This Week in European Tech: Is AI’s spending boom creating real value or hiding bigger risks?

26:32Lomax Ward:Again, you don't know the reality of the reality of the reality, but they were very clear at naming specific tooling that was found and how it worked and how it opened up tunnels through and what it did. So maybe there's more to that. It's a mildly terrifying thought, isn't it? Having a sort of titanium army rather than a terracotta army of humanoids marching around. In your kitchen, in your kitchen spying on you. she she looking looking through the looking glass with his multi with his white cat on his knee you know she's using to make my tea and toast in the morning when i uh i think i think this is the interesting story from the context of early stage startups right and it is that if you're working in this domain the quickest and most capital efficient thing that you can do at the moment is get on a plane to shenzhen basically uh whether you're in the us or in europe um start building out your supply chain there.

27:27Now, clearly, some startups might have the plan that they would do that at the beginning to show proof points, obviously much more capital, capital efficiently, and then would maybe be like, right, we're going to raise this much amount of money to onshore the whole supply chain, right, which is a kind of that's a journey you could go on. In a way, this is shutting down that journey much earlier, which I think potentially is a good thing. But it does mean that those startups earlier on in their journey, you're going to have to raise a lot more money right to start doing you're trying to find more expensive suppliers in the us or if they don't exist which is probably more likely you know doing building this from scratch themselves so sometimes not even the us depending on who you're selling to i mean if you're going to have chinese parts um and you're selling to the uk government i mean at this point pretty much forget it unless you're in something that's absolutely trivial but if you're doing any any deals with the you know mod or even the ministry of justice or actually any major department where you're mission critical then they're not they're not going to want chinese parts and actually in some cases some of our startups have been forced to have an entirely uk uh supply chain well that's where europe's going to that's where the eu's going to i have to say it's ridiculous but i would say that um is it though i'm not sure it is i think i think i think well at least i think at some point if you can get it more cheaply and done properly in europe then fine like you know what i mean and and better but it probably won't be probably won't be cheaper but do you want to be reliant on a chip or something where your entire ability to deliver could be shut down because the u.s goes on i think at some point it's not going to provide your this retrenchment from globalization has to stop somewhere like it cannot be that for example sweden of 10 million people builds everything sovereign that france builds everything sovereign that italy does it won't build everything sovereign but i think where you've got critical where you've got critical industries or you know the us has a strong track record of making political decisions dressed up as other reasons for for blocking sales or increasing taxes on or banning export and i think you know if the us feels that it's got a market being threatened or one of its major companies is being threatened because there's a high growth player in Europe.

29:48I don't think with the current climate, it would bat an eyelid at shutting off supply in order to scupper the growth of that non-US company. And we've seen this over and over again. So it's not just security. The US does all the time sort of talks about free trade and then does exactly the opposite and throws its weight around to protect its home industries i'm not i'm not suggesting that you build you know u.s suppliers into your if you're european government i would say you you might want to have a fire break and not have like not have um u.s critical suppliers in your in your supply chain but stepping back to this is what this means is in a way it's a confirmation of the direction of travel that we're already going on which is that you know supply chains are becoming more vertically sovereign um this is just an acceleration of that this does cause i think the problems that i referred to which is that you know suddenly this this stuff all becomes a lot more expensive you know for example if you know there's one or there's one or two and maybe andrew knows more maybe but you guys know more than me like companies building you know humanoid robots in in the uk and europe right that that's the most humanoid well there's humanoid of course yeah that's a big like capital project You know, of the likes of which Europe's never been very good at, quite frankly.

31:04Northvolt, you know, many, many examples, right? So that's a big swing. Whereas, you know, you'll talk to other startups that are like, oh, you know, we can actually get the, we can buy a humanoid for 40K or 50K plus VAT from China. It's dead hardware and we can then program it, et cetera. So that's a very capital efficient way to build your business, go and start winning customers. but yeah like it's gonna it's gonna be a it's gonna be an exciting time to see whether this

31:32Lomax Ward:is actually even possible i want to talk a bit about ai i think this week i've seen more ai news stories of various flavors than i have in a long time it might be my happy years but it feels like there's been a whole world of very particular different pulls and pushes in the in the ai sphere. NVIDIA backstopping a$250 billion data center project in Ohio. There is more to the open AI hack hugging face than we previously thought. We've seen some noises now from Sam Altman and others about deceleration and a new AI safety initiative, including all the big tech. We saw Mythos, some stories about Mythos and Open 5.

