Chinese AI Models Narrow the Gap, Apps Become the Moat, SaaS Faces Repricing

0:00AI is the biggest disruption of our lifetime. That said, it's just the biggest thing we will experience, or certainly we have experienced. Now, things are evolving as we predicted. The models converge, open source catches up, and apps become the moat. And we talked about that last week. Frontier Labs have to become application companies. And I think that's going to be the red thread running through a lot of today's session. That, I think, is music for Europe. We need a proper inference chip strategy. We need to use public procurement as anchor customers. Yeah, become buyers. 100%. Listen, does anybody for a second think we're not going to need billions worth of hardware CapEx to drive inference for public sectors in healthcare and elsewhere?

0:41Of course we will. I mean, if you've got this incredible toolkit and it's free and you can manipulate it how you want to on a European box. Yep, yep. You can run it for a fraction of the cost, probably a 20th of the cost of Opus. Hello and welcome to Upside, where every week we look behind the headlines and bring you the real stories that affect our ecosystem. Today, Mads and myself. And what are we talking about? What are we talking about? What are we talking about? We've got the EU is quietly cannibalizing its own AI act. We're looking at that. More cyber breaches, but what's really happening?

1:16Sergey Brin comes out of retirement. I put in brackets here again to chase Claude. SpaceX buying Cursor, or will they? Has Europe ceded space sovereignty completely to SpaceX? And Wall Street opening a short on private credit. What is this going to mean for SaaS startups?

1:46This show is not investment advice, and the hosts of this episode may be invested in the funds and companies featured. Now, Mads, my good man, what is happening with you and how are you? Delightful to and delighted to be here. Dan had a good week just back from Amsterdam. Went to the 0100 conference. How do you say that? And I don't even know why. Is it 0100? I don't even know what it is. Do you know what? I've been saying 0100, right? Because we are British and we are in Europe and that's how we say things here. and everybody corrects me and they're like no it's 0100 and i'm like how can it be it's a european conference in europe and why are we speaking american english we want to be more american baby we want to be more american there you go we're manifesting the yanks at time press deep seek 4 has just come out so we're recording on the friday i think it came out last night so this is the deep seek new model deep seek v4 open weights free to the world benchmark benchmarking just behind Gemini 3.1 Pro.

2:46One million token context is standard. That's effectively like a stack of books or stack of novels, working memory as default. So the US banning cutting-edge chips from China for the last three years has pretty much made no difference. The questions, I guess, is the compute moat still relevant? Is the thesis now bleeding? This is the other thing that I was thinking behind the scenes. Is this giving Europe a potential way ahead? Is there something or some way that Europe could catch up on this. But Mads, this is your happy space. What's happening with DeepSeek V4? So many delightful angles here, Dan.

3:20Just taking the time machine back to 2025, January, we all will remember DeepSeek came out at the end of that month, just around the time of Trump's inauguration. Model R1 dropped their first reasoning model and NVIDIA lost nearly $600 billion of market cap in a single session, the biggest single stock drop in history. The low that day was$118. Yesterday's close was nearly$200. So it's up a lot since then, which sort of tells you that, yes, there was a buying opportunity, but markets maybe think that it was overblown and NVIDIA is right back at the top of the throne. Now, DeepSeek is back at it again.

4:00There's a new model, V4. And what's really interesting about it is that it seems to be bloody good. Now, we don't have any independent benchmarks yet because it was just released overnight. But what we have seen from their own in-house evals says that it is on par with Opus 4.6. That is Anthropik's almost leading edge model. I did not know that. That's close. I mean, that's like super close. It shipped on the 5th of February, that Opus 4.6. And we know that we've had 4.7 since then. but is effectively being on par with a model that was shipped just 78 days ago. So if you want to be very crude about it, you can say that the advance that Frontier Labs in the U.S.

4:44have on Chinese open source models right now is 78 days. How the hell have they done it? How the hell have they done it? We'll never know, I guess, but what the hell the heck? Well, we can speculate. There are some things we know and there are some things we can speculate. So one thing is that they are working bloody hard. It's a brilliant team of brilliant scientists in China. They've created some great innovations in terms of how things are made, compiled, etc. It's a very efficient model. It's a model that's trained entirely. So we're led to believe on Chinese chips, meaning they were able to get the semiconductors they needed without having to rely on NVIDIA.

