The AI jobs panic might be wrong

0:00We are now looking at the era of the zero employee AI company. Pulsia this week raised 30 million at a quarter of a billion valuation. Zero employees, nine agents. So yes, AI will be an augmentation and automation and efficiency generator like other technologies before it. But we can't find evidence yet that AI is displacing masses of workers. The data just doesn't stack up. And what people are figuring out is that actually the models are really powerful, but they need a harness to tame them. And you can take the best model from OpenAI or Anthropic, but you still need a lot of application layer on top, the harness, the context, the way you manage the model, the memory, all these things for it to do what you want it to do.

0:40And that is an application. Somewhere in this current cohort of the 15 to 20 AI native unicorns, one is going to blow up publicly by the end of next year, which is going to be a full on 72 hour catastrophic collapse. I think we're possibly thinking about this the wrong way. Instead of being busy regulating new things, I would try and figure out where can we deregulate to let European businesses be more successful. Hello and welcome to Upside, where every week we look behind the headlines affecting European venture. This week it's Mads and myself and what are we talking about? Well, we've got the AI infrastructure winners.

1:15We can now look beyond GPUs. Some good and ugly on the European sovereignty in action front. A zero-employee AI company raises 30 mil. Have a look at that. That sounds interesting. And the Pope weighs in on AI. Plus deals of the week, what to look out for in the week ahead, and some predictions.

1:45This show is not investment advice, and the hosts of this episode may be invested in the funds and companies featured. So Mads, my good man, now we can banter Now we're in, now we do the Prelim banterisation That VCs do so well on their pods What's cooking? Well I hear there's a new car out Did you buy one yet? Yeah, would you get one? No, what were they thinking? Listen, you're not a car dude anyway I'm a car dude I'm upset I'm upset, I think this is This is a reach too far And you can justify it however you like as in Ferrari have got to go and, you know, push the boundaries and maybe this is a new direction.

2:27But this is just looks like a disaster zone. Do you feel it's a Jaguar disaster? Yeah, I think this is monumental. I mean, obviously, whatever, shares slipped 75 % somewhere in that bracket on the launch of the Luce. But I think it's terrible. I mean, I don't think you can style it out in any other way. What else happened this week? What we won't necessarily dig into. we actually will dig into some of this. I was looking at SpaceX postponing Starship and I was just thinking, is this pre-IPO? Is there a reasoning for this? I don't know if you've got any thoughts on the SpaceX postponement of Starship.

3:05They cut the top end of the range from 2.0 to 1.8. I don't know whether the demand is going soft and they're like, well, the last thing we need now is a disaster. You saw Blue Origin, right? I know, it blew up last night. Yeah, I wish the 29th, the recording is the Friday the 29th, and I think it blew up at half 11 last night, yeah. Maybe they don't want one of those. But this is to be expected. I mean, you need these failures now in the early things so that you can iron out all the kinks, is my ignorant view of space launches. You don't need one the day before the IPO. No, and also my take is that unless they can get a spaceship up, A spaceship is the linchpin, as in everything, all of the future launches at the size and scale that Elon needs to make the model work all rely on Starship, right?

4:00Is that a fair appraisal? Yeah, I think so. So it makes complete sense just to kind of have nothing to see here. We'll work out the kinks. Everything's going to be fine. But at the same time, they're still only the last kind of training flights and the last worky-outy bit. So maybe, I don't know, probably not a bad thing to do. Did you see that Gary Lineker, for those not in the UK, he's a famous UK footballer. He's launched a VC firm, which I thought was kind of neat. Nice. Focusing on creatives. Did you see that? You're a football guy. Actually, he was a Spurs guy. Yeah, I used to like him. I've sort of lost my taste.

4:39Oh, okay. Is there a story behind the story? Possibly, but not for the show. Okay, I can see it in your face. The other thing that I saw was Kirkland and Ellis, who are the world's highest grossing law firm. They've put a half a billion dollars aside to build their own AI tech stack. So that's a big chunk of change to build their own AI. and maybe we'll dig into this in weeks to come but is this going to be something that we'll see more of more in-house more local on-prem provision do you in principle do you think we're going to see more of that or is this just pr noise where would you put this so here's what we're going to see and we've talked about applications as where the rubber hits the road for years and there was this furious debate about whether models are going to rule the world and applications are going to disappear into just some small sort of crud.