32:23Lomax Ward:Then there was one I particularly liked the look of, which was Reflection AI, which was the open source. This is Jensen Huang's almost love project, which then they invested in, which was Reflection AI in the States, which is the American open source or one of the American open source offerings. We talk a lot about Chinese open source, but there is open source all over the place. Obviously, Meta, I think, started out. Lama was open source originally, wasn't it? I think. Anyway, Mads, your happy space. What is happening in this wonderful AI week of news? I think you outlined a lot of the interesting bits there.

32:59Of course, the NVIDIA guarantee is interesting, not so much for it being a guarantee, but for the size of it. So you're providing for a company to provide a guarantee of a quarter of a trillion dollars towards somebody else's business is just, I don't know. I can't remember anything else of that scale that I've seen.

33:19Lomax Ward:We take it for granted now a little bit. We take these numbers, I think, just like, yeah, it's okay. These are crazy numbers. It's just a ton of money. Why are they doing it? Well, it's because it's going to be one of the biggest customers, if not the biggest customer. So OpenAI, they're building a 10 gigawatt campus in Ohio, and the whole project is over half a trillion dollars. So imagine that. When you think about some of the other big infrastructure projects we've had throughout history, this is not an industry build-out. This is one project that's going to cost more than$500 billion. What happened with things like Stargate?

33:52Lomax Ward:I mean, I remember when this all kicked off, we were talking about Stargate and other big projects. It feels like they need a new PR story, so they invent a new big bucket of money going into a big mega-project thing. Caustely earnings, man. Got to keep the costly earnings. But at some point, we collectively, on this pod are going to have to go back and measure what's been built against what promised because this just feels like I don't understand where all this money is coming from. Well, apart from NVIDIA, of course. I mean, NVIDIA wins and loses regardless, right? Whether the open models win or the closed models win.

34:28And it's also an investment in relatively near term because these chips are cycled through, well, I suppose with the hyperscale is every two, three years, The top-end chips are a place to keep up. So once you've locked someone, one of these data centers, into that stack, it's not just a one-off from Nvidia's point of view. All those chips will get replaced in that data center, and I suspect in the contract that's wrapped up in these deals.

34:51Lomax Ward:Well, more than that, I mean, CUDA is such a delicious moat. I just can't see anybody else getting it. Forget that, well, the chips are important, but chips aside, the OS that runs the whole ecosystem is… Yeah, the stack. Yeah, it's like what a moat. What an incredible moat. 2006, I think that all kicked off. 20 years, a 20-year play. And if you want to know more about that, read The Thinking Machine. Was it Simon Witt? I forgot who. Stephen Witt, I think you wrote it. It's a really fascinating book about Jensen Huang and NVIDIA. Mads, what else is happening in our AI delicious world? Well, you talked about this post-quantum piece, right, kind of cryptography.

35:33And so everybody kind of colloquially knows that quantum computers, they will eventually break today's encryption. Ten years time. Ten years time, baby. Exactly. Right around the corner. So banking, messaging, government traffic, all of it. And so, you know, what shall we do? And so there's a competition at IST. They run an open competition to pick the replacements, the new cryptography systems that will keep our information safe from the quantum menace. So what do you do? You've got lots of different cryptographers that are submitting to signs. And then the world tries to break those to signs to prove that they can be broken.

36:12And the survivors, so those that can't be broken, they're then the ones that end up inside every browser, every bank, all the good stuff that needs to be protected and defended. And so Hawk was a survivor. And it was two years into the process. So here's a model that's been proposed, a new cryptography model. that was proposed, and it has been surviving for a while. Nobody could break it. And then some chap who isn't even a cryptography specialist, he sat down with Anthropics' new model, and in 60 hours and about$100 ,000 of compute, he halved the effective key size. What does that mean? Well, that means if you think of the key size as the volume of thing that needs to be broken before you have penetrated the defenses.

37:00If it's longer, it's harder. Well, he was able to cut the size of the defense in half in a couple of days, in a few days. So that model is now being deadison, right? And instead of also withdrawn, the whole idea here is that whatever new models we roll out, ideally would have to survive for a few decades. Otherwise, it's not really worth it. Not so much because, of course, you can say, well, we transmit data every day and we can just, you know, upgrade the cryptography as we have new cryptography. But once you crack the codes, you can also go back and look at the old traffic. And it's possible somebody sent something a couple of years ago that actually is, it'd be quite useful to know.

