So what are the actual risks of an AI bubble, and what would that mean for the tools that businesses and firms are investing in right now? We interrogate the various indicators of whether there might be a bubble or not, and discuss the considerations and failsafes that businesses should consider when adopting AI software.
Host
Philippa Lamb
Guests
- Tom Pugh, Chief Economist, RSM UK
- Tasmin Lockwood, tech and environment journalist
- Hothyfa Elbeera, Senior Software Engineer, Claimsgate, MoJ
Producer
Natalie Chisholm
Series Lead
Mark Rowland
Transcript
Philippa Lamb: Welcome back. As they prepare to go public, valuations for top AI firms are nudging a trillion dollars, notably OpenAI, the creators of ChatGPT, and Anthropic, makers of Claude. Money is pouring into the sector, more tools are coming onto the market, but can it last? The possibility of an AI bubble surfaced at the start of last year. It hasn't gone away, not least because of that raft of upcoming IPOs. But how likely is it that the sector will deflate or even crash? What are the warning signs to watch for, and what will the fallout be if the appetite for AI investment does dry up for lost economic growth and for businesses buying AI tools right now?
[Teaser audio] Hothyfa Elbeera: The AI market being overcrowded is not in itself evidence that AI is a bubble. AI's underlying technology is delivering genuine value to organisations. In my opinion, the market is currently separating genuine productivity gains from AI hype.
[Teaser audio] Tom Pugh: You hear reports of AI firms having to throttle usage because demand is just so much for these services. We're still in the kind of experimentation phase.
[Teaser audio] Tasmin Lockwood: If all of these companies start to realise they're spending too much on AI, they're obviously going to start pulling back on their use, or at least I would assume, and then how does that then affect the bottom line or the profits of these other companies?
PL: Tom Pugh is chief economist at RSM—he's back on the podcast to share his thoughts. As is tech and climate journalist Tasmin Lockwood, who's worked with Business Insider, CNBC, and Tech Funding News, and Hothyfa Elbeera, a technology specialist currently working with the government on AI projects. Welcome, everyone.
TL: Thank you so much for having me.
TP: Thank you.
HE: Hi.
PL: Tom, it's fair to say talk of an AI bubble is maybe not as prominent as it was at the start of this year even, but it is still being discussed, isn't it? What are the worries?
TP: This has been an ongoing story now for a while, and every time somebody declares that we're definitely in a bubble, the market goes up another 10%, 20%. Seems to be a pretty good indicator. I don't know if we're in a bubble, I'm not sure anybody does, but I think there's some reasons why actually even if we do turn out to be in a bubble, it shouldn't be nearly as bad if it deflates as some of the previous episodes. There's a couple of reasons for that. One is we get really worried about equity market bubbles or bubbles in general if there's a really strong relationship there between the market and the financial sector. So if you think of the 2008 financial crash when all of the property market went and we got a series of bad loans cascading through the financial system and that meant that banks couldn't then lend to normal households. The whole system dried up and that led to the financial crash. Actually, that doesn't look like what's happening this time. There's not a huge amount of debt borrowing going into this AI bubble, so if it does burst, it should be much more of an equity market issue than a kind of financial crisis issue. I think the other thing to think about is a huge amount of this money is actually going on infrastructure, so it's going on data centers, it's going on chips, it's going on energy, et cetera, et cetera. So if this does pop, we will still be left with all of that compute and energy power, which will put itself to a good use. Think the kind of the railway bubble where railway stocks are incredibly inflated. They all crashed, but we were left with that underlying infrastructure.
PL: But on the other side, I guess it's fair to say these companies, they're not making profits, are they, right now? All these multiple IPOs, there is a sense, isn't there, that when companies go public it's because they're reaching the height of their value, it's a good moment for them. So is it not fair enough to raise the question?
TP: Actually, a lot of these companies are making incredibly large profits. If you think about Nvidia for example, it's making huge profits, and the increase in profits is enormous every quarter.
