I read this Substack column last week comparing the timing of subprime resets in 2006-08 to the payments that the AI companies will soon have to make for data centers as they open. (Sorry, I would credit the author, but I don’t know who Mr. or Ms. Groundbreaker is.) Most data centers are being constructed with contracts with the AI producers where they first start having to pay lease obligations when the data centers become operational.
Anyhow, I was first taken by the argument, which is essentially that the timing of the collapse of the housing bubble was an entirely predictable event, because we knew that the teaser rates on hundreds of billions of dollars in subprime loans were due to reset to much higher rates starting in 2006, with the volume rising in 2007 and 2008. Since millions of homeowners would be unable to pay the higher reset rate, their mortgages would soon go into delinquency and default. The column sees this as parallel to the trillions of dollars of lease obligations that the AI companies will have to start paying as more data centers start operating in 2027 and 2028.
That originally struck me as a neat parallel, but then I remembered a bit more of the housing bubble history. It’s true that millions of people with subprime mortgages hit a wall in 2007-08 when their teaser rates reset to much higher rates, but that was not a wall that first got erected in 2006.
There were hundreds of billions of subprime mortgages with teaser rates issued in 2004 and 2005. If someone had constructed this graph in either of those years, we would have seen a reset wall in 2005 or 2006. It’s true that the wall would have been somewhat smaller in those years. There were fewer subprime loans, and their value was less, since house prices rose rapidly in those years, but the graphs for both 2004 and 2005 would have still looked ominous. (The big jump in 2007 is an inevitable result of the timing. People who reset a mortgage in 2005 or 2006, prior to the construction of the chart, will not face another reset for at least two years.)
In prior years, millions of people who had taken out subprime mortgages were able to refinance into new mortgages before the mortgage reset to a higher rate. That was no longer true when house prices stopped rising. This meant they were stuck facing the higher reset rate, which they were not able to pay.
There is a great scene in the movie The Big Short, where the Steve Carell character is talking with a stripper who owns five houses that she intends to flip. She is explaining how this makes sense. She has subprime loans with relatively low teaser rates, which she can afford to pay. When Carell asks her what happens when the loans reset to higher rates, she explains that she just refinances into a new subprime and gets the teaser rate again. He then asks her what happens if she can’t refinance, at which point she is jolted into reality.
I’m raising this point not to criticize the analysis, which is very useful, but to take issue with the idea that the collapse of the bubble is somehow locked in by the payoff schedule on the data centers being built. Undoubtedly the impending flood of lease payments from the AI producers will be a massive hurdle for them to overcome, but it’s worth thinking through this dynamic more closely.
Let’s assume that everyone reads the Groundbreaker analysis and agrees it is correct. The AI producers have committed themselves to payments that there is no way on earth that they can pay. Does the bubble continue to inflate until they have to start making good on their data center leases?
That seems unlikely. After the hyperscalers read Groundbreaker’s analysis they will know that they will never be repaid for the data centers they are paying hundreds of billions to construct. At best they will be able to get partial payment through renegotiating a contract with a company facing, or actually in, bankruptcy. This would almost certainly leave them with large losses on their massive investments.
Facing that prospect, Alphabet, Microsoft, Amazon and the rest would almost certainly look to renegotiate their deals now, before they throw still more money down the toilet. Perhaps they would not be able to do this. OpenAI and Anthropic may be operating with the view that bankruptcy is inevitable, so all the risk lies with the hyperscalers, but the most profitable companies in the world presumably will not watch themselves slow walk into the abyss without putting up some sort of fight.
Anyhow, this gets into areas of law that I don’t know and contracts that I have not seen, but the point is that the belief in the huge profitability of AI, like the belief in ever-rising house prices, is what drives the bubbles. As long as those beliefs persist, the bubbles can continue to grow.
If people still think AI will be the most profitable thing the planet has ever seen, when the lease obligations come due, Sam Altman will still be able to go down to Wall Street and get hundreds of billions, or even trillions, for whatever garbage he puts on the table. When the people controlling the big bucks no longer accept his story, that’s when the music stops. There is no fixed schedule for the collapse.
And with that – here are the latest numbers on P/E ratios and market cap for the big AI companies:
You must log in or register to comment.
I don’t know who Mr. or Ms. Groundbreaker is
Ed Zitron, maybe? I don’t think he does Substack, but his personal blog is very Substack-like. And a lot of what you’re talking about above feels very on brand for him. His YouTube channel “Better Offline” (Invidious link) is highly recommended as well.