37:40And so you really want the model with a lot of longevity. And Anthropic was able to poke some holes here. So this is all very interesting for security buffs. We've seen some pricing moves on the models, Opus 5 has come out at half the price of Fable. What's your take on Opus 5?

38:01Lomax Ward:I've not, I've not, I mean, I've not seen, I don't use it like you do, Mads. I've not seen the 4.8 to 5. I don't know if there's any meaningful story around this as a release. I think maybe the meaningful story is when you look at benchmarks, it is, you know, on some benchmarks it's on par with, and on some even it scores higher than Fable. And this is counterintuitive. How can a cheaper model, a model on half the price, be better? I think the real story here is that, yes, on some narrow tasks, it is a really, really good model. It has to be handled slightly differently from some of the predecessors.

38:40So actually, if you have got context documents, if you have skills, if you've been setting up scaffolding to manage the way you use your models, you may have to rewrite some of that stuff because the new beast has to be operated in different ways. Okay, so there's a learning tax there. But I would also say when it comes to some of that completely off the wall, connecting dots in ways you hadn't even thought about, Fable is still the winner. That doesn't necessarily show up in some of the benchmarks. And I would say it's the day-to-day workhorse. Opus is formidable. If you're going to crunch some really, really hard problems, Fable is probably still the winner.

39:16Well, we're not seeing the big leaps, are we, between the major model launches that we saw before where you were getting like, you know, a 510x improvement that was very perceivable when you used it. And a lot of the focus now on these newer models, even the major releases, are sort of actually how the models use the existing smarts, in quotes, in real-world tasks. So that might mean sort of longer horizon coding with much bigger, being able to modify and hold and understand much larger code bases for analysis or actually how some of these agents actually work or reducing hallucinations slightly or increasing the number of tokens to the long context, so like millions of token usage, or perhaps the multimodal elements, rather than a sort of fundamental step change in this model is so noticeably smarter.

40:05And that's certainly what I've seen in usage. And when you look at some of the benchmarking, what's been happening, I'd be surprised if we saw another huge elite without, as you mentioned, Dan, a sort of another technology coming along unless someone's going to really step change on the existing stack. Yeah, but I think there is some truth to that. I think two other angles, it might just be interesting to keep in mind. I think it's only a month and a bit since Fable was released. And I do think, I do consider that a step change, a fundamental step change in terms of AI's capability and how it's used.

40:41We've also seen a much, much shorter release cycle. So you have models come out, it used to be every three or four months, now it's one to two months. And so, yes, every increment will maybe seem a bit smaller, but they are more frequent. So, you know, I agree with you. You can probably see in some benchmarks an asymptote. But I think in real world use, we are still seeing very big step changes on some of these generations. And I would say I really hope fable up there.

41:08Lomax Ward:Listen, I know this is a really hard question to answer, but how many more step changes do we have with this current stack? I don't know that anybody can answer that question. I think if you're inside the labs, they'll still say, look, we've got stuff that you can barely dream of that we have in tests now. I think one of the markers, obviously, is I mean, there were some there were some other stories this week about the Hugging Face hack, where there were other organizations that were also hacked, which didn't report like Hugging Face. and what i saw dario talking about some of their sandbox models actually escaping getting an internet connection and then hacking inverted commas other other pieces of kit so maybe there will be some more restriction layer and maybe we won't see not that it can't be done but we just won't see the step changes because they're too dangerous perhaps a lot of the step changes early on in the early versions were driven by simply adding more internet content and adding more text or adding more GPUs.

42:19And that's what drove some of these big jumps. And obviously, that's a lot of diminishing returns in that you come to the end of how much content you can throw at these models to make them significantly, obviously, better compared to the very early models where they'd only got a fraction of the content they have now. So some of the improvements early on were driven by what had been fed rather than a sort of fundamental shift in the technology being used. But I think Mads is right that we will see more and more of these, more and more of the tooling be optimized so that it does work better and more efficiently with fewer mistakes.

42:56But that is different to the fundamental smarts of these models getting better.

43:00Lomax Ward:Well, we need it to come down to the actual application layer. It needs to come down now to more real world. And I was looking at, I forgot what the website is called, but it was something like remote labor or something. I'll post it in the pod notes where they were looking at how many roles or how many jobs plus roles have actually been fully automated or do a good, AI does a good job of. And it's a lot less than we would think. I think Fable came up top at about 13 % of completed roles. So the actual automation, that application layer, there's still a lot to do when it's actually looking agentically at running a task or a role.