5:24Two years ago, we said that the compute mode was CUDA and GPUs, and now there are alternatives. That's really interesting. The question one might ask is, what about Anthropic? I mean, we've been praising them and saying how great they're doing. My suspicion is that in the short term, they're probably fine. We know that DeepSeek, they don't have enough capacity to be able to scale. They've already rate-limited the new model. Our suspicion is that Western enterprises won't be quick to send their proprietary data and their proprietary code to Chinese-hosted models. Why? Well, Anthropic, they've formally alleged industrial scale distillation, meaning that they're essentially saying, look, the Chinese labs used Anthropic models and effectively Anthropic IP to build the open source model.

6:09But they have to, right? I mean, this has to be, otherwise, how the heck are they doing it? I mean, yes, that is certainly the way it seems. And so if you know that that is the way the model company runs its business, i.e. IP infringement, are you going to send all your IP to them? And the guess is probably not. But that said, it's an open source model. Yeah. Download it from Hugging Face. Yeah. And anyone can host AWS. The economics will speak, right? I mean, if you've got this incredible toolkit and it's free and you can manipulate it how you want to on a European box. Yep. Yep. You can run it for a fraction of the cost.

6:49Probably a 20th of the cost of Opus. Wow. Wow. So that will have an implication. Now, things are evolving as we predicted. The models converge, open source catches up, and apps become the moat. And we talked about that last week. Frontier Labs have to become application companies. And I think that's going to be the right thread running through a lot of today's session. And that, I think, is music for Europe because that's where the opportunity still lies for us. Again, we're going to talk more about that later. Loving it, loving it, loving it. Well, let's see how, I mean, one random question. Do you think this is going to affect any pricing in markets?

7:31I think it's going to be an excuse to re-look at NVIDIA share price again. How could it not? Or is it just expected? I mean, the cat and mouse game is quite extreme, right? So maybe it's all priced in and expected. I don't know. I don't know if this is going to have a rumble. Fuckets run on fear and greed. We can all see the secular traction of AI. There's going to be more AI. We're going to be spending more money on this stuff, right? And so everybody is piling into this, even though we've got issues around oil prices. We get war in the Middle East. You get all this stuff. Stock markets are still up because AI, right?

8:07It's everybody wants AI. And so every now and then people are going to look for excuses to take a little bit of risk off. And probably an argument that today could be one of those days. Maybe, maybe not. If not, there'll be something else next week. But the secular trend will still be there. There will be more AI. Yeah, yeah, yeah. I mean, and there is a reality to this technology that is unbound, I would argue. But we'll see how the winners win and the rest of the field doesn't do so well over time. I want to talk about Monday the 28th, which is, according to reports, when the EU is going to effectively be gutting its own AI Act.

8:49So this has been dressed up as simplification, but I don't believe that that's what's really happening. What's going on behind the scenes is Parliament, Council and Commission are sitting down for their second trilogue. So this is all somewhat slightly controversial insofar as they do this behind closed doors. So nobody really knows what's going on. To backtrack, the original EU AI Act entered into force on the 1st of August 2024, but it was a milestone basis of enforcement. However, under the new deal being looked at, implementation of standard and high-risk systems is going to slip to December 27, as an example.

9:26AI embedded in regulated products, i.e. medical devices, machinery, vehicles, slips to the 2nd of August 2028. But the kicking of the can isn't the whole story here. They want to move the entire juicy high risk as in AI embedded in regulated products out of scope completely folded into other safety laws, which actually smells like pretty much a repeal. So kicking the can, getting rid of some of the really tough bits forever. It looks like the EU's data protection team kind of lost the arguments to the commercial team and the competitive crowd. Is that a fair assessment? What would you challenge?

10:08What does this mean for us and our ecosystem? How is this going to pair out, do you reckon, Mads? To me, it looks like there is a recognition that a mistake was made, and now we're trying to rectify it. We're dressing it up as a number of other things. It's a bit of technicalities. We're moving something from one regime into another regime. We're saying, oh, it's about simplification. Look, it probably at some level feels better than just saying, we were wrong, let's fix it. But that is probably in reality what it is. How would this kind of affect us? Is this going to in any way affect investing startups or how our ecosystem works?