5:34It's basically just a database layer with an LLM on top, but that's going to replace all applications. And what people are figuring out is that actually the models are really powerful, but they need a harness to tame them. And you can take the best model from OpenAI or Anthropic, but you still need a lot of application layer on top, the harness, the context, the way you manage the model, the memory, all these things for it to do what you want it to do. And that is an application. And so what we will see going forward is, yes, we'll see a ton of vertical applications in lots of spaces that the frontier labs can't occupy themselves.

6:11And as we have always seen, we will see smart firms at the frontier investing in in-house technology. it's many years since many of the biggest investment banks ended up having more software developers than they had bankers, right? Because they figured out quickly that technology is really, really important to their business model. The lawyers are figuring that out as well now. And so they're piling in. So actually, I would say what's happening here is exactly what you would expect if you expect applications and the application layer to be really, really important for value creation. I mean, half a bar is quite a big chunk of change.

6:48It's meaningful. Yeah. I mean, you don't know what is just getting the headlines and what the reality is. And also, I would assume that if you've got a bank of engineers building whatever in old school times, that would be a big chunk of change over a period of time as well. So maybe it's horses and horses. Yeah, 100%. And you think about how productive developers are today. You can build a lot of software for 500 million. yeah well see i think i've got my theory is and we kind of spoken about this before is that we'll see a lot more on-prem and we'll see a lot more local provision to build out the actual application layer i think this is going to be i think we might see more of this would be maybe i'll put that in one of my predictions of the week not today about some of it today right we've got deals of the week and other stuff will definitely be a recurring theme yeah and uh last on my not to go into list but I wanted to just throw it into your foray, is the Deal Room Ecosystem Report.

7:48I know you're not a report fan, but they've mapped out 325 cities. They call them ecosystems, but effectively it's just it's 325 cities across the world. Looking at investment, innovation, talent and outcomes. I don't know if you pulled out any nuggets from this. I did see that London has regained the top spot from Paris after losing in 24. Israel slightly down. Most of the Chinese cities apart from one slightly down americans obviously still owning the roost did you see any do you see it for starters and anything in there that you'd pull out or uh no i think anecdotally it confirms what we have seen on the on the ground which is that london is really heaving right now the volume of deals we're seeing the tat the quality of the teams the activity is is the highest it's ever been and that's a good thing yeah yeah um i've seen a few linkedin Europe bashing articles this week, comparing, for example, Anthropics Rays to European Venture.

8:42And you kind of can't compare. They're not comparing oranges and oranges. And I think it's interesting to see how we take the applications and how Europe makes its own, carves its own path in the world as these things settle in. And let's settle in to what we are going to discuss on the docket, Starting with, we've got some interesting AI infra stories this week. Now, for most of our discussions about AI, we've basically been looking at NVIDIA and all the GPU stories. Now, we've got three, four different infra stories, starting with memory. So let's just pick up on this one first. capex it looks like just to tee it up it looks like capex is going wider into memory software and beyond and starting on the memory side we've had a couple of companies cross the trillion mark what start give us the insight into into what's happening in memory man let's start there it's super interesting how memory has suddenly become the hottest thing on the planet and you're absolutely right we saw micron and sk hynex cross the trillion dollar threshold this week memory is It's one of those businesses where, you know, for the history buffs will remember that Intel used to be a memory business.

9:58And they got out of memory in 1985 because memory was always a very low margin, highly cyclical, very CapEx intensive business. And so in the early 80s, Japanese suppliers like NEC and Hitachi and Toshiba, they crushed Intel on price. And then DeGro basically said, listen, you know, we've got to get out of this. And that's why they pivoted to CPUs. And then they had 30, 35 years of excellent results being the world's dominant CPU maker. So that sort of saved them. For 40 years after getting out of memory, Andy Grove and Intel were right because it was a famously cyclical business. It was boom and it was bust.

10:39And every time there was a boom, everybody would pile in and add capacity. And then when it was bust, all those people that added capacity during boom, they would die. So companies that were, if you look back 25 years, come up with the major DRAM players, Toshiba, Powerchip, Promos, Nanya, Inotera, they're all gone. It was the same pattern every single time. And there were sort of three survivors left, three survivors standing. It was SK Hynex and Micron that are specialists, and then Samsung, which is just generally a powerhouse. They could do a lot of things and have a fabulous balance sheet.