43:36Lomax Ward:And we'll dig into more of that another day. So much of this is engineering. It's engineering rather than R &D. And I think if you work with the models at the front end, and I may just disclaim, we're not in kind of 100 organizations that have access to the best models at all. So I wouldn't claim that we have the absolute bleeding frontier, but we pick up the new models as soon as they're released. We use them for a lot of stuff. So many of the challenges around turning the capability into real-world value is engineering. Let me give you an example. So the models now, they can code and they can develop and they can build pretty much anything you want.

44:15And, you know, we built software factories. We have things setting up now that can almost autonomously build new capabilities for our teams. and people come up and say, look, we want a piece of software that can do X, we want something that can do Y. Now, these are built as autonomous graphs, right? These are autonomous systems that'll build the code, check it, verify it, security test it, implement, and then deploy it. What can happen if you're not careful with this stuff? And we saw that in a specific instance we had, was that some of the checkers were checking code, finding a bug, issuing a correction, then pushing that back into a queue that then would go to another checker that might find another bug somewhere saying okay well we also need to fix that so we're creating a new tick and we're creating a new raising a new thing meanwhile that all started to cascade and fan out and this the models the systems were able to find new things they want to correct sometimes very small things sometimes even a little common than a file that the production of new code outpaced the ability of the fund, the pipeline's ability to integrate and deploy the new code.

45:27Okay. So, you know, some of you may be familiar with the paperclip story, which is, well, you know, what is going to lead to the end of humanity from AI? Well, it's that we have an AI. We tell it that the most important thing is to maximize the production of paperclips. And the AI then says, well, that's great. I can take every single material in the world and turn it into a paperclip. When I'm done with that, I can find a way to convert the mass of humans into paperclips, right? And it'll just continue until we're all extinct and then convert the whole universe into paperclips. And we saw some of that in this software production factory, where the factory became so good at identifying issues it wanted to solve that it completely clocked up and fanned out.

46:07And suddenly, there are just hundreds of issues that are all competing with each other and creating a mess in the code. So if you're not careful, this stuff can really just sort of take on a life of its own. What does that all mean? This is not an R &D challenge. This is engineering, and this is what we're seeing everywhere. This is what corporations are seeing when they're struggling. It's because the data is in the wrong format or the systems haven't been configured or set up correctly yet. But the capability to actually use the AI to automate a lot of stuff, well, it's there. It's just we haven't just completed all the engineering work to integrate and fix and debug.

46:42But it's happening. and it's coming and we can see it.

46:45Lomax Ward:Lomax, Andrew, anything else that you'd pick up on the AI news cycle this week before we move on? No, I mean, one thing hidden away, which is a little bit back in the CapEx side of things, there was an announcement that Meta put out with BlackRock around yet another data center. This was a$14 billion data center in El Paso, Texas. BlackRock will hold 80 % and Meta will hold 20%. So it's just yet another example of this off-balance sheet financing. that's happening so it's just it's i think it's just something to watch uh and maybe we can do a kind of bigger segment on it in the future is well it makes sense right as i mean i i think i think i'm pretty sure that meta had spent everything bar 800 million or something i mean there were there were a 96 percent spend like hardly any free cash flow i think so it makes sense then so how do you then how do you get what you need built well you do a deal with somebody well you do a deal But also all of that doesn't appear on your books.

47:43Lomax Ward:It's off balance. Yeah, yeah, yeah, yeah, yeah, yeah. There'll be lots of offsetting, I bet. Deals of the week, boys and girls. Mads, what is your deal of the week? The deal of the week is a nuclear deal. And I feel like maybe I'm slightly going in on Andrew's turf, but we know a little bit about it. So I'm going to steal it anyway. It's nuclear turbines, but it's a Manchester deal. It's a£50 million seat on brand with the Zeitgeist. There you go. It's a BAE Systems spin-out. It was spun out from their launch pad platform. And it's a terrific story. So Jeremy Osten, he spent most of his career at BAE Systems.

48:32And he was a principal engineer on maritime submarines until last year. And his job there was the machinery that turns reactor heat into electricity. One of the problems they were looking at is, you know, sort of saying, look, we think the nuclear industry, they've been thinking about how to turn nuclear fission into energy in the wrong way. You know, it's been all about how to create heat and then figuring out later what's happening with the heat. And they sort of flip the script on the head and say, look, what is the cheapest way to generate power once we already have the heat? And so, you know, what they've actually built, well, in a normal nuclear reactor, you're going to boil some water and then steam is going to spin a turbine.