10:49Or is this so what? No, I do think it's important. There was an argument at some point that the EU AI Act could protect European AI, but I just don't really buy it. Making it super hard for our homegrown companies to compete is not going to help us. And so I think this is about rectifying a mistake, but I do think we need to go on the offensive to actually win. You can't just correct mistakes. That's not enough. We need a proper inference strategy. We need to use public procurement as anchor customers. Yeah, become buyers. A hundred percent. Listen, does anybody for a second think we're not going to need billions worth of hardware CapEx to drive inference for public sectors in healthcare and elsewhere?

11:34Of course we will. Let's find a way to buy some of that locally. We need single market unification, 20th regime for companies. We've talked about it so often. There's so many places where there's still friction. We need to find a way to unlock the European pension capital. We can get more money into innovation. That's important. And we need to think about how we can get more energy to our data centers. Is that a nuclear for compute? Is it SMRs next to data centers? We need to find a way to move, kind of remove those barriers that stand in the way so we can get the energy we need. I think those are the five major blocks we need to get right.

12:11Kind of fixing the mistake of the EU AI Act, that's just sort of just the beginning, right? That's just cleaning up. Now we actually need a strategy to go forward with ambition. Right, let's move on. Cyber breaches. There's been a few more. I've always been quite big, and I've had this as one of my predictions before on this pod, that we're going to see some pretty horrific cyber stuff in 26 and beyond. But I think this year we're going to see more. On Thursday, the UK's technology minister told the House of Commons that the health data of 500 ,000 UK Biobank volunteers, which is DNA, blood sampling, medical scans, lifestyle information has been leaked and is for sale on Alibaba.

12:51However, an Oxford academic who's been tracking this stuff says there's hundreds of exposures. And I know that you've got some insight onto this, Mads. So we're going to come onto this particular breach, non-breach. The other news story this week was France's National ID Agency also confirmed that they had a breach. But this is a much larger scale. So this is 19 million records up for sale on the dark web. Data from ANZ, which is the French ministry's, the department that looks after driving licenses, national identity cards, passports. Basically, the French state's own identity registry is up for sale.

13:27And that's one in three adults. And there were a few other things I saw in the news cycle this week. Mads, what's your take on cyber and the nefarious things going on behind the scenes? There are actually some different stories happening, as you point out, and they're quite different in nature. So around the cyber breaches, you're absolutely right. We had the breach with the French identity records, and they're being run by Atos, which is kind of an outsourcing firm, which is insolvent. So you've sort of taken this very, very precious data. You've given over to an outsourcing company, which is on its knees, and expected them to be great at protecting it.

14:05And that's probably not the best way to do things. We've seen similar things in the UK. We saw Capitan 2023. they lost 6.6 million records. And again, there was an issue. Actually, the breach was detected quite quickly, but then it took a long time for action to be taken. So the first safeguard had been put up, but then the process was broken in terms of making sure that things were followed up on. So the pattern here is that governments outsource data management, and then the contractors fail, and then you have kind of some real issues, real repercussions coming out of that. Now, the things you can do to address that, and we're going to come back to those in a minute.

14:43But I think let's just jump over to Biobank because that's a story that many people have looked at through the same lens, but actually it's very, very different. The Biobank was started in 2003 at Oxford, and the whole idea was let's gather all this data on health, on where we sort of say here are people's health records, here's the DNA, here's all the stuff we need to gather to then give scientists the ability to research with great data sets. Now, back then, it's hard to remember, but 23 years ago, the world was a different place. And at the time, the view was, look, it's much better to have tens of thousands of researchers from all across the world access this data so they can help use the data to create new cures.

15:32And so the data wasn't kept proprietarily. it was opened up to the world and you kind of have to sign an NDA. And if you did that, then you could access the data. And not only could you access the data, you could download the data. And lo and behold, 23 years later, we now find out that somebody downloaded the data and then put it up for sale. But you must ask, well, if you have let this data out to tens of thousands of people, that there shouldn't be somebody nefarious in that mix is kind of maybe a bit naive. And so I think there's been a trade-off there. Now, what's interesting is that China, they have a similar data set they've created actually partnering with Oxford so they could see the benefit of the approach.

