Even banks and hyperscalers are now sounding the alarm about the AI bubble

KETTLE From international banking worries to the market state of canary-in-the-coal-mine Oracle, the AI bubble is sure looking taut. The Bank for International Settlements, often referred to as "the central bank for central banks", said in a report at the end of June that it was worried the AI bubble was nigh on to popping and taking the global economy with it. Oracle, the hyperscaler with arguably the largest exposure to the AI bubble, has lost more than 40 percent of its share value in the past month and recently outlined all the ways it might suffer if this whole AI thing doesn't pan out. If you ask our systems editor Tobias Mann and senior reporter Tom Claburn, those factors and more make it seem like the AI industry could be on the verge of a massive contraction, and that's the very thing they chat about with Kettle host Brandon Vigliarolo on this week's episode. You can listen to the latest episode of The Kettle by clicking on the player above, as well as on Spotify, Apple Music, or YouTube, or read the transcript of the latest episode below. It's been lightly edited for clarity. Brandon (00:05) Welcome to the latest episode of The Register's Kettle Podcast. I'm Brandon Vigliarolo, and this week the topic is one that neither we nor the investors of the world can seem to stop thinking about, and that's the possibility that the edges of the AI bubble are beginning to come into focus. I'm joined this week by systems editor Tobias Mann and senior reporter Tom Claburn to talk about the latest news in this space, and there is a lot to talk about, so thanks for joining me guys. Tom Claburn (00:30) Yeah, thank you. Tobias Mann (00:31) Good to be here. Brandon (00:32) So let's start with a piece our UK editor Paul Kunert wrote, and he couldn't be here for this episode, so we're going to have to pick it apart in his place. And that's the Bank for International Settlements' report that it's worried that the AI bubble could pop and take the global economy with it. Tom, can you explain a bit about what was in that report and what they're worried about? Tom Claburn (00:51) So the Bank of International Settlements is sort of like a central bank for central banks or the UN of finance organizations. Brandon (01:02) So their opinion matters, basically. Tom Claburn (01:13) Yeah, it's not like some person, a blogger having an opinion. This is a big dea. The report raises the specter of all the previous financial manias like the British railway mania of the 1800s and the canal investment bubbles and the dot com bubble. Brandon (01:26) Yeah, the dot com boom, obviously. Tom Claburn (01:28) And it draws a comparison saying that all these things shared a common trait, which is that they attracted a lot more capital than the resulting industry could actually produce. And if that's the case with the AI bubble, there are going to be a lot of people who have invested billions in capex who are not going to see that money back. That's presumably not really the best investment scenario. And I think that that's on a lot of people's minds as all these major hyperscalers keep pouring money into this, and in so doing, raising all the costs for the equipment that they're putting into place, making it harder for everybody to get RAM and even consumers to, in the end, buy computers because there are just no supplies of RAM left. Brandon (02:08) My wife just bought a MacBook Neo the week before the price went up. And it was like, oh my God. I just sent her an article, and I was like, "Well, you should be glad you bought that now," then she had some issue with the payment and she was worried that it wasn't going to go through and they were going to reset the order and then she had to pay more. But luckily that got cleared up in time. It's affecting everyone. To put some of these capex forecasts in perspective, I think Amazon is saying that it's planning north of two hundred billion in AI build outs this year. Microsoft's looking at one hundred and ninety billion. Google a hundred and eighty billion. Meta one hundred and forty billion. I mean this is a lot of money being tossed around for potentially no returns, right? Tobias Mann (03:02) Well, some of it is empire building with the major hyperscalers. Because if you look at their business models, they can afford to lose a sizable portion of that. Their investors aren't going to be happy, but their investors won't be happy if they're not spending that money either. It becomes a catch-22 for them. But in the event the bubble pops, you can bet Google, Amazon, Microsoft are going to ultimately be fine and have built their empires. It's everybody else that is the problem as far as I can tell. Brandon (03:40) Right, which is what the BIS hit on, right? A lot of that report kind of focuses on what could happen to the hyperscalers, which like you said, they'll survive. But it's all the people sort of downwind of them, not just consumers and not just small equipment manufacturers and small businesses that rely on their components and their cloud. I think the report mentioned, Tom, like suppliers, construction firms, all these different people who could potentially be affected if these projects collapse. Tom Claburn (04:11) Yeah and it's not just the people in the industry that are building these datacenters, it's the assumptions that rest of all these customers who are going to be then buying these AI services. What if they don't come or what if they have a different idea of what they want to spend? And you're starting to see that with a lot of pushback from people just working on another story about that includes some comments from Palantir CEO Alex Karp, who's talking about how enterprises are very unhappy with this very gate-kept world of frontier AI where there's very little transparency and control. Ultimately enterprises skew towards being able to control their spending, being able to switch providers, and it's not always possible at any given time, but that's what they're going to push for, and so there's going to be a lot of pressure on everybody in this ecosystem to deliver open source models, affordable