49:15And this all sounds quite old fashioned. And to some extent it is. Steam is not exactly a new invention. So the steam engines. The problem is these these steam reactors, they're old fashioned. And so they deleted the notion of steam completely. And instead, the reactor heats compressed air directly. So the hot air goes straight into a gas turbine, and that's the kind you might find in a jet engine or a gas power station. So they've created a whole new way to turn the heat into power, and that's interesting. Now, it's a 15 million pound seed deal. It looks small for nuclear. You'd expect it would be a lot bigger, but there's still some R &D they're working through, and so that is what this is about, is proving some of these ideas they have taken to the next level before they go out and raise the big money to actually build the plant.

50:05But I just think it's wonderful. This could be the route to kind of small nuclear reactors that genuinely work.

50:14Lomax Ward:Love it. I've got a much bigger round, early stage round. I've got$152 million Series A in the UK, crazy kids. It's the UK robotics startup Humanoid Series A, and they've just hit unicorn at 1.35 billion valuation. Well done. That was mine. Lomax, have you got a deal of the week? Agon, which is also encroaching on Andrew's turf, which is a defence company, broadly speaking. It's a UK project. It's building synthetic environments for autonomous systems to train in. Big$30 million seed. I've done a pre-seed before and this one was led by Northzone. Nice, nice. Some big deals. Andrew, did you have one?

51:04Yeah, another quantum deal, actually. And this is really just more of a comment on the sector that Europe is piling a lot of money still into quantum. Zurich Q has raised the 25 million seed. They say they've got a new architecture for guess what quantum chips that you can scale more rapidly. And obviously, you know, we've got a horse in the race here. So it's just interesting that the appetite for quantum in Europe isn't dying. And I just hope the European government start writing big checks as customers for these companies, because we've got a real opportunity to have a global winner when one of these technologies really takes off.

51:39So I'm happy to see that the appetite for quantum, which is, as you say, nearly with us.

51:45Lomax Ward:We've got to be in it, right? We've got to be in it to win it. And just go back on the ability to crack RSA. And we talked earlier about post-quantum encryption. I mean, we probably are now two, three years away. Iron Q has said that it'll have an attack on this in 2028. Previously, most people agreed that you'd need one, if not two million qubits, given the current error correction levels, to crack that RSA 2048 encryption. I think everyone accepts now that for the better technologies, we're perhaps only down in the hundreds of thousands of qubits. So we're edging closer. We're edging closer.

52:23It will be this decade, I think.

52:25Lomax Ward:uh well we'll keep you to that there you go i love lovely prediction uh mads what's happening in the week ahead any earnings reports anything else happening what do we need to look out for well this week was the big week we've got one coming up on monday that's palantir and certainly an interesting insight into what's really happened with multiples and valuations at the forefront of of all the things i saw the um i saw the nhs report that palantir haven't actually saved anything they haven't created any value yet i don't know if that's just a moment in time or if you saw that report i've seen other reports saying exactly the opposite i wonder how much of this is weaponization ahead of negotiations yeah yeah yeah yeah fair okay well obviously we have not heard we've not heard the last in that story um and as we've been talking about here the last few weeks there are finally some uk european alternatives that are being funded to go and take on this market, which I just think is exciting, right?

53:21Competition is great. Anthropic IPO, still, we don't have a date yet. Should something drop around news next week? If our Anthropic bill is anything to measure by, I think it's still growing exponentially.

53:36Lomax Ward:We are feeding the machine. Well, on that note, ladies and gentlemen, boys and girls, thank you so much for your time. Any final closing reports before we close off? apart from have fabulous weekends and happy august on your yachts everybody that's all we have to say right when we're swanning around the lomax is going to be a pirate ship in august which i love yeah with his with his car and his arr and his dry powder well i'm going to be away next week i'm going to be away with the kids but it won't be on a yacht unfortunately wait for some time I'm going to a hidden undisclosed location away from the sun.

54:18Away from keyboard. AFK, baby. Well, he also has to get away from all the cameras, the media that generally follows Lemeck's around. I know. It's tough. It's the new haircut. It just draws in all the ladies.

54:32Lomax Ward:Right. Thanks, guys. Take care, everyone. And we will see you next week. See you in the next one.

54:45Lomax Ward:Upside!