the overall traffic. So 96% of the traffic is from our users. And then, you know, if I had transaction data, I could talk about that too, but we're seeing more and more people going from building to actually succeeding with their run and grow part of their business, which I see as a great success for us, that we're kind of being outsourced by our users in terms of how much traffic they get.
What advice do you have for people like myself and many people I talk to who are maybe not the sort of cutting edge of early adopters of coding apps? And when I think about what kind of app should I make, it's sort of daunting in a way, even though I know technically it may be very easy because of AI. But, you know, I think many people sort of like, well, what should I even build if that's not really my job? Do you really think that the world is full of, you know, untapped potential of AI entrepreneurs?
And what advice do you have for people who that doesn't fit into their, that's not what they wake up thinking about when they start their day?
Yeah. Well, I think one of our biggest challenges is to help people understand what they can do with the product. Like, we call it the blank canvas problem, right? You come to the product and you go to lovable.dev and it's this kind of blank prompt box, and you ask yourself, wait, what do I do with this? Right? And underneath that we have like all of these capabilities that let you build all of these amazing different things. And some people are very much leaning into that and experimenting and learning. And I would really encourage everyone to do exactly that, where it's really hard to kind of sit here and explain to someone. I, you know, I can give you a bunch of use cases and so on, but I think to get this kind of visceral feeling that you can only get by using AI to create, that is irreplaceable. So I would just say, go and try it, try different things, and look what other people are doing. It's amazing, people are using it for kind of creating personal software, software for their family, software for new companies, and also software for existing companies where they're optimizing and creating new products.
Great. Well, thank you so much.
Yeah. All right. Thanks for having me.
Please welcome to the stage Uva Hostman, founder and CEO of Stark, and Luchana Lakandrew, partner at Sequoia Capital, for a conversation with Bloomberg's Tom McKenzie.
Janna, really excited for this panel. Obviously kind of gold-plated VC, Sequoia, with an incredible portfolio and one of the most exciting defense tech companies in Europe, and only something like two years old. You're well known for the drones that you've been producing, but also the AI software platform. And today I think you're announcing, you've been pushing into this multi-domain space at Stark, Uve, and today announcing a new addition to the maritime space. So maybe you can tell us a little bit about the Vant 24 which you're announcing, I think, now. This is breaking news.
Absolutely. Thanks so much for having us. I mean, it's not too long ago, you know, defense didn't really happen on panels like this or on stages like that. Really appreciate that. So yeah, like we started two years ago, and obviously against the backdrop of technological change that is happening on the battlefields, and we, as you know, Europeans, as Westerners, we need to embrace that change and make sure that we also turn it into real defense and deterrence so that hopefully we never really have to use them. And the drones, the aerial drones, the flying drones, was a very, very clear example of that. But that technological change of software-driven, inexpensive, mass, scalable, unmanned technological solutions that turn into military capabilities, they need to exist in the maritime domain just as well. And the Vant 24, the boat that you just saw, is a 24-meter boat that can carry other unmanned boats. There are no people on those boats. They can be, but they don't really have to be. And I think they are the perfect accompanying vessels for our frigates and cruisers and destroyers that we have in our navies, because the trend clearly goes into something that we call the hybrid navy, and we want to be spearheading that revolution together with others, where we see real deterrence being created by using our established military structures but really enriching them with those new technologies that are first and foremost software-driven, AI-driven, but of course also really materialized in physical structures.
And I want to get more from both of you on the hybrid navy picture, but just on the specifics of what the Vant 24 can do and will be able to do, just give us a little bit of detail.
So we have Aventa 6, we have Aventa 12, we have Aventa 24. The larger, and that means the length of the vessel. So the larger a vessel gets, typically the better it can operate and survive in higher sea states, so in more treacherous environments. So think not only about coastal waters, but, you know, the North Sea, the Arctic Sea, and so on and so forth. And then I think we see a big trend towards modular payloads, which is in line with how we should be thinking about defense. Not always defining in five years we want to be doing exactly this and then we break down the engineering process, but we want to have flexibility because of the innovation speed, because of the innovation cycle. So if you see the Vant 24 picture, you see standard containers on the back of it which can carry different payloads. You can use different, you can put other unmanned boats on it and deploy it from there. So that modularity, I think, is at the very, very core. And then of course, you know, those are systems that cost a thousandth of the price of a manned frigate, which takes years to produce, and those are systems that can be deployed within months. So there is of course a big speed element. We need to create mass right now as Europeans, and this is definitely a contribution to that big push towards software-defined, AI-driven mass for the battlefields.
Ljanna, you were telling me in the green room that really you invested from day one in Uve and the team at Stark, and you've doubled down, you've tripled down on that three times now. You and the Sequoia team have backed Uve and Stark. You could have picked any drone company, any defense... There's more than a hundred European, I believe, European drone defense tech companies out there. Why back Uve? Why back Stark?
