Front Burner - If AI is a bubble, what happens when it pops?

Episode Date: November 17, 2025

With the absolutely massive amounts of money tied up in the AI data centre boom, it's not hard to see why people fear a bubble. That worry has come into sharper focus in recent weeks, following comments from OpenAI and some big moves on the stock market.This recent round of bubble fear isn't about the tech itself. Rather, it's a growing realization that the boom is being funded in a way that’s starting to resemble some historically devastating bubbles of the past.Paul Kedrosky is a partner at the venture capital firm SK Ventures and a research fellow at MIT's Initiative for the Digital Economy. He explains why changes in the AI boom's financing are renewing fears of a bubble bursting, and the massive potential impacts if it does.For transcripts of Front Burner, please visit: https://www.cbc.ca/radio/frontburner/transcripts

This ascent isn't for everyone. You need grit to climb this high this often. You've got to be an underdog that always over-delivers. You've got to be 6,500 hospital staff, 1,000 doctors, all doing so much with so little. You've got to be Scarborough. Defined by our uphill battle and always striving towards new heights. And you can help us keep climbing.

Donate at lovescarbro.cairbo. This is a CBC podcast. Hey, everybody, I'm Jamie Prosson. You know, how can the company with $13 billion in revenues make $1.4 trillion of spend commitments, you know, and you've heard the criticism, Sam. We're doing well more revenue than that. Second of all, Brad, if you want to sell your share. is I'll find you a buyer. I just, enough.

If you are OpenAI CEO, Sam Oatman, you're clearly getting a bit tired of having to answer the question of whether the AI industry is a bubble. Well, simply saying enough hasn't put the question to bet. Here is Bill Gates. There are a ton of these investments

that will be dead ends. At the end of October, hedge fund manager Michael Burry, the guy who predicted the 2008 housing collapse, posted on X, quote, sometimes we see bubbles. Sometimes there is something to do about it. Sometimes, the only winning move is not to play. And these are not the only signs. So why? What has changed in the world of AI that has got investors and experts feeling so spooked? If it's a bubble and it pops, what economic damage will it cause? Paul Kodroski is with me today.

He's a partner at the venture capital firm SK Ventures. He's also a research fellow at MIT's initiative for the digital economy. So he's been watching all of this very closely. Paul, hi, thank you so much for making the time today. Sure. Glad to be here. So as I mentioned, people have been saying that AI is a bubble for a while now. This isn't new. But why have we been seeing this renewed wave of headlines and people being concerned about the bubble popping over the last couple of weeks? Yeah. So the way I try to frame it for people is that there's really two different ways to talk about a bubble and they get very confused. And so there's the idea that AI itself is a bubble or a mania or it's some kind of, you know, furry animal that we're going to one day be embarrassed. We all got excited about. And that's a mistake. It's not. It's actually a really useful technology. There's going to be lots of uses for it, blah, blah, blah. The way in which it's a bubble is in the sense that it's got people so excited that they're overspending in building out all. All of the things that go along with AI, you know, most notoriously, these things called data centers, these big hulking warehouses full of chips.

But there's a whole host of things that go along with that. And so there's this massive buildout underway, and that buildout is very expensive. That very expensive buildout is financed in a bunch of different ways, which is increasingly making people nervous. So that's the latter part of the bubble, the second bit. But how all of this is going to get paid for, and to my mind, and maybe more importantly, are we building an absolute boatload too much of this stuff, given what we know about what's going on? So the second part is what got people excited

in the last, say, a couple of weeks. Yeah, and just could you tell me a little bit more about the concerns people have around the lending, right? And how the funding model, as I understand it, has changed somewhat. Yeah, so whenever we were only, and I say only with air quotes, spending tens of billions of dollars building data centers, these large technology companies, the Googles and Metas and others, and even the private ones, like an Open AI,

we're spending mostly out of cash flow, meaning out of their profits. These are hugely profitable companies on the order of hundreds of billions of dollars, and they can ante-op tens of billions for data centers without having to go out and raise more money because they just take some of their earnings and use it exactly for this. But what's happened is this has become a much more expensive project. You've probably seen some of the numbers get bounced around. There was a report last week from J.P. Morgan.

that on current projections, it might cost $4 trillion over the next, say, five years to build out all of these data centers and related infrastructure for AI. Well, that's not going to come out of companies' earnings. Just mathematically can't happen. And so what's begun to happen this year and really accelerated in the second half of the year was the technology companies increasingly went out and found partners to raise money. And those partners are usually what are called private credit firms. So these are like banks that aren't regulated.