PL: But Anthropic isn't?
TP: No, Anthropic isn't. So if you are looking for an indicator of a bubble, a pretty good one tends to be if you get a wave of IPOs within the sector. Historically speaking, that tends to indicate that you might be at or around some sort of peak level because you probably don't IPO if you think your company is undervalued.
PL: But that's what we've got happening now, isn't it? We've got Anthropic coming up, SpaceX, and OpenAI all talking about it?
TP: Yeah, this could well be an indicator that we're reaching some sort of goal, or this is not a hard and fast rule by any account. It tends to be, you look back and there's a reasonable correlation, but there are lots of episodes where it's not happened. It could be a chance that actually if you're a private investor within these firms, you've on paper made a lot of money and you'd like to realise some of that. Could be that actually the cost of building out all of these data centers is enormous. To build out a full data center is up to £80 billion with associated infrastructure. You can't really fund that effectively through the debt market, so the way to do it is you have to go to market and release some of that kind of pent-up equity that you've got. So it's not a hard and fast rule. You look back through previous episodes, there's a correlation there, but correlation isn't always causation.
PL: Tasmin, what do you think?
TL: Those are really interesting points, and we are seeing in the private markets a huge liquidity crisis, so people are desperate to free up some cash. So of course, there will be investors pushing these big companies to go public anyway. As to whether there's a bubble or not, yeah, for sure, nobody knows. But I think if it ultimately does deflate or burst in a dramatic manner, there will be like a big correction in the private markets for sure because you have all of these smaller— but still big in terms of the private markets— AI companies holding their breath, waiting for these bigger dogs, Anthropic and OpenAI, to go public and see how they're received before they ultimately do as well. So yeah, speaking to whether a spate of IPOs would be an indication of a bubble, I think once these ones are underway, we will definitely see a lot of smaller companies pursuing IPOs. Apparently from my conversations with investors, companies would traditionally look at IPO-ing when they've got around, I think it was like £250M ARR, and today it's more like £500M. So companies are just staying private for longer in general, and this is basically adding pressure to the liquidity crisis. So yeah, there's definitely going to be a wave of IPOs if these ones are received well by the public market.
PL: That's the thing, isn't it? Because even if one of them isn't received perhaps quite as well as expected that's going to put a question mark over the whole sector? Would it be fair to say that?
TL: Yeah, I think so.
PL: What about something entirely different? What about another, say, DeepSeek style announcement from China? Something completely out of left field again.
TL: I think actually that would create more interest in AI companies, because there's this whole sovereign AI play, right? And we all want the next big AI company to be born and bred in our respective countries. If China is coming out with competitors who are thinking differently and managing to build these crazy AI models on less powerful hardware, then that shows us that there is a way to do this that is different from the way that we've been building AI. So surely that would then light a fire in the bellies of our entrepreneurs and our investors and our policymakers to try and do this differently. I think this is a fundamental problem with Silicon Valley as well. Everybody looks the same, they speak the same, they've had the same education, there's no diversity in thought, and that's why we are struggling to come up with anything that's actually innovative.
PL: Hothyfa, there's been a lot of talk, hasn't there, about diversity of thought in this and that, actually the levers of these organisations are all concentrated in a bunch of white male Americans. But it is a crowded market. Is that good? Is that driving innovation or just healthy competition, or is it a problem?
HE: I do believe that AI brought a breakthrough into the market similar to the internet. Is the AI market overcrowded? I suppose yes. I do believe that's a normal thing. I believe we've seen the same thing happen with cloud computing, mobile applications, and software as a service providers. Most providers may not survive the long run. The AI market being overcrowded is not in itself evidence that AI is a bubble. AI's underlying technology is delivering genuine value to organisations. In my opinion, the market is currently separating genuine productivity gains from AI hype. I think we are in a provider bubble, not necessarily an AI bubble. Some vendors may disappear, but the capability remains valuable. If there is a correction, I expect AI to become less visible but more embedded within organisations. Organisations will not be talking about AI strategy. They will simply be using AI as part of their normal operations.
