AI Is Hitting Constraints: Memory, Power and Geopolitics

0:00But universally, what people are saying is we are capacity constraint. If we had more GPUs, we could grow faster. I've seen loads of reports this week on energy, on memory, on GPU, on planning. There's been a number of undertones around that topic. The bull case here is space compute. We talked about constraints on Earth and Musk keeps coming back to, look, there's just not enough power. And so space is the only place we can get the power we need fast enough to tool up on delivering the AI people want. But my assumption is also it's not just about power. It's about freedom. They don't have planning constraints, build constraints, people constraints, regulatory constraints.

0:38This is just like we can do what we want in space. Is that fair? SpaceX going public is super exciting. Another company that must go public is Anthropic. It's the most exciting company in the most exciting space. And it should be a public company. And people should be able to buy into that business and should have transparency into what's happening. What we've also seen is that there is so much appetite to throw capital at them. They're talking about a$50 billion round at kind of open AI prices. Soon there'll be anthropic prices. They'll be the market setter. Well, there you go. Hello and welcome to Upside, where every week we look behind the headlines to give you the real news affecting European venture.

1:16Today, it is Mads and myself. And what have we got on our agenda? We've got big tech earnings. Reports are out. Lots of politicking this week, especially at the Frontier Labs. More AI model drops. What's hot? Will the EU and the UK AI Act converge? Energy. Should the UK get small modular reactors online? Plus predictions and deals of the week. My dear.

1:56this show is not investment advice and the hosts of this episode may be invested in the funds and companies featured now quite an interesting news week but first up have you been how's the how's the energy in the room are you good are you bad are you upside down where are you fabulous come the sun has been out it's great it's beautiful week it is it is nice once the sun hits uh once the sun hits london it's a different place um i just i we don't normally do this but i thought mate i was quite intrigued by the by the news cycle this week most of this we are not going to dig into watching king charles tell jokes at the at the dinners and at the y-house i thought was quite quite nice i wonder how much goodwill he's bought i saw the the whiskey tariff has been dropped so that was a we got one thing i guess and for whatever that's worth shootings at the correspondence dinner obviously the hormuz crisis is still still with us i think brent at time of press was hitting 126 dollars the uae has left opec um rates to be held i think pretty much universally i saw i don't know if you saw this that canada has launched a sovereign wealth fund did you see that yep um and i also saw from the court transcripts with the musk um altman shenanigans.

3:06I thought it was quite interesting that Tesla is not currently pursuing AGI. I don't know if you saw that as well. Anything else caught your mind this week? I think he's been flip-flopping a bit on that, Musk. I think he said something different just a few weeks ago on X. He said, Tesla will be one of the companies to make AGI. And then something different under oath. But okay, that's fine. You can have a difference of the horizon you're talking about. Slightly Trumpian kind of say whatever he needs to say at the moment in time to call a reaction maybe. I think, listen, I think just talking about Trump, I think the Charles visit was brilliant.

3:42I think it was so well choreographed. I don't know if you noticed that he had the Navy. He had the, well, those too. But he had the Navy troll through the Naval Registry and they found a ship called HMS Trump. and they found the bell from that ship and they gave it to her. He must love that. He must love that. I mean, I think it's so clever. Yeah, yeah. Just grooming and loving like Tim Cook taking, what did he take, a golden disc or something? Golden iPod, I forgot what it was. Lovely. Well, let's dig in. First up we've got on the docket is markets, Mads. So this is kind of your happy place. So big tech earnings week this week.

4:27Microsoft, Alphabet, Meta, Amazon, all published figures. I think at Time of Press, we're recording on the Friday. I think Apple is releasing today, I think, or last night. Yeah, it was released last night. Okay, so tell us, what's cooking with markets, and how does this pair back to Europe, and what do we need to keep an eye out for? Well, the five big beasts have all reported. So as you said, Microsoft, Alphabet, Meta, Amazon, and Apple. And you sort of had two up, two down, and then Apple is actually looking like it's trading up here in after hours. And what we can see is that the cloud market, powered by AI inference, is absolutely on fire.