16:17The data set's been created there, but under Chinese law, that data cannot leave the country. And so they've taken a very different approach to data protection, whereas kind of the UK approach was just come get the data, download it, you know, access it, do with it what you want. So whether this is naivety of history decades ago that's catching up with us now, I think it's probably more that than a cyber breach as such. Although, of course, in the public discourse, people are happily mixing it all into one big thing. I think it was quite interesting to read that the data had been de-identified and not anonymized.

16:55So my assumption is data has been downloaded. It's then been un-de-anonymized, as in you can match various data sets and then create a full data set, which you can then sell. So I think also there's this 23-year-old mindset which doesn't play in the modern world. So I think there's a little bit of catching up as to how we think about running these programs and who has access to what. I'm sure that's already been thought through, to be fair. yeah i look i think all this uh sort of becomes relevant in the context of national identity cards right because we've been talking about that in the uk and the government has been talking about this idea of the brit card will it ever happen do you think i mean i can't i just can't see the brits actually ever going for it because we've been so kind of mind bent and and we've got such a negative attitude to even though we have passports and driving licenses and credit i mean And if you want to find somebody, it's not hard.

17:53But we've got this resistance to these things, haven't we? Yeah. And I think there's a lot of fear mongering. And perhaps also some people just don't like the idea of what they'd see as big government and don't want government to be able to look at this data across different databases. I do think if you want a modern economy, a digital ID is not optional. You can't have a digital economy if government can't verify who you are. You must have a way. Preaching to the choir, my man. I totally agree. And if we look around the world, Estonia was Soviet era until 1991. And they've been running digital identity at scale now for more than two decades.

18:36So I don't think we really have an excuse. I think some of the things that got right in Estonia that we should copy is it's a distributed architecture. It's a data stays in individual agencies. It's not a central honeypot where you can go and just siphon off everything. And they've been very, very good around the auditing and the tracking and everything that gets queried on you is locked, is visible to you, free on your phone. So if somebody, if a public agency wants to query something, you can see exactly what's been asked for and when. It's state-owned IP, but it's open source. So that means instead of having sort of a big Fujitsu horizon where you have some contract or an Atos, right, that doesn't really know what they're doing, you have a lot of open source that's running this, meaning that it's been much easier to have different vendors compete.

19:24It's been easier to identify bugs. And you have selected disclosures by default, which I think is super smart. So the Estonian system allows you to prove you're over 18 without revealing your date of birth, as an example. Right. So there are lots of smart things you can put into this if done right. But I think it's clear that the approach that was taken in the UK was almost the opposite. People started by saying it's about checking right to work. Because we don't need a new system to do that. But there are probably things we do want help with, right? Making it easier to file your taxes, easier to get your NHS records, easier to file things at Companies House.

20:02Companies House is still struggling to know who we are when we file things. It sounds nuts. And they've tried to build now their own ID system. I mean, it's like, instead of having one, you end up having 27. How is that going to be safer? It's just bonkers. It sounds a bit nuts. Yeah. So the question for me is not whether we should have a digital ID, but I think it's what architecture can make it as safe as possible. And there's definitely still work to do there. I just wonder if having that conversation openly with the public would sway any opinion. I don't see my wife's going through some health issues at the moment.

20:36and I don't understand why her health records aren't on her handset. Why? And I know there is an NHS app and we have private health care, but there should be some way of tracking, even on that simplistic basis, because every time she goes to a doctor or a health professional, they haven't seen her records and she should have those and be able to talk with the people directly. So I don't know. It just feels so obvious and so simple and so straightforward. So I hope we learn. Yeah. We're going to talk about Sergey Brin has come out of retirement yet again. He is now running a strike team inside DeepMind, and they are looking at closing the coding gap on Anthropic.

21:17So this week, the information broke the story. The leaked memo was as such as Sergey Brin saying, to win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers. So every Gemini engineer is now required to use the internal agents because they were using Anthropic clawed agents, which I think is a bit embarrassing. So they are now being told to use their own kit rather than using clawed to ship. Anthropic uses allegedly AI for nearly all its own code. Google uses AI for about half, it looks like. Brim was also framing that coding is the unlock to recursive self-improvement, which I think we would all agree is the real prize.