PL: Tom mentioned a lot of this money is going into, literally bricks and mortar. I'm wondering about the practical factors that might limit that expansion, things like planning consent, community resistance. The US and Ireland have seen some, I think it's fair to say, bad outcomes from these data centers. There's been a lot of press about it. Is that something that could hold back development in this sector?
TL: We saw OpenAI 'pause' Stargate UK citing energy costs as a result. So energy is completely intertwined with the infrastructure built for AI. Then, yeah, when it comes to community resistance, it's super interesting in the UK, we've had the government intervening when local councils have declined planning applications for data centers. So that in itself is super problematic, and we actually had earlier this year the government apologise for one of those instances saying that they shouldn't have intervened essentially.
PL: So they don't have a clear policy on this, do they? It does seem to me the government's in a bit of a bind with this. I don't know, what do you think, Hothyfa? Because on the one hand, this is a sector that could really drive the economic growth the government is very preoccupied with producing, and yet voters don't want these data centers on their doorstep because they've heard about water contamination, water use, power use. How's that going to play out, do you think?
HE: First of all, I do believe the government is investing into AI. We do see many departments across government launching AI projects as well as using AI within their daily workflows to speed up workflows in general. Getting to the point of building data centers, I do believe there's going to be a lot of limitations, whether that's in energy or water. This will come alongside a lot of complications. I'm not definite on how the public will view these actions.
PL: Tom, what do you think? We've got a case study close at hand in Ireland, haven't we? Power demand there in, I think it was 2024, pushed up residential power bills by €350 per household because it's Europe's preferred hyperscaler location. The pull on the power on the grid there was so immense that it pushed up domestic bills. That's not going to play well, is it?
TL: They had to introduce a moratorium as a result, which has since been rolled back though.
PL: What do you think?
TP: Look, it's not going to go down well, but we come from a starting point where we already have the highest energy prices in the developed world, essentially.
PL: So that makes it even worse, doesn't it?
TP: Well, precisely. But if you were looking to build data centers, the UK, unless you had a specific reason to build in it, would be top of mind simply because the cost of energy is already so high. If you're running compute for ChatGPT, it doesn't need to be in the UK. There are some specific instances where you might want to.
PL: Okay. I'm interested in this because I think there's over 100 applications currently, as I understand it, on UK soil for data centers. So presumably, there is an appetite to build these things, but I'm not clear myself on whether they need to be adjacent to the businesses that are using the tech.
TL: So some applications do need to be close to the point of use, right? Mostly because of light latency. You don't want a lag on your services.
PL: So the further away it is in geographic terms, the slower everything runs. Is that essentially what you're saying?
TL: I'm sure there's a more technical person that could give a more accurate definition. But that's what we're talking about. Maybe Tom would be able to.
PL: But when we talk about latency, that's what we mean, right? So, that's why?
TP: Well, exactly. There's a couple of reasons, isn't there? If I'm looking to edit something that I've written in ChatGPT, I don't really mind if it takes me 20 seconds or 30 seconds or a minute, it's not the end of the world. If you're trying to run a high frequency trading algorithm using AI, well then, yes, it matters. Fractions of a second matter. You certainly would want to be as close to your AI hub as possible. There could be regulatory reasons. If you want to run AI analysis of NHS health centers, for example, you may very well find that you have to run that within the UK.
PL: Because you don't take your data across borders?
TP: Yeah, I assume it would be a requirement that you'd have to keep that data within the UK. You couldn't just send it off to a server in America, for example. But for a lot of applications, you can just run them in the cloud wherever that happens to be.