models, pricing that's predictable and not have this crazy situation that you have now with someone like Anthropic where things just get turned off, or you try and do a security-related query and it just says, "I'm sorry, Dave, I can't do that." It's not really a viable long-term scenario that people are going to use this on a regular basis. Brandon (05:26) Right. I mean, and that's the same with token costs keep fluctuating and going up. I think Karp in that interview, which I think aired on CNBC on the first of July, he was talking about another big thing, and this has been a concern since the dawn of large language models, is where is the data? If you're a customer and you're basically using one of these big frontier labs to host your stuff or to process your stuff, I mean, they're ingesting all that knowledge, right? Whether we can debate over how they're using it, and their terms of service are always kinda very specific, but there's a concern. If you're giving your company secrets over to these large firms, that there's the potential that they could benefit from those things too. It just seems like there's a million and one possibilities. Obviously Karp was pushing Palantir as the solution to that. Tom Claburn (06:30) Right. Brandon (06:31) Whether you believe that or not is another thing altogether. I mean he made a lot of really good points in that interview. He called himself Crazy Karp. Every time I hear him speak, I always am kind of reminded of just how manic he can be. But he always makes some pretty good points. Tom Claburn (06:46) I think ultimately if you look at what Apple is doing with its private cloud computing, which Google has also copied, probably recognizing that that's going to be a real thing, ultimately a lot of these workloads are going to have to move toward that model where the data is kept out of these large frontier clouds and is processed to some degree on premises or in a datacenter that the company has control over because it's just not acceptable to send off all this sensitive data and hope for the best. I think Apple's playing a good long-term game with its hardware story and I think that there will be other players in this space that have similar offerings. Brandon (07:22) They're going to have to. I feel like there is a lot of souring on companies like OpenAI and Anthropic and those frontier labs. A lot of people are just kind of fed up with the inconsistencies and what have you. Speaking of the big labs, OpenAI, I feel like in this current kind of bubble environment that we're living in, OpenAI is kind of inextricably tied to Oracle, right? Through some of the big deals they've made. And it was one of the big hyperscalers I didn't mention up above because Tobias, you wrote a story this week about an SEC filing from Oracle recently that basically is an outline of all the things that might go wrong. And it's extensive. So what were they talking about? it was huge. Tobias Mann (08:24) It was a long list. So in the view of looking at SEC filings on a fairly regular basis, you see a lot of boilerplate. This was not boilerplate. This was actually fairly specific to the challenges that they could potentially face. OpenAI has entered into this partnership, initiative, whatever you want to call it, called Stargate, about a year and a half ago now, and Oracle was one of the major founding partners alongside SoftBank and MGX. As it turned out, Oracle was going to be one of the primary financiers of this datacenter build-out. And to the tune of $300 billion of this $500 dollar project over this indeterminate period of time. The numbers just keep getting bigger though. Brandon (09:28) Right. And I think Oracle was mentioning, I think last November, about they were going to have to borrow $25 billion a year to afford to pay for this, right? And that was again, without any guarantee that OpenAI was ever going to turn enough profit to pay back its share. Tobias Mann (09:37) Right. OpenAI, in case anybody's forgotten, can't pay its own bills. It has to find somebody who's willing to provide cash in exchange for equity or debt. So it's the kind of thing that you find yourself in a situation where you are renting property because Oracle doesn't own its own datacenters for the most part, and then leasing it to a customer who says, "Yeah, I'm good for the rent. it's going to be ready in 2027, 2028 right? Okay, yeah, we're going to need a lot though. Does that work?" And just going ahead with it. Like this is totally going to the company that doesn't make money and is entirely reliant on venture capitalists to fund this, is totally going to be able to pay their cloud bills. And I think anybody who's been paying attention over the last year since Stargate was announced has been scratching their heads going, how exactly is this supposed to work? Brandon (10:54) I always cite this $300 billion of investment with OpenAI as a prime example of just the sort of absurdity in AI financing right now. It's just top of the list for me whenever people are like, "Well, what's wrong with AI financing?" I get into conversations so often where Oracle and OpenAI come up because I'm just like here's a picture of this. This is the quintessential example of the insanity of this. Tobias Mann (11:21) And so now with this SEC filing we now get Oracle saying what we've all kind of been asking, how is this risk even remotely manageable? And the reality is it does really feel like Larry has put his fingers over his ears and is going "la la la la la, everything's fine." And they put themselves in this position where they're kind of damned if they do, damned if they don't. Brandon (11:51) I think you said that exact thing in the story, right? If you don't build this empire, your investors are going to be pissed. So you gotta build the empire, but there's a good chance you're building it on a swamp, right? Where foundations are not going to be stable. Tom Claburn (12:04) And every so often you hear stories like the one I think came out in the Financial Times today where about OpenAI floating the idea of the US government taking a five percent share in the company. I suspect that that's the worst case