There's a host of them out there. Most of them emerged, ironically enough, after the last financial crisis. And they go into these partnerships with the large tech companies to build the data centers. And these are done, and it's very opaque and sometimes mysterious way, using things like special purpose vehicles, these legal entities that only exist to build the data center. And I have a stake in it as a technology company. The private credit company does. We put in debt. to put it back on my balance sheet as a tech company. So when it comes time to say, well, how much

exposure do you have to all of this? I get to say, hey, hey, not my table. I don't own that. It's over there, which is kind of a shell game, obviously. But nevertheless, that's the change that's made things much riskier, really accelerated things in the last, I'll say six months, but even in the last two or three months, it's really dramatic how this has changed. And it's this change to a more debt financing away from just using cash flow. I remember back in October, I think it was, there was this diagram in Bloomberg that was making the rounds, right? And it was showing the way that a lot of money seems to be flowing in the industry at the center of the diagram was Nvidia, which is the main provider of the chips that use, that these AI models run on. And, of course, the single biggest company on the U.S. Stock Exchange.

And then there are all these arrows, right, showing the cash flow in various directions between Nvidia and like all these other companies, OpenAI, Microsoft, Oracle, et cetera, et cetera. And, you know, it just struck me as just huge amounts of money moving around between different players in the industry and, like, it just seemed incredibly circular. And is that different than what you're talking about now? Because I know people were concerned about that at the time, too. Yeah. So that's another piece of the problem. So there's this other aspect, which used to be called back in the days of the telecom bubble, back in the dot-com days, this was called vendor financing, where I, extend alone to you as in video. It's not really alone, though. It's just money I'm giving you

so you can buy my stuff. Now, back in that period, Nortel, Canadian company, was at the epicenter of that. So that has gone on previously. It rarely ends well. But it's just another aspect of the current mania that we also, in addition to this prodigious amount of debt that's flowing into these technology companies to allow them to go out and buy and build data centers, there's also an aspect by which some of the money is very circular in that I extend credit to you. It could be in any different form as an NVIDIA, say, to Open AI. But in exchange for that, guess what you have to buy all my stuff? So it becomes very circular and very hard to tell where the reality is anymore because all the money just goes, as the Bloomberg diagram showed,

it just goes round and round and round. So that's another aspect of the risk. It just masks the size of things and then creates this illusion of even faster growth, which in turn drives more fundraising, which in turn drives more debt, and the more debt there is, the more risky things get. I wanted to also talk to you about something that caused a stir earlier this week, and it was a comment from OpenAI's chief financial officer, Sarah Fryer. She was talking at this Wall Street journal event and she essentially said that this is where we're looking for an ecosystem of banks private equity maybe even governmental um the ways governments can come to bear meaning like a federal subsidy or meaning like just first of all the the backstop the guarantee that allows the financing

to happen she then walked that back did she ever in a LinkedIn post saying open AI is not seeking government backstopping, and Sam Altman chimed in as well, saying that they are on track for a $20 billion run rate this year and will hit hundreds of billions by 2030, money that will fund its $1.4 trillion in compute contracts. The Trump administration AIs came in and was like, we're not going to bail out anybody if you kind of implode, right? And why did this talk of government backstopping seem to freak out so many people? Yeah, because it's got, it gives you acid flashbacks to the financial crisis, I would think, because in the sense, that's what Fannie and Freddie were, the big federal, if you will, mortgage banks in the United States. They were backstopping very risky mortgages, which in turn all were packaged out and syndicated and kind of blew up the global economy. But at the back of it all were these two federal entities that were, in a sense, making it easy for people to make really bad loans. Well, we don't want to reenact that again.

is the thinking. I'm not surprisingly. And so to the extent that government would be in any way backstopping loans to data centers, which are inherently a very risky proposition given how rapidly these things become obsolete, or inability to know what future demand looks like, all those kinds of pieces makes these loans very, very risky because, again, they have to pay off the loan using the proceeds from the center. And if the centers are turning out, we're building 2x too much, well, that's not going to work because the math doesn't work on that. And so here's the thing, though. only really said the quiet part out loud. There is this idea, and it's not just in the U.S.,

it's in Canada and it's worldwide, that we have a national interest, is the claim in AI. There's this idea of sovereign AI and so on. So the idea that Open AI would kind of take it to the next step and say, well, if this is so strategically important, you should be willing to stand behind what we're doing in some form or another. Now, saying you're going to backstop loans, maybe have been a one step too far. But listen, the U.S., in the form of executive orders and in the form of explicit policy is really pushing the idea that this is the new global international arms race. And, you know, I see the same thing happening in Canada. I see it in Western Europe. I see it in China. So it really is just saying the quiet part out loud.

So our AI minister, Evan Solomon, has talked about sovereign AI quite a bit. The federal budget that was being voted on this week promises $1 billion. in funding to the AI industry over the next five years and then also recently our Prime Minister, Mark Kearney, even called pipelines. It's so boring.