PL: So we'll likely end up with a mix, won't we, of data centers on UK turf and use abroad. But thinking about the ones that will come here, there is a direct conflict, isn't there, for the government in terms of their green energy targets. There's a direct conflict with pulling on the grid for other public infrastructure projects, hospitals, all that sort of thing. There's only so much power in the grid. All these things, are they checks on potential progress in the sector?
TP: Yes, to some degree. But if you look at the UK energy situation at the minute, we have an enormous amount of renewable wind capacity, certainly in the north, that currently we are unable to make effective use of. We spend billions of pounds paying wind companies to turn off every year because we can't overload the grid. So one potential application is that you site things like data centers, which are heavy energy users, where there is the available power rather than trying to shift that power to where industry is. So yeah, I agree with your kind of broad point that the UK energy system is incredibly expensive and there's not a lot of excess consumption ready to go. But I think if you're building new data centers, there is the possible opportunity there to take advantage of the energy infrastructure that's already been built. This is an idea that's been mooted, around not just data centers, but relocating. If you relocate heavy industry to where the wind farms are, you can get discounted power, for example.
PL: Tasmin, in fairness, obviously in the US, AI built their own power stations. There's just more space, there's just more opportunity to do that sort of stuff. They're not really so feasible on a little island like ours. But the companies are investing in energy research because they know they're going to be enormous users. So they are in a way, driving innovation in a way that's useful for all of us?
TL: It's actually super interesting because the sort of broader market sentiment on investing in renewables was actually quite damp. Nobody was really into it because renewables were so cheap that there wasn't a strong investment case, right? So actually because of this increased demand from data centers, it strengthened the case for renewables and it has encouraged the build out of renewables. However, at the same time, we also have investment in coal in the US for instance. China's long been building out its coal in parallel to renewables as well. But of course we need power to these data centers, but it's whether renewables or fossil fuels will get there first essentially. Lots of data centers are being built with the intention to be initially powered by gas and then move to renewables when that is able to happen. But I'm skeptical as to whether that transition will actually happen, right? I was doing some preparation for the podcast and I was reading this government document and it turns out there's no actual definition of what a data center is. There's no standard industrial classification code so that we can't really actually track it. So then there was this other piece that I was looking at that said 100 UK data centers plan to burn gas to generate electricity. So that's an immediate red flag and it doesn't necessarily correlate with our climate goals at all.
PL: So another tension between economic growth and climate goals, as you say. Hothyfa, circling back to the idea that, as we say, this is a very crowded space right now in terms of products and companies. I'm wondering how businesses can best protect themselves economically and technologically when some of these tools fall away as they inevitably will in a competitive market. If you've made your choice, where are you left if that provider goes under?
HE: Businesses need to build capabilities. They do not need to build dependencies. They need to avoid building systems which are tightly coupled to a single model.
PL: How do they do that?
HE: Having fallback AI models, for example. If an AI model is no longer with us or if it's not the best performer, we can fall back to a different AI model within the same infrastructure, within the in-house infrastructure, I suppose, in that business. So building abstraction layers and not directly wiring the business to one provider. For example, the business is able to switch without rebuilding the entire system or design systems where LLMs or open source models can be swapped. Business logic needs to be separate from AI providers and lastly retain ownership of their own data.
PL: What's your sense of how much understanding there is of this across business? Because obviously there's a huge drive to adopt, and this is new to everyone, and I am wondering whether people aren't already investing in systems that actually will fetter them when these challenges arise. So I suppose what I'm trying to get at is the key procurement questions they should be asking?
TL: I'm not sure about the procurement questions specifically, but I've seen a host of sort of new technology startups spinning up that are doing this sort of middle layer thing where they are offering a service, very particular sort of use case, and they're tapping into lots of different models behind the scenes, but their customers don't necessarily know which models are being used. The tech company, the tech startup is basically doing all of that legwork for the customer.
PL: It's a hub essentially for them?