11:18And then suddenly AI and GPUs happen. And this has been the absolutely nuts pattern we've seen over the last two years because you've had a compounding growth in demand for memory. First of all, GPUs are famously memory hungry. You need a lot of memory to run the big AI models and it sits in the GPUs. And so you've seen an explosion in the shipment of GPUs. In 2024, there was 3.4 million units shipped. This year, we're looking at 8 to 9 million. So you've seen between a doubling and a tripling of the capacity, growing roughly 50 % a year. But what's happening in the meantime is that as AI models have gotten bigger, the amount of memory you put on each GPU has also grown by about 80 % a year.

12:10So you've had more units needing much more memory. And in a capacity-constrained industry, that's gone bananas. How many new players come in this time, Matt? Do you know? None. It takes years to stand up this stuff. Yeah, yeah. And so what we've seen historically is that people pile in in boom times, and then they go bust in bear times. And that's why nobody has really been adding new capacity because it takes too long. And everybody is looking at this and thinking, even if we add capacity for the stuff that's in most demand right now, well, guess what? We're going to be at a next generation of memory two years from now when that capacity is being stood up.

12:47I thought the Chinese might. I thought the Chinese, I mean, it might be something that, obviously they're looking to be like a properly self-contained sovereign nation. I thought we might see something. They have been hit by the HBM, kind of the high bandwidth memory shortage as much as anybody, because even if they can stand up and create GPUs, they've not been able to give access to the memory. So it's a gap that's been difficult to close. It's not going to close this year. It's probably not going to close next year, maybe not even in 2028. And so that's what we're seeing in these memory makers and why their revenue is just going through the roof and margins are exploding in ways we've never seen before for memory.

13:25At the same time, something else has happened. The capacity discipline almost never held, as you said. Historically, someone always defected and said, okay, great, it's boom time. let me add more capacity. That's why it was a dreadful business for 40 years. Now, today, you sort of have an agreement amongst the three that nobody is piling in big time. And there is a suspicion that the buyers of the memory, so the semiconductor chip, the GPU makers, they don't want there to be masses of new capacity built because they're worried that the memory suppliers are just going to go bust in a few years.

14:05So you almost have a bit of a tacit agreement. There's a super interesting pattern. And it seems like margins are going to be high in memory land for some time to come. My theory before I want to talk about Snowflake and Salesforce in a sec. But my theory is that we are going to not need the capex spend. I've said this before. We've had a bit of a tussle over it. But I don't think we're going to need the 100x growth in this whatever capex spend from whomever. we're going to need 5x, 3x, 5x, something like that that we are currently seeing. But it feels like, I mean, obviously it makes sense for the RAM people and the memory people to not then, you know, go crazy on supply.

14:43But I'm just wondering, in general, all of this CapEx stuff, all of the stuff that we're talking about in this segment, I mean, I can't see us needing this ad infinitum. Does that, do you agree? No, but that's a truism. So look, that's a truism. Look, I think as with memory, if you look at the wider world, Every time you have a space that's hot, everybody piles in, everybody invests, and at some point you're going to overbuild. It's just human nature, right? There's always going to be somebody who defects and invests more and wants to try and grab that marginal bit of capacity and a marginal bit of market share.

15:19And then the market turns or you reach a new plateau. So I think that happens for sure. You're saying we're not going to need 100 times more. I think that's fair. I mean, if we are thinking about hyperscalers now, if we just take those 700 billion of CapEx investment, I mean, they're not going to need 70 trillion, right? I mean, they're not going to need 100 times more than where they are. But even if it's five times or 10 times more, that's still a massive explosion from where we are today. Yeah, yeah, yeah. No, listen, I was looking at the rough numbers. I put it in the three to five bucket, but I'm always a bit conservative about these things.

15:49I mean, back when this whole thing started, we were all talking about Pascal's wager. if you remember that, right? Kind of the whole idea of, you know, am I praying to God? Does he exist? If he exists, it's good. And if he doesn't, it's not going to hurt me. So I might as well. And there's been a sense that for many of the hyperscalers, it was almost like a Pascal's Wager. It's more important to do it and be in it than if the train does continue and you're not in it, you're sort of your dad. Let's talk about Snowflake and Salesforce. So Snowflake popped 37%. Even Salesforce beat estimates. And this is how I'm positioning this as part of the broader kind of foray into CapEx and AI infra, the whole infra piece.

16:34So we're now looking at more hardware provision and more software provision. So everybody seems to be, you know, all boats rising with the tide or whatever the phrase goes. Tell us a little bit about what's happening on the software side. Yeah. So we talked about Salesforce last week, kind of previewing that they had earnings coming up this week. And some of the things I said to watch out for is, you know, is this a cesspocalypse redux? Is agent force actually growing? Or is this emissary hidden in the numbers? And I think it's kind of cesspocalypse. If you look at Salesforce, they're down sort of 30 % a year.