It's not actually for a big part of the country. It is because it's, look, it's, don't worry, we're on it, we're on it, like we're on it. But there is this, you know, a big topic of conversation here and suggested that not boring, pipelines are not boring to a lot of people in this country. No, yeah, sure. He suggested that the real economic payday for Canada right now

is in data centers and AI. What some in the room will unlock on the data center side, the intelligence infrastructure side, will have a much bigger impact on productivity in this country. We'll have a much bigger impact on our standard of living. It's an easy conversation to have about a pipeline because it's one thing we can see. But the reality is that there's much, much more of the Canadian economy and there's much, much more to the future of Canadian economy. What do you make of him framing it like that that Canada desperately needs to build up its domestic AI industry or risk getting left behind or not as profitable as other countries?

I think that's a dangerous line of patter that's aided and abetted by the technology industry, which likes the tax subsidies that come from getting relief on capital expenditure to build data centers, from getting relief with respect to taxes on the real estate they're willing. getting relief with respect to long-term power agreements. All over the U.S., this is happening constantly, and the claim is that these are the factories of the new industrial revolution, and so we have to have these things, which is, of course, nonsensical because all, say, inference supply,

whenever you ask a question to a chat model, like OpenAI, all of the Canadian inference supply could be done with, I don't know, two data centers in the corner of Abilene, Texas. So the notion that I have to do this domestically or it won't happen, is demonstrably wrong. There's infinite capacity to supply all of that. So then you say, well, we have a uniquely Canadian way of doing this. It's possible, I suppose, but it seems unlikely because

training these models is a prodigious exercise requiring exabytes of data and years and years of effort. It's not something you're going to do just because you have a couple of extra data centers in northern Alberta. So it's not about training and it's not about inference and it's demonstrably not about the job creation because these things basically run lights out after you finish the construction phase. So I defy you to come up with a rationale for why it's strategically important to spend billions of dollars on data centers

other than it takes the cost again off the balance sheet of the technology companies themselves. This ascent isn't for everyone. You need grit to climb this high this off. You've got to be an underdog that always overdelivers. You've got to be 6,500 hospital staff, 1,000 doctors, all doing so much with so little. You've got to be Scarborough. Defined by our uphill battle and always striving towards new heights.

And you can help us keep climbing. Donate at lovescarbro.cairbo.ca. Are your pipes ready for a deep freeze? You can take action to help protect your home from extreme weather. Discover prevention tips that can help you be climate ready at keep it intact.ca. You mentioned the 2008 crisis earlier, and I just wonder if you could talk a bit more about it because it's probably on a lot of people's minds listening to you. You mentioned some similarities, but like how might what's happening right now be different from 2008?

So it's different only in the sense that it's all of that and more. And this is the thing that I think people miss about the current episode. They say, oh, you know, these crazy technologists, there's another goofy bubble. They're doing crazy stuff again. It's not dot com. It's something else. It's data centers. No, that's not really the case.

So if you go back to the history of modern financial euphoria's bubbles, if you will, they tend to localize around a couple of things. So there's a great technology story. Well, this one has that. there's loose credit, the easy availability of financing. Well, this one had that, and so did the global financial crisis, and it was mostly used for real estate. It tends to have an element of government involvement in many cases. Well, this one has that, and so did the financial crisis.

There's all these pieces that come together, but usually each of those are discrete bubbles. This one has all of the pieces that make a single bubble combined in one. Real estate, technology, loose credit, government. And this is what makes this bubble different. What makes it different is sort of the bubble of bubbles, which is why it's so much larger and growing so much more quickly than people realize and why people all seem to see it from different perspectives. Real estate people see it as a real estate thing. Tech people see it as a tech thing. Government sees it as the new industrial revolution. There it is right there. This bubble combines all the things that may have made the worst bubbles in the last 150 years into one neat package. I suppose one kind of counter-bubble argument that I've heard is that we could get this build-out of infrastructure in these data centers that seems a tremendous overkill at the moment, but that it might be useful later, right? So, for example, in the dot-com era, you saw the massive expansion of fiber optic networks, which are now the backbone of the modern internet.

Or further back, you've got the railway mania. Yep. Or canals or take your pick. Yeah. And it did actually lead to something useful later on. And what do you make of that argument, that we might be overbuilding these data centers now, but that they will be useful someday? It actually mostly shows why this bubble's different because there's nothing.