TL: Yeah. They never really need to fully understand. They're just using this shiny new tool, and this new tool is pulling in whichever models it sees fit for the request of the customer. Then ultimately if one model does die, then the next one will spin up and get plugged into this middle service. So it is like a nice buffer for the end user. But if there is a bubble and it does burst, how much of these smaller AI companies will still exist anyway is another question, but at least you haven't then sunk all of your time and resources into learning how to use the model, like embedding your systems within it or embedding it within your systems and back, and then be left facing chaos when and if it does die.
PL: That's exactly what you're talking about, Hothyfa?
HE: Yes, correct. To add to what questions should businesses be asking themselves before they buy any AI is, for example, can we switch providers or who owns the data? What happens if the pricing changes? Can we run a different model tomorrow? In some cases, can the solution operate partially offline? So having internal RAG systems where if your AI system does fail, it can still actually operate within that locally. So businesses should not make a five-year procurement decision based on who has the best model this year. I believe the biggest procurement mistake is not choosing the wrong model, it's designing a solution that cannot evolve alongside AI.
TL: Which is a fundamental issue because all of these companies that are building on top of AI models, their business potentially has to change week after week, depending on the changes that the foundation models also have. So you're constantly fighting fires if you're not able to build a robust enough product.
PL: I'm also wondering about supply chain because obviously, AI is in use across organisations of all types and sizes and territories. So how can companies build in some safety for themselves there around the way their supply chain is using AI?
HE: I suppose many AI systems depend on a number of essential components, so a model provider, a cloud provider, a vector database, an orchestration platform, a monitoring system, and third-party APIs. One provider failure can create a chain reaction. Many organisations believe they only have one AI when in reality they have multiple different providers who, if any of them fail, the AI capability will be lost as well.
TL: I think going back to the IPO question as well, just when it comes to supply chains, I've spoken with investors who have thought these IPOs are so inflated, those price tags are just insane. The real value lies in the supply chain. It's those smaller companies that are built out of Asia, built elsewhere in the world that are actually doing the components and the hardware. If you wanna make a buck on AI, that's where they suggest investing because that's the fundamental bread and butter of the industry, it's the infrastructure.
PL: And I guess the real bread and butter of the industry is the fact there is enormous appetite for these services, yeah? Is that the fundamental we really need to be looking at?
TP: This is one of the surprising things. We see the enormous amount being spent on AI infrastructure build-out and colossal sums, hundreds and hundreds of billions, it'll be in the trillions before long in the US alone. But yet, at the same time, you hear reports of AI firms having to throttle usage because demand is just so much for these services. We're still in the kind of experimentation phase now. We're trying to work out where AI is and how firms can best use it to boost productivity. The tech optimists will say that'll be 2030 before we're all jobless, and in reality, it'll probably be a lot longer. But I think fundamentally, AI is going to change the way that we all work, especially in the professional services environment, which is very well suited to an adoption of these AI tools.
PL: It's fascinating, isn't it? Has there ever been a technology that was adopted so immediately and so comprehensively while it was still in development? You talked about railways and automotive and there was a watch and wait, wasn't there, to those technologies? But this, we've grabbed it up, haven't we? The whole world has grabbed it up.
TP: This is a long-running trend, though, in terms of the speed from invention to adoption for each of these technological revolutions becoming vastly quicker than before. It took 100 years for electricity to make its way into factories, and then it took however long, ~30 years, for the internet to really make itself present and we don't know quite how long AI will be, but it's going to be a lot quicker than 30 years. So this has been the quickest one, but this is just a continuation of a trend that we have seen.
TL: I do feel like we are maybe seeing the beginnings of the end of experimentation, however.
PL: Do you? What makes you think that?
TL: It was Uber, I believe, who basically blew through its annual AI budget in the first quarter because its employees were token maxing. I think they basically have said they need to roll that back a little bit.
PL: There's a nervousness about spending that much on it?