17:08They had a pretty good quarter, but they're still, the shares are still soft. And I think there is a sense that, look, there will be a potentially seat-based erosion because the whole notion is agents, they don't buy licenses. And if we need fewer people to do work, we need fewer Salesforce licenses. So the stock has gotten punished for that. So that's a shift. I mean, effectively, they're pricing a shift in the model from SaaS to usage. Is that a fair analogy? That's exactly right. And we have seen a growth of AgentForce, which is their new agentic solution. The headline says that agent force has grown 200 % year on year.

17:46But headline numbers, headline revenue is at about a billion, which is still relatively modest in Salesforce terms. It's a much bigger company. Look, there are a lot of questions around that business and kind of how well it's going to do in the future. Now, at the same time, over at Snowflake, we're seeing exactly the opposite story. Because Snowflake, they've had a phenomenal revenue growth in their new product line called Coco. Now, what is Coco? It's CortexCode, which is an agent data assistant that lives inside Snowflake. What does that mean? Well, it essentially means I've got Snowflake, which is a huge data repository, and Coco is effectively like a cloud code or kind of an AI I can use to ask questions of my Snowflake data warehouse in real time, and then it'll write all the SQL queries.

18:38It's not going to hallucinate. and it'll produce proper insights for me on the back of that. And again, this is back to where we spoke at the top of the segment about the value of applications. Coco runs on top of Opus from Anthropic, but it's not just Anthropic. It's not just the raw model. It's all the intelligence and the harness Stoflake is putting on top to make sure that it can go in, query the database and the data warehouse in the right way, not hallucinate any rows or columns, make sure that the data that comes back is actually, it's a proper answer because it has to be deterministic.

19:17It can't hallucinate anything along the way. You query your data warehouse, you want to make sure that the answer you get is true. And so that's just been a resounding success. Revenue has already eclipsed a billion and that's within a few months of launching it. And the CFO there, Brian Robbins, he said that Coco has been the largest driver to the increase in the forecast of the year. So they're hugely bullish on AI's impact on the business and the market is really liking it. I've not heard of Coco. I'll have to have a play. Very quickly, I want to talk about the S &P rule changes. Now, index rules, and I know this is American and everyone needs to forgive me, but I think this is important.

19:55So index rules are super dry. They are not sexy until you realize how this is going to affect all the IPOs lining up in 2026. and this automatic buying piece for all the indices, mainly for retail. So the S &P is about to halve the seasoning period and waive profitability for mega caps. Sounds dull, bear with me. What does this mean for the IPOs coming up? I think this is going to be mega interesting for SpaceX specifically, but how would you break down why this is meaningful? And then maybe there is a European angle, I'm not so sure, but very quickly, the S &P rule change what's cooking here.

20:32And a massive part of the stock market is today traded passively by indices, either directly by robo-trackers that just buy the index passively or by other investors that are looking to track the index closely because they're measured on their ability to beat the index. And so the S &P 500 is an index, it's the world's most important index, and so many people are tracking it. And the way to track it, you can try and replicate it in other ways, but of course the only way really to track it is to buy the shares that are in the index. And so if I'm in the index, a lot of people have to buy my shares because I'm in the index.

21:06And the way to track the index is to buy the shares in the index. And therefore, if I'm not in the index today and I get to join the index, a lot of people that otherwise might not have bought my shares need to buy my shares. And that becomes automatic demand. And so what will happen is when companies, they get admitted to the S &P 500, there's a lot more demand and the share price go up. And reversely, if they get thrown out, well, there's a lot less demand than the share price goes down. And of course, if you are SpaceX and you're looking to IPO with a massive IPO and could command that$1.8 trillion market cap, if you have a lot of demand lined up from people that must buy your shares because you're part of the index, that's really valuable.

21:46And so that's what the whole discussion is about. I was also quite keen to understand all the lockups. I haven't really dug into the S1 or any of the details around the SpaceX listing, but I'd be so fascinated to see what comes in, how that affects markets and pricing, and then what happens as lockups unlock and people sell and maneuver. I think it's going to be a very, very interesting year in 2026. Yes, and will the pace of rapid pulling out of hats be able to keep up with the end of those lockup periods? Yeah. Right? Because, I mean, no, but listen, Colossus, I mean, you think back to the whole story about X and XAI and how, you know, Musk obviously, he stood up XAI and then merged it with X and then merged that with SpaceX.