There's no commonality between building out of fiber optic network or laying railroad tracks or building canals and building data centers full of GPUs because GPUs have very short lifespans, not just because, technology changes really quickly, and that's true. You're probably turning over the chips on a three to five-year basis, generally speaking, inside of data centers. Railroads are not like that. I don't have to turn over the rails every three to five years. I may have to weed them now and then, but that's about it. This has nothing to do with turning over the technology changing. So even if I don't use a railroad line for five years, and I come back and I say, you know what,

I want to run some trains down that track. You can do that with impunity. That is not the case with the GPU-driven data center. They can't just sit there and then be turned on and suddenly come back to life and everything will be fine. The technology has changed. And there's an even more insidious component,

which is that if the data center is used for training, so let's say I was using it to train some of these large language models, that's kind of like using your car for like a 24-hour race at Le Mans. It's running that thing so hot that it breaks the... It's just like a car breaks down quickly

from being used for endurance racing. Chips degrade quickly. if they're used for training. So it's not three to five years anymore. If I've used my data center for training, it's often 18 months to 24 months. So now, again, we have this problem that given that 60% of the cost of a data center is the chips

and their lifespans are very short, the notion that we can build all of this and it'll all be, we'll figure it all out four or five or six years down the road. No, it's actually the reverse. You're going to have to make all those capital expenditures a second time or most of what you spent previously will be useless for this reason of the relatively short,

relatively short lifespan of the technology underneath it, as opposed to canals, railroads, fiber optic cables. You know, let's say this bubble pops. Like, I just, how could this play out? And, you know, in your estimation, how bad could it get? Because I'm just thinking, too, that at the center of this are companies, that also are quite diversified, right? And do make money doing a lot of other things.

So could it be as catastrophic as 2008 or the dot-com bubble? Or are you saying it could be worse? Oh, it could be way, much worse. For me, it's more like the late 1920s where you have the combination of this. So at the time, there were two big technologies that were making people incredibly optimistic and that led in part to the crash of 29,

the subsequent economic decline. And one of them was electrification, This idea that we have this ubiquitous technology, rural electrification, that everyone's going to connect to and it's going to change everything. And it did, but we vastly overspent, built too many speculative companies, many of which were packaged into forms that all failed after the crash of 29. And the car was another technology that came along at the same time that helped the two together conspired to make people incredibly optimistic, drive this incredible stock market gains. Well, we've seen something similar in the lead up to this current episode, something like 35% of the S&P 500. is AI-related companies now from a market cap standpoint, something like pushing 50% of the gains of the S&P 500 over the last four years are all tech-related. I mean, this is insane. We didn't

even see this at the peak of the dot-com period, this concentration in such a small number of companies. So that unwinding has a negative wealth effect, makes people feel poor and they spend less. But at the same time, something like 50% of recent GDP growth in the U.S. has been data centers. So it's been like a private sector stimulus program. that nobody talked about. The unwinding of that is the exact same. It's a contraction in the economy and takes you into a recessionary period. So the idea that somehow we can remove all of this, the thing that was more than half of GDP growth, the thing that's driven all the stock market gains, and have that not have consequences is naive. And then the thing that makes it

even worse, of course, is that now that it's being significant debt component attached to it, well, the debt then becomes collateral for other things. And that is increasingly catastrophizing through the economy because the financial providers who are providing the debt and they turn around and securitize it. They cut it up into pieces and sell it to other people. So the debt finds its way. It's not just sitting with a bunch of, you know, billionaires at private equity firms. That debt gets sliced up and ends up everywhere. So you've got pieces of this, not just in equity markets, but debt markets and everywhere else. Wow. That makes me feel quite nervous what you just said.

I really, the question I want to know is when do you think we're going to start to see it unwind, but I suppose that's not a fair question to ask. No, I mean, that's the problem. And, you know, I hear a lot from hedge fund manager and others who have the same question. And, of course, the answer is, A, if I knew I wouldn't tell you, because I'd probably just, I'd go out and play Michael Burry or something. But even Michael Burry apparently doesn't know, and he's decided that it's too complicated for him. He's just quitting.

Yeah, he's like, I don't understand this. Right. I'm taking all my toys and I'm going home. So I think what we saw in the last two months was a really nice, sort of an opening realization of people saying, you know what? These technology companies are more at risk than we thought. There are these esoteric instruments called credit default swaps that are used to measure how risky your debt is. And historically, technology companies, people didn't even realize that their debt could be risky. And Oracle, one of the larger technology companies, their credit default swaps have doubled in the last three months.

That's a sign that it costs roughly twice as much. to ensure their debt against default. This all sounds very wonky, but the point being, the idea that technology companies risk of defaulting on debt could double in three months, there you go.

There's a sign that finally there's a growing realization that something dramatic has changed and it's that an entire sector of the economy as much riskier than it once was. And I think it'll get riskier before the wheels really come off, but is it more than two years away,

a year away? I doubt it. Okay. Paul, thank you so much for this. I really appreciate you coming by. I hope you'll come back soon. Sure. Thanks. All right, that is all for today.

I'm Jamie Poisson. Talk to you tomorrow. go to cbc.ca slash podcasts.