5:07Massive revenue growth. You had Microsoft up 18 % year-on-year, Alphabet up 22 % year-on-year, with the cloud parts of those businesses just growing phenomenally. Alphabet, or Google, if you will ,'s cloud business was up 63 % year-on-year. Wow. Yeah. I mean, it's just absolutely firing. And what's happening at the same time is CapEx is also going up. We've been talking about that over the last quarter is that even though it was already at a high level, they keep raising the bar. And universally, what people are saying is we are capacity constrained. If we had more GPUs, we could grow faster. I've seen loads of reports this week on energy, on memory, on GPU, on planning.

5:50There's been a number of undertones around that topic. So I'd say from sort of some of the bubble talk we had throughout 2025, I think a lot of the naysayers are kind of coming around to the fact that there is a lot more to this than just a flash in a pan. Meta have also raised their CapEx budgets, but obviously don't have a cloud business. And so they're getting punished for it stock wise. Whereas the share is down more than 8%, whereas Alphabet that are growing very fast, they're up nearly 10 % to an all-time high. Alphabet is just the darling right now. They can do no wrong. I mean, if you look at the bets they have made, okay, so they have deep mind that they're making the Gemini model.

6:36And, you know, we know that Gemini, they're selling a lot of it, but it is not sort of seen as the most popular model in the market right now. Well, guess what? They've got nearly 15 % share in Anthropic, which obviously is. They also happen to be a major shareholder in SpaceX that's about to go public. So, I mean, it's just, you know, if you want to buy just one stock to sort of be in everything right now, it seems like Alphabet is the bet. Obviously, no investment advice and all your studio and your search and all that stuff. Yeah, all that good stuff. Memory, what's happening on the memory side?

7:11Well, so memory is interesting because the shift, we've seen a shift from power being the main constraint to actually memory seeming to be the constraint. And still, okay. Yeah, yeah, you've got a Triopoly. You've got the SK Hynix, you've got Samsung, you've got Micron. And they were up last year, but they are up massively over the last two years. And they are really kind of at the end of the series of choke points that you have around the semiconductor manufacturing equipment, kind of the packaging expertise that you have at TSMC, and then actually the ability to make the memory. And some of the discussions that have been is around, you know, China, Chinese AI chips, can they make their own inference?

7:49It seems today that it's not even, even if they make their own chips, they have such a massive constraint on memory, and it takes so long to stand up these memory fabrication facilities that that becomes a chokehold in itself. Interestingly, just going back to earnings, Apple, they had a really, really good quarter. There's been lots of talk about, you know, Apple obviously not being in the CapEx game, not being in the AI game like others. I'm sure everyone's seen the little memes and the little graphs. Well, Tim Cook, he came out and said, look, we are in the AI game. It's going to be personal.

8:23It's going to be private. It's going to be fast. It's going to be on device. I'm so excited by this. As you well know, I think they are going to come out strong. I do. Is it John Turman who's the new CEO? He's the product guy. I just hope he pulls it. It'd be great if his first time on stage is pulling something that actually bloody works. I'm bullish on them. Well, that would be fabulous. But I'd say even without that, Cook, he has certainly had a strong quarter here to report. iPhone sales up 22%. Greater China, which was seen potentially as a weak point, up 28%. They're talking about India as a massive market.

9:01So people are positive around Apple and the stock's going up. Now, why are we talking about all this? Well, back to memory. Because even for Apple, they're talking about memory being an increasing cost component. So if you look at the bill of materials of the hardware they make, especially on the computer side, well, memory is becoming an issue there for them. So not something that's holding them back, but it's just memory is pervasively the choke point of the global AI industry. Well, Tim Cook, he is the supply chain guy. I've nearly finished Apple in China, which is a great book if people haven't read that book.

9:37I didn't realize how instrumental and how much he worked on the whole supply chain piece globally. It's a fascinating read. Do give it a go. But it's interesting to juxtapose this with what's happening in the wider economy because your stock markets have been doing quite well. And it may be a bit counterintuitive. We're talking about the Straits of Hormuz and we're talking about conflict in the Middle East and rising inflation and issues around oil. And still stock markets have been doing well. And it's really because AI is powering these companies to perform so extremely well that that is not just holding up the market but actually pushing the market.

10:16They're true bellwethers. But what people want to know, though, is, I mean, I guess when when will Wall Street meet Main Street? I think that's that's the challenge that most people have, because you can't have this forever growing market state when the real world hasn't quite caught up and inflation is going to is going to bite backsides. Because all the reading that I've been doing about how the Hormuz is actually going to affect us, it's not going to come this year. It's going to come next year. energy prices will go up, fertilizer and everything will start to impact the real world. And I don't know how that's going to marry.