Firstly, thank you so much for having us. I always enjoy being on a panel with you. So, thank you for giving us the opportunity today. We were fortunate enough to partner with Stark when the company got started, and we've had a strong thesis on defense in the US for a long time, and then our thesis in Europe developed as well, and we thought that the trends in Europe are as strong, or honestly arguably maybe even stronger, given that governments had been underinvesting in defense since the end of the Cold War, which now is changing, and our thesis is panning out. So once we had that top-down view that there was an opportunity to build a neoprime in Europe, we went and looked for the right team, and we were looking for a combination of a few things. We were looking for a combination of a team that has the gravitas to speak with governments, understands the procurement processes, because this is very different from building your typical AI software startup that sells to enterprises in the Valley. So we were looking for people who really truly understand that, have the connections, are in that field, are well regarded in that field, but also your more typical startup requirements of being able to build great AI technology, bleeding-edge technology. And we thought that the combination of Uve and Johannes, the CTO, and the rest of the team just made a lot of sense. We also thought that there was an opportunity to build a large multi-domain platform. So it's actually such a pleasure to be here on this panel today as you launch a new product, because this is how we saw Stark evolve since day one. We didn't think we were backing a drone company. We thought we were backing a multi-domain neoprime. And why, you may ask? We think that in Europe in particular there's an opportunity to concentrate talent and capital. Actually, I think in this field capital really does matter, talent and capital in, you know, a handful of companies that can emerge as these very, very large businesses. I will also say, you know, in our industry now and again you have an opportunity to partner with a company that you really think can become truly large and actually do good for society, and this was the case here. So we see Stark as very much a platform as opposed to a point solution. And, you know, if you go back to our discussions back then, we didn't think of it as we're backing one of the 100 drone companies. We thought of it as we are backing the future neoprime that can become a multi-domain platform. So it's really nice that we're seeing this in action today.
And that neoprime, maybe we can unpack that a little bit. Does that, do we assume from that that you are going to be a challenger? Are you going to cannibalize, you're going to take market share from the Rheinmetalls, the BAE Systems, or you partner with the primes? What is your definition of neoprime and what that looks like for the defense space in the years ahead?
So I think in order to build a new defense company that can be a proper pillar for the deterrence architecture of the continent, this has to be the ambition, right? You have to do four things. One is you need to be close to the battlefield. So you need to really lean into that learning opportunity that you have, the iteration cycles, the new product variants that you have to launch almost on a daily basis. I think you have to be venture-first. That means for me that based on that information you need to be willing to use venture capital and funding sources like that to pre-fund the research and development. And this sounds obvious to a lot of people in a lot of different industries. It is not so usual in defense, where people have been relying on research development contracts where the government essentially pays for your R&D. We don't think this is the right way. We think it's not a good model for the systems that we are building. Thirdly, you do have to think about the necessary requirements that make this business so serious, which is certification, supply chain sovereignty, and so on and so forth. It's a very, very important aspect of almost like aerospace quality engineering that you need to bring to the table. And then fourthly, you need to think about industrial scalability and deterrence. When a big four consulting operations group came into our production facility and said, like, oh, this looks like the production facility of a washing machine. And I don't know if they tried to insult us, but for us it was, that's exactly where you want to be, because if I tell you, hey Tom, we're building 5,000 washing machines this month, next month we're going to build 15,000, you're going to be okay. Yeah, I can believe how this is going to happen. So you need to act like this. So those four things really have to come together. And I would say that not all primes that we have at the moment are doing well in all four dimensions. Let me, you know, we don't see enough response, I think, when it comes to those new modern AI-driven systems, military capabilities. Now my point is this. We will need all of us. You know, there is, yes, we have seen big unlocks in spending in Germany, for example, in other countries as well. I think we have seen a fundamental understanding that we as Europeans need to step up. Those are all good things. We've seen change, positive change, procurement behaviors, indoctrination, and so on and so forth. And that for me means that now is the time to team up with people. You know, I think fundamentally for us as Stark, the ambition needs to be we want to be a good partner for other people. We want to work with other people. We work well with large primes, and we have some great examples that we value our partnerships very, very, very much. And we also are the prime for smaller companies that we bring along. If you believe in, you know, software-defined warfare, interconnectivity, a kill web, you should also believe that companies need to be good at cooperating with each other. Otherwise it kind of wouldn't make sense. So all this to say, I don't think we're going to need to elbow anybody out. I think we need to partner better with...
But Luger, you think there is going to be consolidation in this space? You think there needs to be consolidation?
I do. Again, to take a step back, we very much have a platform vision in defense. I should say I don't think that there will be 20 defense companies that are going to be worth, you know, tens of billions of dollars. I think there will be a handful of companies that are being built now and that are emerging and have a chance to become very, very large. So from that perspective, I think this particular industry has different dynamics from AI or from SaaS back in the day. I think what those companies that have a chance of emerging, that the neoprimes are doing well, is concentration of capital, concentration of talent. And, you know, current primes, I'm sure they're attracting great, great talent. But I will say when it comes to really attracting the smartest AI people, the smartest young techies, I think a company like Stark is just doing an exceptional job. I will tell you something else. You know, we have a talent team at Sequoia that helps our companies recruit, and a couple of years ago when we reached out to people, the interest in defense was not as strong as it is today. You know, we heard, oh, maybe it's not an industry I want to enter for ethical reasons. Today it completely flipped. When we reach out to people, they say this is an industry I want to be part of for ethical reasons. So I just think that the mission, I just think that talent has become so mission-driven. So to answer your question, I do think there will be consolidation because a platform like Stark has access to capital, has access to the smartest people in Europe, and a lot of the smartest people in Europe want to be in defense because it's important for our society. And once that happens, they can go and build the next product, like in this particular case, and they can go and consolidate other smaller platforms as well, other smaller point solutions, I should say as well. What we're seeing in VC right now is that there is a lot of funding going into defense, and that's a great thing actually for a company like Stark, because some of these companies at some point may look to be part of a bigger platform like Stark, and that's a great spot.
Is it looking frothy right now? Valuations a little too stretched? Too much money going in or...
There is a lot of money going into defense. Absolutely. There are also great companies being built.
Yeah. I want, can I touch on sovereignty? This moment where, depending on the reports, either Trump or Musk reduced or cut off Starlink over Ukraine for a period of time. We just had Louis Mosley on from Palantir, that certainly in Germany there's been a pushback on Palantir being used in the Bundeswehr, for example. Are we still over-reliant on US tech? How much is that reliance? Can we close the gap? How do you think about that in terms of the sovereignty challenge?