22:02And whichever labs AI writes the next generation of AI fastest compounds first and wins the race. So lots at stake here. But that also raises the uncomfortable question for us, which is none of our lot in Europe are even in the conversation. But if frontier compounding velocity is the remote, where should, I mean, maybe we can tie back at the end of this, Mads, to where should our sovereign AI bet be placed? But where would you start on this one? I think it's really interesting to see what's happening with Google. If you recall, going back to the middle of November just last year, Google dropped Gemini 3 with deep think mode and with the anti-gravity IDE to compete with Cursor.

22:42And there was a comeback story that really landed. And for about a week, everybody was talking about how Google was back in the race after a rough year. I remember us talking about it. It's like, oh my God, it's catmell. Google is back, Google is back, Google is back. And then a week later, Anthropic Shift Opus 4.5. And it was just like, okay, nobody's heard from Google since then, honestly, right? And it still amazes me how far open AI has fallen from grace and still hasn't, in my mind, regained the mantle. It's still Anthropic up there, but I find it really fascinating to watch. Yeah, I mean, right?

23:19So Anthropic shift 4.5, and then in January 4.6, and then now they've released 4.7. So we've just seen so much velocity from Anthropic, and it's been looking like Google was standing still. So clearly they know that they've got work to do. But here's what's interesting, because on one hand, you've got sort of that slight panic. And yes, Brent has now come out of retirement, I don't know, the third time. He was ready. Wheel him out. We need another PR story. Wheel him out. That's fine. Look, what do you do when you have a 300-foot yard, right? I mean, you've got to go see it every now and then.

23:57But here's the thing. Google owns 14 % of Anthropic. Yeah. They have just in the same week, they have inked a deal with Anthropic to sell$10 billion worth of TPUs. Okay. So on one hand over here, they're panic stations and we've got to compete, compete, compete. But on the other hand, oh, Sony also owns the shares in the company and we can sell them a lot of GPUs. I think it's pretty clever. I must admit. Yeah, yeah. Well, there's the PR story, isn't there? And then there's the reality that goes on behind the scenes. I mean, people probably do know these days. I mean, when Apple was like, you know, we are, you know, the smartphone provider back in the day.

24:38And actually, all of the devices were filled with Samsung screens and Sony this and other provision of that and Google licenses for the other. So there's the PR piece, I always think, with these things and the narrative that the department wants to push. And then there's the real commercial reality that goes on behind the scenes. Yeah. So, look, I think Google is definitely – they have it all to do on the AI model front and on the agentic coding front. But if you look at the full stack, there's still some smart strategy that's happening. If you look at the models, what we've seen is that even with the anti-gravity IDE, so with Google's coding tool, you can still use the anthropic model under the hood.

25:22What about Europe? Yeah, you asked about Europe. Look, our players and applications, not foundation models and not frontier compute, I think. That's okay though, right? Yeah, I think that's fine. Look, in the end, it's always applications. I think so. Go back 50 years, right? in the 70s, five German IBM engineers said, we want to make applications. And IBM was like, who wants to make applications? We just sell compute. It's fine. And then they went off and they created SAP, which, of course, today is worth as much as IBM. And so I think you're going to see exactly the same here. Everybody is moving up the stack and into applications.

25:58We have another story coming up in just a minute, which is going to be exactly talking more about this move to apps. So everything, I think, is evolving in the way we've been talking about. I mean, I've always been bullish on European apps and the actual application and not just on public services, but in private markets, too. It's just this is this is to me is the super exciting place and obviously slightly talking our book because this is where we invest. But I think this is this is from a European perspective that the real magic is is coming to force now. But let's talk about that very deal because you're talking about SpaceX striking a deal with Cursor for$60 billion.

26:36So on Tuesday, they announced the option to acquire Cursor. This is an option. This is not going to happen pre-IPO. I think it would have just muddied the water from what I can gather. But this is looking at a$60 billion valuation deal later this year. Sorry,$60 billion price point, actually. Or to pay$10 billion if they break the contract just for the privilege of working with them. But this, looking behind the scenes, is not$10 billion in cash. It's 10 billion in compute access, I think, mostly. So there's some other, from what I can gather, some kind of contractual finaglings that make this quite an attractive deal for SpaceX.