22:41and it was just seen as a big CapEx drain, but he built those two Colossus data centers and he was just able to make a deal with Anthropic to sell spare capacity to them. Isn't it three data centers? Isn't he got three? It's just Colossus 2 and Colossus 1. Maybe there's a third one on the way. The important thing for the story is that he was able to do that deal and that sort of doubled the revenue of SpaceX overnight because something else he did over in X to stand up a data center, he figured out he didn't need. So it's those rabbits that, you know, where does it come from? Selling the future, baby.

23:14Always selling the future. FSD's coming. Don't you worry. Well, it turned out the future became the current really quickly, right? Because this is real revenue that now is flowing in. And the question is, what other rabbits are awaiting that could grow revenue by 10, 20, 30 bill over the space of a short amount of time? And if he has those, well, things might look better. It's all right. We'll get SpaceX to buy all the Cybertrucks. We'll get Tesla to buy all the AI. So it's fine. We can just keep the whole circular trade circling. But that's not what happened here, if you think about it. No, no, no.

23:49I'm being facetious. I'm being facetious. And that's fine. It's easy to do. It's easy to poke fun at all that stuff because it is a bit funny. But I think it's some of those wrappers that we sometimes don't see coming that are there suddenly. And then they sort of change the game. it just always feels a little bit of kind of post-rationalized nonsense to me when when you look behind some of the dealings with elon and then all of a sudden he's a magician because he's managed to whatever sell sell compute to anthropic i think i think it's not so much judgment personally but then i'm a little bit down a little bit down on the man for reasons that aren't always valid let's not go there because there's no point uh i want to talk about very quickly opus 4.8 it's out It was out two days ago, I think.

24:34We were on the 29th, I think. Was it 28th? I thought it was 27th. Anyway, whatever. It's a newbie, newbie on the block. Is it a thing? Is it a step change? Looking very, very quickly, there's some of the benchmarks and some of the measures. It looks like it's beaten 5.5, 3.1. They've teased Mythos as the next frontier model, but what's the scoop with 4.8? Look, there's just a lot of excitement for the model. 4.7 got a bit of flack. The new model of people are saying, you should just call it the Opus 5. It's that good. Some people are saying we still prefer OpenAI's codex, but the Anthropic model 4.8 is the best and it's just fabulous.

25:14And people are generally really, really excited. The vibe checks are great. If we look at what's happened at the same time is they've closed this new race and they've now eclipsed OpenAI as the most valuable frontier lab. It's kind of a$965 billion valuation. So it's a nice headline to come out with that, come out with this new model at the same time. Run rate revenue looks like they might be at$50 billion run rate by the end of the quarter, and it'll be the first fully profitable quarter. And as I said last week, I think Anthropic is... It's incredible. I think it's the most important company in the world right now.

25:52But who would have thought? I mean, it was only yesterday that OpenAI was the wunderkind. And it's just, I mean, who knows how things are going to leapfrog and think the world will change. But that just seems casmic. Listen, I think Anthropic is bringing the best out of open AI. And they are back and they're working it. And Codex is a great harness. And what we've seen is, so one of the things that's been discussed this week is who has got access to Mythos. When it came out, of course, U.S. government, large U.S. banks, U.K. as well. The AI Safety Institute was one of the early labs to get a test run of the model.

26:29UK banks have had access to or are getting access to methods to help proof their security. Not so much the EU. And of course, that's been a huge discussion point. Sam Altman, on the other hand, has been quick to say, well, we will provide you with our side of the model. Always the opportunist. Well, there you go. Bless you. But isn't that great that competition is keeping people honest? I think it's a great thing. And we'll talk a little bit about kind of European sovereignty, I think, is coming up. But we're touching a little bit on it here because obviously these frontier labs are structurally and systemically so important because the technology is so powerful.

27:07And what is a whole continent to do if it can't access the good stuff? Yeah, we're going to go beyond talking about it in the sovereign sense. We're going to be talking about the Pope talking about it with an anthropic rep. I mean, Chris was by his side. So there's a there's a very European angle here. But let's talk about the sovereign piece right now, because there has been this week some articles and some news stories about European sovereignty. Not all good, but there's three. Let's start at the top of the list. So the Dutch government blocked an acquisition. They stopped the IT giant, the American IT giant Kindrel from acquiring Solvinity.