10:52I don't know if you've got any thoughts on that. I think it'll happen before then. I think you're going to start to see inflation trickle into inflation sooner if there isn't a resolution. And obviously, if we want mainstream to do well, the general economy has to do well. And that is back to all the things we know. You've got to have well-functioning capital markets. You've got well-functioning commodities markets. You've got to have access to energy. We've got to have well-functioning labor markets and kind of all the hard work. Like even the magic of AI can't wish away all these other things.

11:19But it's interesting that, you know, in a situation where one would have thought, look, it could be a bit dour and markets could be down. Actually, they are in many ways up. Yeah. Well, let's stay in that world. I want to talk about Frontier Labs and politicking. I've had a bunch this week. I've seen a number of news stories about some shenanigans. So we've got Musk versus Altman, as I'm sure everyone has seen. My favorite quote so far is from OpenAI's lawyer, William Savitt. His opening statement was, we are here because Mr. Musk didn't get his way at OpenAI. So he quit saying that they would fail for sure.

11:55But my clients had the nerve to go on and succeed without him, which I'm sure was nice and baity for Musk sitting in the seat. We've got China's blog, the sale of Manus to Facebook. Now, you might know this, Maz, but I can't quite work out how the Chinese state has got such tendrils into a Singaporean company. It also felt a little bit tit-for-tat-ish. So the Americans accusing Chinese labs of industrial-scale distillation, i.e. how can DeepSeek do so well so quickly with an anthropic. So therefore, they blocked the sale of Manus. I don't know if there's any truth in that. It felt a little bit tit-for-tat-y.

12:33So you've got Google investing under 40 billion in Anthropic, as you've already mentioned. That's another interesting chapter in the circular financing compendium. And then quietly behind the scenes, Europe has been absolutely killing it. Ineffable, recursive superintelligence, some big deals, some big stuff. Is there a gap with all of this stuff going on and people falling out for Europe to slot in? I don't know. But where would you start on the politicking piece? As you say, it's been a super interesting news week. I mean, let's just pick up on the whole Manus piece, because you're right. People might recall Manus was one of the first agentic AI companies coming out last year, and it was a Chinese company.

13:13And they made a beautiful front end, agentic front end. And people said, oh, great, is that Chinese AI? But it was all anthropic under the hood, right? I mean, it was because they really thought about how can we make this not just something cheap and cheerful, but actually use the best components all the way up the stack to build a delightful user experience. It truly looked like something that had come out of Silicon Valley. Now, the founders quickly made a decision to say, look, we are going to relocate, re-domicile from China to Singapore, gives us more functionality. Which is common, right?

13:41I mean, that's quite a standard path for some large Chinese organizations out here. Yeah, I don't think we've seen it that much in sort of the startup scene yet. Right, okay. And there are certainly – this is part of what's leading to some of the issues you raise. What then happened was you had Benchmark Invest into Manus, which is a kind of U.S. leading venture capital firm putting money back into effectively the Chinese startup ecosystem. It was something that people had been pulling back for for a while, but it was felt, look, it's a Singaporean company now. We can put money in. And then there was a sale to Meta of that business.

14:24Now, as you ask, how can China assert that kind of legal supremacy, legal sovereignty over an entity that's now Singaporean? Well, I think a couple of things. Basically, what the Chinese are saying is that the IP was relocated and transferred to Singapore without approval. And it's subject to export controls. And there's a travel ban, there's an exit ban on the Chinese national founders. And so when you block the transfer of the IP and you block the movement of the people, it is hard to sell an AI company. So, yeah, that is. Wouldn't that have come up in DD? Wouldn't that have come up in, wouldn't that be a possible outcome if you're a large US venture firm looking at a deal like this?

15:16Well, the question is how much is retroactive? You know, I think you talk about whether this is tit for tat around distillation. I actually think that the parallel here, the analogy is NVIDIA. It's export controls both ways, right? The Americans are saying that you can't buy our best chips. The Chinese are saying, well, you can't buy our best engineers. Yeah. Da-da-da-da. Yeah. Yeah. And so, yeah. What else on this? So we've also got the Google Anthropic deal. Do you want to talk about that? Yeah. Yeah, we talked a bit about it last week when it was still at sort of at a 10 billion kind of quantum.