scenario for a lot of these companies. Like, "Well at least we'll get the government to bail us out and we can shed the risk and let the public take care of it." But I don't know that that's really going to be a very appealing option. Brandon (12:31) I mean you can paint it as a sovereign wealth fund, right? That's what it seems like they're all doing. OpenAI was like, "We're doing this for the public good, right? So you can all benefit from our eventual collapse." Tobias Mann (12:41) Right. The problem with the bubble is that when it pops - I think at this point it's a question of when and how bad, not a matter of if - Oracle could find itself holding the bag because it is signing these commitments with all of these datacenter partners, Crusoe being one of the big recognizable names in this, the builder behind the Stargate facility in Abilene, they have made all these commitments and they're in a position to honor those commitments, but it's all based on a promise that OpenAI is going to come through and actually pay for it. If OpenAI can't pay its bills, the other $155 billion of committed capacity, or promissory capacity that they're entitled to provide, that's probably also not happening. And so it's just like I said, it is like leasing property to release to somebody else who may or may not be able to pay it, but probably can't. Tom Claburn (14:00) So it sounds a little bit like the We Work business model. Brandon (14:05) We all know how well that went, right? So let's talk about some more of the specifics they cited, because there was more than just the ties to OpenAI that were risky. I think you mentioned there's no guarantee that those releases are going to get paid or even renewed, right? So they might be on the hook for then unpaid leases that get abandoned. So you have a whole bunch of extra compute. Tobias Mann (14:31) Right, right. OpenAI is arguably the AI flag bearer of this whole bubble. And if they can't pay their bills, like I said, who is going to take that space? Who is going to be able to fulfill that commitment? So there's definitely that, they might be okay if OpenAI says, "Hey, we can get a big better deal over here," because that seems like a very Sam Altman thing to do. But if they actually can't pay their bills, it is an open question of who is going to take that capacity. Brandon (15:10) Yeah, probably no one, right? I mean Anthropic? Who? xAI? xAI can't even fill their own datacenters with their own loads, right? They're hearing all the time about like how I've heard multiple stories recently about how xAI is leasing space to other AI companies 'cause they've got all this extra hardware lying around that they're not using. Tobias Mann (15:27) They thought they could engineer their way out of the heterogeneality problem they deployed, Colossus, and they used a couple different generations of accelerators, thinking that they were going to be able to make it work. Turned out to be really rather problematic and it just became easier for them to go build it properly with Colossus too and lease out that original space. That's my understanding of how that actually played out. Brandon (15:34) Okay, gotcha. That's good to know. I didn't understand the specifics there. But let's see, what else was Oracle warning about? Needing to balance their demand estimates properly, I think. Which again it seems like be a function of the space and whether or not those leases are going to paid. And then there was down to the fact that they might have trouble just getting the datacenters they need to build built in the first place. I think there was mentions of permitting. The worry about a growing number of moratoriums around the country, power sourcing issues, all of this kind of combines to not look good for Oracle. And like you said, they could very well get stuck holding the bag for this. Tom Claburn (16:36) There's certainly a political sort of populist sentiment that's growing against AI datacenters that makes this even harder in a lot of communities. I mean there are a lot of places where people just don't want this to happen and that's not something you can just sort of wish away. Tobias Mann (16:56) Yeah, there's a lot of factors that Oracle has no control over. Or any of these big hyperscalers don't have a lot of control over. Power is one of the big ones that, especially if they're leasing, they might sign a lease agreement with a company like Crusoe who says, "Yeah, we can build that." But then that Crusoe may run into permitting issues. They may not be able to secure adequate power. The utility boards may say, "yeah, no." And if they push ahead, they could then alienate themselves and create controversy that is too much for the customer, Oracle or whatever hyperscaler that's trying to lease it, to even touch. If they say, "Yeah, we can build it," and then don't say that they're going to have to use a bunch of portable gas generator plants outside of it that are just belching smog and just completely alienate...or is completely incompatible with their sustainability goals, it just becomes this situation of just, "Yeah, we thought we could get capacity, but we can't, and so we can't actually realize the revenue that we made all these decisions based on." Brandon (18:00) And it seems like Oracle investors are already starting to realize that maybe before the company did, right? The stock's not doing great, right? More than they're down, I think, more than forty percent in the past month, is that right? Tobias Mann (18:25) Yeah, since they reported their Q4 earnings earlier last month. And at that time, even then things were, their stock was slipping pretty precipitously in spite of the fact that revenues were up and their profits were healthy. The impact and kind of exposure that Oracle has is arguably among the worst of the hyperskillers, because as I was saying earlier, Meta has a massive business for advertising that can happily run on GPUs. It may not be as profitable, but they can find a use case for it. And arguably that's where most of their money is coming from right now, anyway. All of the other cloud providers have similar business models to that effect that will carry them