TL: They're spending so much on AI, yeah. It's like, what is experimentation and what's actually useful? Of course, the price tags that we're seeing with these big companies, they make sense as long as the business fundamentals also match that, right? So that there's actually value there. But if all of these companies start to realise they're spending too much on AI, they're obviously going to start pulling back on their use, or at least I would assume. Then how does that then affect the bottom line or the profits of these other companies? Then there was also, I believe it was Amazon, that had an internal leaderboard of who was using the most AI, and they basically had to scrap that because their employees were hacking the system and just doing really silly requests with AI just so that they could be on top, and it's just incentivising the wrong use of it. Also when we think about that, yeah, in this broader context and with energy and climate and water use, that's really problematic.
PL: That is interesting, Tom, isn't it? This idea that businesses, they've adopted this, and the budget is running away with them, and are they really clear about quantifying the business wins from it, the productivity gains from it? Do you feel that businesses you speak to are really understanding the benefits in financial terms of this investment?
TP: Not yet. I think it's still quite difficult to find examples where people have adopted them, and it's been a real kind of dramatic cost saver. Please send me any if there are.
PL: It sounds like a worry if they aren't.
TP: Well, I'm sure they're out there. But this comes back to this phase. We're in this kind of implementation, experimentation phase.
PL: But the budgets are big, aren't they? I mean, I hear what you're saying, but this is serious money.
TP: Yeah, I'm not denying that or playing it down, but I think it represents the kind of the risk reward relationship here in terms of, yes, the upfront investment is significant, but so are the potential rewards.
PL: So it's a trust contract right now. We think this tech will work for us, we're happy to pay for it now until it does?
TP: I don't even think it doesn't necessarily have to be trust. You have to take it on the risk reward basis. If there's a 20% chance or 50% chance it works out, then you allocate the kind of associated budget to that. I'm not sure anyone would be throwing their entire budget at a risky use case, and if they are, maybe you're a particularly high risk appetite board. But there is no guarantee with any investment you make that it's going to pan out to be particularly profitable in the same way that you've modeled, and I think the uncertainties are probably wider, but I'm not sure that the fundamental point is any different.
PL: Interesting, though, the points that Tasmin was making about Uber and Amazon. These are very substantial companies, and it sounds like they're just reigning it in a little bit, not quite happy about the way it's being used.
TP: Well, I think with the case that it's, “show me the incentives and I'll show you the outcome”.
PL: So it's the way you manage the systems?
TP: It's exactly the same way. If you incentivise people to just use AI as much as possible, then that's what you'll get. If you incentivise them to use it in a way that enhances productivity, then hopefully that's more what you get.
PL: Hothyfa, is that your experience?
HE: Yes. Companies are, I suppose, investing in AI capabilities with it are actually encouraging their employees to use AI as well, and many of them have been hit with higher than expected invoices or hitting their cap early on into the year. Also, to go back to a very important point which Tom raised around jobs. So I believe businesses need to use AI to understand where human effort is spent before deciding where human expertise is no longer needed. In my opinion, AI will change jobs. Some roles will shrink, new roles will emerge. Many businesses are measuring the cost of savings of AI, but few are measuring the cost of losing institutional knowledge. The riskiest approach is removing people before you understand the operational limits of the technology that the companies are adopting.
PL: You make an excellent point. We're going to have to wrap it up here. There is so much more that I would like to talk about, but I'm sure we'll come back to it. Thank you all very much indeed.
TL: Thank you so much.
TP: Thank you, guys.
HE: Thank you very much.
PL: If you'd like to dig further into this, you'll find links in the show notes as usual. If you're an ICAEW member, remember to click through from there to the ICAEW site to log your listen to this podcast as CPD. Next episode, we'll be hearing about one such AI adoption story. EY is rolling out AI agents across its audit and assurance teams. We'll be speaking to them about how they're actually using them and how they're keeping them in check, as we've just been discussing. If you have any thoughts you'd like to share about today's episode or the podcast in general, you can do so via our contact email. Here it is, podcasts@icaew.com. Don't hesitate. Thanks for being with us.