15:53And I think right after we had gone to tape the pod, it was then up to 40 bill with 10 bill down now and another 30. They're subject to some milestones. I just think it's exciting. I think I feel about Google a bit the way you feel about Apple. And I think it's super smart. I am I'm always kind of intrigued when you see pledges that start with up to so we will invest up to 40 billion. Well, is that 1 million or is that 40 billion? I don't know. Nobody knows. It just feels I'm still always a little bit kind of sniffing around for the for the bullshit PR of we're going to spend a trillion on CapEx.

16:32We're going to invest a 50 billion to this. I wonder what the reality will be. It just feels a little bit feel a bit nebulous, a little bit ethereal, all this kind of numbers being thrown around. Yeah, I think there's a structural cap on the ownership that the deal is allowing Google to take, which is 15%. I think Anthropic has been very smart in the way they have both Amazon and Google as cloud providers, so they're not locked into either one. But obviously, there's a very close relationship and collaboration between the two because Anthropic now does a lot of their work on Google TPUs. So they're using the Google stack.

17:05And everyone needs everyone to a degree, right? I mean, Apple still needs Samsung, even though there is a lot of, well, there was a lot of animosity back in the day. I mean, everybody still needs everybody to create a functioning ecosystem. It's probably a good thing. Yeah, I think it kind of ties everyone together, I think, in a healthy way. So we need to go into a little bit of what about us-ism. So how does this come back to Europe and what does this mean for us? We've just had some phenomenal announcements. I think you teed it up nicely. We've had Ineffable, which is David Silva from DeepMind.

17:39That's massive. It's massive. $1.1 billion round at plus$5 billion in valuation. In Europe. This is here. This isn't Valley. It's great. No, right here in London. He is just a fantastic technologist. I mean, he was the creator of AlphaGo. He was the architect of AlphaZero. I mean, we've watched those documentaries. We've seen and followed what's been happening. Now, his pitch to the world is, you know, we are not going to train another LLM on the corpus of the Internet. We're going to have reinforcement learning from scratch. So what does this mean? It's essentially saying the way that AlphaGo was trained, i.e.

18:21by teaching it the basic rules and then having it become world champion by playing against itself and learning the rules and figuring things out for itself. That's how we're going to create the next generation of AI. Can it even work? I don't know, it sounds bonkers, right? I mean, how is a model going to figure all that out with all the corpus of the body of world knowledge to train it? But if anybody has got the right to go and figure it out, it seems it's David Silver. And the who is who is in that round. Yeah, that's great. And Recursive and Amy, I don't know if you have any notes on those guys as well.

18:55Yeah, so that's more about AI that approves its own architecture. It's about, you know, once you have AI out in the real world, How can you make sure that it continuously improves and becomes better? We've seen a lot of failures in enterprise deployment. And, you know, we're trying to put AI in and actually make it work. And, of course, if you have a static model that can't learn on the fly, well, that's a huge impediment. And so Recursive and Superintelligence is making a pitch there. Again, it's a$500 million round. Love it. Love it. But then everything is kind of double, triple, quadruple in our world.

19:28I remember looking at our early notes when we first started our first fund, and I was looking at helping founders get to their first million. It just sounds so quaint now. It just feels so out of date, but there we go. I also want to pick your brains on some Frontier models. It looks like we've had a number of new drops. What's happening on the Frontier model base? Yeah, look, it used to be big news, big announcements when new models were out. And we've seen so many come out now. X, they launched the new version of Grok 4.3. On the 17th of April, you've seen the new Kimi K 2.6. GPT 5.5 came out.

20:10And then we had Opus 4.7 come out just before. So it's the densest release window on record. And so the cadence itself stopped being the story a couple of quarters ago. now monetization is and monetization is really kicking in i saw um i saw very quickly i saw a number of buddies now really loving uh 5.5 yeah i don't know if you've had a little play or whatever but it felt like there was a lot of love for all of the anthropic products and the direction and oh my god this is amazing whether it's talking about mythos or 4.7 or whatever but it feels like there was a little bit of okay they're not out of the race just yet so this is obviously still interesting stuff and everyone is still kind of cat and mousing here.

