Imagine it’s 2004 and I explain to you what a mortgage-backed security (MBS) is. I show you how they work and how banks are using them to sell groups of mortgages to investors. Even with this knowledge, it would’ve been difficult to foresee the financial crisis MBS would help create in just a few years time. After all, there’s nothing particularly dangerous about MBS. They weren’t new either. The first MBS was issued in 1968.
But as lending standards loosened in the 2000s, riskier mortgages were underwritten, packaged, and sold (as AAA rated MBS) to unsuspecting investors. This once harmless structure became the catalyst for a financial meltdown.
Well, the year is 2025 and I think I’ve found the MBS of our era. It’s called conduit debt financing.
Conduit Debt Financing = The New MBS?
I first heard of conduit debt financing through Matt Levine’s piece about how big tech companies are paying for their data center build outs. I’ve since read this excellent primer on the topic as well. Here’s a simplified version of how conduit debt financing works:
- A tech company wants to build a data center, but they don’t want to use their cash to pay for it.
- Instead of borrowing the money directly (which could harm their credit rating), the tech company creates a new company, called a special purpose vehicle (SPV).
- This SPV borrows the money from investors, builds the data center, and then leases the data center back to the tech company.
- If the tech company ever fails to pay the SPV, the SPV keeps the data center as collateral.
In short, the tech company gets its data center without the debt and investors get a highly productive asset. What’s not to like?
Like MBS, there’s nothing inherently dangerous about conduit debt financing. Similarly, it’s not new either. Municipalities and companies alike have used such structures since the 1980s.
But, with the massive speculation around AI investments, such an arrangement could lead to problems. For example, what if the tech company decides to stop leasing the data center? In Meta’s recent $27 billion deal with Blue Owl Capital, they have the option to cancel the lease after only four years (but will need to pay a capped termination fee to do so).
Technically, the SPV could just lease the data center to another tech company. But what if the specialized infrastructure or the chips inside the data center have become obsolete? What happens if the demand for compute falls dramatically? What will the SPV investors do then? The equity holders will obviously get wiped out, but who will pay back the debt holders?
If the SPV debt holders were all hedge funds, this wouldn’t be a problem. But they’re not. Since the SPV will typically have a credit rating just below that of the tech company, institutional investors like pension funds and insurance companies are eligible to become debt holders. And, if some politicians had their way, retail investors would be eligible as well.
Unfortunately, when these kind of investors lose lots of money, things can get really bad. As Bloomberg recently reported, PHL Variable Insurance Co., a private-equity backed insurance company, couldn’t make payments to policyholders due to a $2.2 billion shortfall. What happens when other insurance companies can’t make payments to policyholders because their SPV debt defaults?
It’s hard to imagine such an outcome when GPUs/TPUs are the most sought after product on Earth. However, just because a technology is becoming popular doesn’t mean it will outperform in financial markets. As Jeffrey Gundlach explained in a recent episode of Odd Lots:
Probably the biggest thing that changed the economy, and the world, in the last, I don’t know, 150 years was electricity. Electricity being put into peoples’ homes was probably one of the biggest changes of all time.
And, of course, around 1900 people realized that electricity to homes was coming. And so electricity stocks were in a huge mania and they did incredibly well. But the relative performance of electricity, relative to…everything else but electricity…peaked in 1911. Houses weren’t even broadly electrified by 1911.
Gundlach is exactly right. Only about 16% of U.S. homes had electricity in 1911, yet that was when the relative performance peaked for electricity stocks.
More importantly, a technology bubble can burst even though the underlying infrastructure is still incredibly valuable. Ben Eidelson and Anay Shah noted this in the Stepchange podcast when discussing the collapse of the DotCom bubble:
And by 2001 everything collapsed. Advertising folded. Startups folded…PSI Net, one of the largest ISPs, had already collapsed.
And so was this the death of the internet and the death of value all this fiber that was laid? Clearly not.
It was the death of a particular moment and the overbuild in a bubble. But, in fact, the actual infrastructure that was built out would prove immensely valuable as services that mattered matured and business models matured.
These examples demonstrate that the demand for a technology can remain high even if the economics don’t make sense for a period of time.
Today, it seems like there is no feasible scenario where demand for data center compute could be lower. But this is how every bubble bursts. One big assumption turns out to be false.
When What You Know For Sure Just Ain’t So
The big assumption in the GFC was that housing prices couldn’t fall across the nation at the same time. This is important because MBS were generally geographically diversified and the homes themselves acted as collateral in the event borrowers didn’t pay their mortgages. So, it was assumed that if home prices dropped in California, home prices in Texas wouldn’t be impacted (and vice versa). When this assumption turned out to be false, things started to come undone.
Similarly, the big assumption today is that the demand for compute cannot go down. If that turns out to be false (somehow), then a lot of high-flying AI stocks will fall back to Earth.
Of course, I have no idea how this will play out. The number of events that could derail this theory are enormous. More importantly, conduit debt financing is still small and doesn’t currently represent an existential threat to the financial system. But, with the trillions of dollars that tech companies plan to spend on data centers in the next few years, small can become “too big to fail” in a hurry.
I’m not an expert in conduit debt financing, but when something seems off, I take note. And how tech companies plan to use SPVs over the next few years seems like one of those things that will be viewed as obviously problematic in hindsight. Because if these big tech companies aren’t taking the risk, someone else is.
Is it Wall Street? Is it Main Street? Will it be the Fed when they have to buy up “distressed” data centers?
I know this might seem a bit far-fetched now, but the CFO of OpenAI has already discussed the possibility of a government backstop to chip investments, though she later retracted it.
I may come off as a bit alarmist in this post. But if a bunch of SPVs default in the future, don’t say you weren’t warned.
Thank you for reading.
If you liked this post, consider signing up for my newsletter.
This is post 478. Any code I have related to this post can be found here with the same numbering: https://github.com/nmaggiulli/of-dollars-and-data