through, even if the big promise of AI turns out to be even more of a nothingburger than 5G Brandon (19:28) So I mean, so what so what are we seeing? Is Oracle the canary in the coal mine for this whole thing, so to speak? Are we looking at the edges of that bubble starting to form and firm up and get closer to popping? What do you guys think? Tom Claburn (19:45) I can't think of another company that is exposed in the same way. There just are a lot of sort of things that need to align for things to really pop. I mean right now there's still investment happening and talk about the golden goose of AI. But I think that give it another quarter or two and there's going to be a lot more pressure. And I think that the geopolitical situation, a lot of things are just not going along in a way that look good for things. And I think we may see a few of these companies really feel the heat soon. And a lot of this also depends upon how long the big frontier companies can sort of sustain the illusion that this is really something that's affordable and accessible to everyone before they have to raise prices to really account for the cost of business. Brandon (20:46) I mean, they've already done that, right? I feel like we've seen token costs go up quite a bit recently because they're not getting enough uptake or they're starting to realize they've got enough there that people actually want to use this. But I mean, six of one, I think when you look at it in terms of what companies are saying versus the reality. Tobias Mann (21:04) I would jump in to say I think Tom's onto something here. I think we're coming up on the squeeze where the need to demonstrate profitability is going to become a major deciding factor of how much longer this bubble lasts. Because a squeeze can go two ways. You squeeze, you raise prices, and you see how sticky the products have become. If you're lucky, they're really sticky and people are going to put up with higher costs because it has wormed its way into their lives or their business practices and they can't actually walk away from it. The risk is if you have to raise prices too high, you can run into a situation where the market just can't bear the burden. And that I think is where we reach the limit of how big this bubble can grow. Tom Claburn (22:06) Right. And you start to think about how often you hear about Chinese models, it's a direct reflection of that. And when prices go up for OpenAI and for Anthropic, and all of a sudden people are going, "Those Chinese models are not so bad," that's telling you something. The shift towards open source or whatever is going to push against the ability of these companies like OpenAI and Anthropic to generate a profit. And their hope seems to be, "Well we'll just get the government to ban those everything that's not competing with us," but I don't know how well that's going to work out. Tobias Mann (22:46) Well, especially when, if you're a business you decide, I'm going to go build my new cybersecurity suite on Mythos. And you're live for two weeks and then the US government can come in and say, "Well, actually no," and now your company just implodes. I mean talk about risk. Brandon (23:05) Yeah, and that ban lasted what, two weeks, so in the meantime, your company may have collapsed, right? And all of a sudden now it's back and it's available again. There's no reason to believe that the government won't keep being fickle about this as new advancements come up or as the next company says, "Well, we've got this amazing new model that can do this amazing new thing" and the government's like, "Well then we're not going to let you release it until we get our hands on it and test it and et cetera. Tobias Mann (23:33) Yeah, it's going to be interesting over the next couple of months as to how this plays out and the way the market adjusts to this kind of new reality. I've said it before, but we are starting to see the hardware change because the cost of running these models is just so intense and you can't run them at a loss forever. Tom Claburn (24:01) And you also see this in the software too. I mean, everything that I see in terms of development that has to do with how you harness and use a model is geared towards using it less. These are innovations like, "How can we capture more context so that the model doesn't use as many tokens and how can we basically reduce our spend?" Brandon (24:26) I don't know, it's just kinda funny. How can we use this amazing thing that we swear is going to transform our business less? Tom Claburn (24:29) Right. How can we use it less? And people are talking about "Well, you don't actually need a frontier model for more than 25 percent of your queries.That kind of parsing of how you want to do your workloads is very telling. And if the big companies aren't paying attention to that, they're going to find very quickly that someone else has come up with a way to sell this at a much more transparent, affordable, and reliable way. Tobias Mann (24:59) It can't be understated how complicated this technology is. From the ground that it's set on to the facilities that are built, to how you provide power to it and how you run those services. This is just insanely complicated. So the number of levers to pull and buttons to press in order to optimize this are many. The problem is that's a slow process that so long as there's lots of money doesn't typically happen. Once the venture capitalists stop dumping cash into this, I think we'll see that a lot more. Whether that's enough to sustain it, it's so hard to say. Brandon (25:48) Yeah. I mean we'll see, right? Like you said, we're coming up on the squeeze here. A couple more quarters or maybe a year before a lot of these bills are going to come due and people are going to wanna see results and we'll see what happens after that. There's still websites, there's still railroads, there's still electricity, so clearly a bubble popping doesn't mean it'll go away. It's just going to have to find a way to do things more efficiently. And no matter what happens, we'll probably be right here on The Kettle to talk about it. So thanks for tuning in this week and we will talk to you soon. (R)