20:53100%. 100%. And isn't competition a great thing? I mean, I think, you know, Altman, he has declared, I think he declared a couple of code reds actually over the last 12 months. Who hasn't? There you go. Well, I think there's no code red at Anthropic. I think they've just been, you know, just doing the do. Yeah. But so that's been interesting. You're seeing the API cost, the token cost of the new GPT model from OpenAI, it's double the price of their previous model. So they're really putting on the screws. Now, Anthropic has been doing something maybe nefarious. I don't know if I'd call it, but they've been quite clever.

21:34I can feel you catching your words. What do we call this? What do we call this? Everybody that I know have been noticing what happened as they rolled out Opus 4.7. it's obviously frightfully smart and it uses more tokens. And so two things have happened simultaneously. It's more token hungry. And they've also lowered some of the thresholds for when you're kind of all you can eat consumption or what kind of limits kick in. So people's anthropic bills are just going through the roof right now. And again, we'll see whether you get some pushback because of that. But of course, it's so good. So we pay and we use it, right?

22:18Because it creates so much value. But it's just really interesting that they were able to grow at the level they grew without this. And now they're putting on some of the screws around monetization. It's effectively a price increase. Yeah. And we should see. I mean, the expectation is they could hit 100 billion in revenue this year. That's incredible. Well, yeah. I mean, let's talk a bit more about that later. What else is going on on the model front? Anything else? No, I think that's key. So the Grok 4.3 model, it's still beta only, and it's charging$300 a month for it. There are no benchmarks yet, and we don't have any separate API pricing as far as I know, so Musk is still working it out.

22:57We've seen a little bit of movement on Mistral. They have a new model out. You can buy it off their API server at a much, much lower price than Anthropic Sonnet model, and they've shipped their own Vibe agents as well. So Mistral is pushing things through. interesting i don't think i don't see them competing globally yet but i think for sovereign ai in europe there is a massive market for what they do yeah because it will at some point become a case of utility versus price i imagine oh 100 and as you can chop and change and as it does become more commoditized and as there is such healthy competition i wonder where the utility layer will stop and you go look this is what we need and therefore you find that the best price for that, but we'll see.

23:42Yeah, but then the value moves up the stack, right? I mean, as we can see, if you have the right harness, and with harness, I mean the thing like clawed code, the thing that sits on top of the model and really helps you make the most of it, the right harness can 6X the performance. Yeah. So we're talking about eking out 10%, 20 % for every new model release. Well, you can 6X by having the right scaffolding, the right structure around it. That's much, much, much, much more. So that's where you move from model to middleware, to application. Yeah, but my point is more a case about, you know, what does a business need?

24:16What does the utility need to do and to function? And at what point will you make a call? This is good enough. We don't need the latest, greatest. And I'm super intrigued as to where open source layers into that. Yeah, but I think you're sort of already given part of the answer there, right? Because I think you're going to be looking at lots of different use cases. And for lots of use cases, you don't need the most fantastic model. I mean, a lot of the stuff we do in business every day is not quantum mechanics. It's not rocket science. It's reconciling invoices. It just needs to work. It just needs to work, right?

24:49Or processing KYC documents or application forms. There's so much stuff that could be done without the frontier level stuff. So absolutely, that's what's happening. I've been keeping track of which IPOs are kind of going to happen in 26 and which are going to fall into 27. or perhaps not even happen. But SpaceX looks like it's still going ahead this year. What's the latest with their listing? So we know that the S1 has been filed confidentially and we're expecting it to go public mid-May. What does that mean? It's effectively the filing that has the data, the information that investors need to look at to figure out whether they want to invest in this IPO or not.

25:37And so why is that exciting? Well, it would be the biggest IPO in history. Nothing short of that. The target is somewhere in the window of$1.75 to$2 trillion, a raise probably around the$75 billion amount. What's UK GDP? Is what,$3 trillion? Yeah. Is that fish? If you squint? So this is$2 trillion in one filing, potentially. It's a massive number. And it's a massive number also when you put it in the context of the revenue the company is doing. Yeah. 2025 revenue was 15 to 16 billion. But it's Muskian revenue. It's Muskian revenue. So it's okay because he's going to do something in the future that we just don't know.