My take on what Ed has to say. I really hope he’s right about his original timelines (which were roughly “end of 2027”-ish). And I want the catharsis that would come of a massive spectacular collapse. And it all makes sense what he has to say. (A hell of a lot more sense than do the claims of, say, Sam Altman or other AI hypemongers.) But it was years ago that it seemed plainly obvious that Tether was going to topple and go to zero imminently and that was going to quickly bring Bitcoin, Ethereum, etc all down with it. That’s not how things have gone. Bitcoin is surprisingly high right now. And Tether doesn’t seem to have wavered at all, despite how manifestly obvious it is that claims of Tether being 100% backed by real USD are pure fiction. The blockchain bubble does seem “popped” at least in the sense that it very rarely enters the news now and I haven’t had any coworkers buttonhole me to sell me on investing a million dollars in Gary-Busey’s-Fish-Flavored-Feet-Coin. (Exaggerating, but you’ve probably had the experience I’m referring to. If not, I envy you.)
Again, I hope Ed or whoever it is you’re posting about is right. I’m fucking sick of this AI bubble. Listening (obsessively) to his content gives me hope (or perhaps “hopium”). But it pains me to say I wouldn’t be surprised if it didn’t turn out quite like he or anyone else expects. The AI bubble is basically just a pivot from the blockchain bubble anyway. It’s possible the AI bubble will pivot again into something else. Or maybe the U.S. government will start bailing out companies like OpenAI and Anthropic. (If they do that, I’d be surprised if they didn’t try to be sneaky about it. Make some cover story about how it’s vitally important to U.S. security that they grant OpenAI or whoever some big contract that just happens to be just enough to keep them from going under.) Or maybe the VCs are dumb enough to keep throwing money down a rat hole for longer than anyone would expect. There are all kinds of possibilities.
In fact, Ed has recently been admitting to having been wrong about earlier timelines and now refuses to make specific predictions regarding timeline.
I was convinced in 2024 that models plateued. But that hasn’t been the case. LLMs really are useful. But there is whole lot of work to be done to be actually useful for areas outside of programming.
I’ve heard multiple times that LLMs have “improved” in recent times. But anything you use LLMs for is something you forget how to do yourself. If LLMs could be trusted to work as well as humans do but without any human supervision/correction then humans forgetting how to code wouldn’t be a problem. (For corpos, that is. For humanity, it would be catastrophic. For corpos it would be… profitable (except that LLMs are super expensive to run, of course, so maybe it remains to be seen whether they could be profitable even in that capacity).) But so long as humans still need to know how to code, I think LLMs will be worse than nothing even for profitability (and would be so even if running/using LLMs cost nothing in electricity/compute/subscription/etc cost). And I can’t see stochastic parrots ever getting good enough to run without supervision.
It’s not redeemable. The amount of computing power coming up is such that they will be making top of the line movies, podcasts, books, and other such entertainment. But even now, we’re just swimming in ads and entertainment. And the population is shrinking. Plus they practically destroyed the personal computer market in the process. So basically they went on a long vovoyage and burnt the boat while getting started on a bigger boat that they could not finish.
Additionally, I could not possibly be worth enough money to anyone to have them pay for equipment to keep track of my daily motions. I’d like to believe that people in power aren’t that stupid.
So maybe it’s aliens? Maybe there’s an asteroid coming our way? Otherwise, there’s no reason to do all this clearly wrong things.
AI was largely being used in places where it was useful already. large LLM’s might not be as competitive in these niches because smaller purpose built ones might strictly outperform them in cost, throughput, or control. Especially with hardware costs. If LLM’s want to be the only game in town, it would have to at least really do what it claims it can. A smaller crappier model is probably going to win out over a larger less crappy one for the cost they need to justify the investment.
AI entertainment isn’t profitable either. people hate it. Even when they can’t tell, it’s something they think about and consider. It doesn’t raise the ceiling, it pushes the floor down. For every successful AI story, there’s thousands of failed attempts aiming for the same thing. People will be flooded with bad AI and look for better content, I know I’ve been using YT a lot less. AI will also obfuscate if views come from real people or not, either by accident or people trying to game the system with it. Imagine a billion buckets, but only a million drops of water. Eventually a person needs to decide who’s deserving of water and who’s not. People will just work harder to push the AI down the queue. It’s happened before. The equilibrium will find the average, and the average will shift past what AI can provide without more investment.
These models aren’t beholden to us either. It seems like the miniscule money they make compared to the debt is more interesting than the fact these models will be used the break the very system they’re built to serve. What if a prediction machine was instructed to delete all the Bitcoin it could hack in that moment. It’s so obvious it’s not really going to make more for us than it’s going to cost us. Except the investors can’t believe anything else anymore, because it’s too late to go back. We aren’t collectively screwed, but it’s far too late to stop this from affecting a lot of us.
