Skip to content
Paul Kedrosky
6 min read

SPVs, Credit, and AI Datacenters

How a new credit bubble is building in AI data centers

SPVs, Credit, and AI Datacenters

Updates and Erasures:

Last week I wrote about how AI goes through three phases:

  • AI imitates us
  • We and AI co-evolve
  • We imitate AI

A recent Bloomberg piece was a reminder that this is already happening, and there are unexpected implications: Call center employees are increasingly irritated at being accused of being AI bots.

By the time Jessica Lindsey’s customers accuse her of being an AI, they are often already shouting. For the past two years, her work as a call center agent for outsourcing company Concentrix has been punctuated by people at the other end of the phone demanding to speak to a real human. Sometimes they ask her straight, ‘Are you an AI?’ Other times they just start yelling commands: ‘Speak to a representative! Speak to a representative!’

To adapt Kurt Vonnegut, and so it goes.

Rough Notes:

AI Data Centers as Credit Risk

I was struck late this past week by Meta's rumored $29b fundraising for a rapid buildout of more AI data centers. The company is supposedly talking to various private equity firms, looking to structure it as $26 billion in debt, $3 billion in equity. This will buy you a lot of Nvidia GPUs.

But everyone is buying a lot of Nvidia GPUs, or Google ones, as the following WSJ graph shows. More were bought in the first quarter of this year than in any full year before 2021.

But let's return to Meta's AI datacenter spending, because it is instructive. A friend asked me, "Why do that? Don't they have the money?" And that got me thinking. Yes, they do, but that "having the money" doesn't matter illuminates the current moment in instructive ways.

Consider this from the FT article:

Private investment groups have increasingly been pitching investment grade corporations on alternative financings to traditional corporate bonds or loans. Such deals, including the Intel transaction, are often structured as special purpose vehicles or joint ventures, where the asset managers take a large minority ownership share in the vehicle. The company contributes assets to the venture in exchange for the capital — either in debt or equity — that private investment firms provide.

There is a lot here, so let's unpack it. It's saying that companies like Meta, which can raise money from banks at low rates any time they want to, increasingly choose ... not to. Instead, they turn to private investment groups—private equity, essentially—who can create custom financing for the project. And for which the company pays a significant premium over investment grade interest rates. How much more? As much as 200-300 basis points, or 2-3%. This is a juicy return on investment-grade company debt.

So, why would an investment-grade company agree to do that? They do it because the capital needed for these buildouts is so large that doing it with orthodox balance sheet debt, or by issuing sufficient equity, let alone spending your cash, would make a mess of your balance sheet.

By structuring it this way, via special purpose vehicles (SPVs) in which they have joint ownership, companies like Meta don't have to show the debt as their debt. It is the debt of those guys over there, that SPV. Not us. Granted, they retain shared control, and they get to use the AI data center, and nothing there happens without their say-so, but still. It's not ours.

This is accounting trickery, of course. It is a transparent attempt to raise large amounts of money without balance sheet damage by putting the debt in a vehicle you indirectly control, but that, for accounting reasons, doesn't have to be disclosed as your debt on your balance sheet. The accounting term of art is "control without consolidation".

Now, turn the lens around and look at it from the standpoint of private credit providers. Their problem is that they have a lot of capital, but they want higher yields on it—and they want to get that yield without taking more risk. Their solution: create customized financings for investment-grade companies like Meta. They can charge higher rates, but still feel confident that they will be repaid.

At the same time, an increasing percentage of private credit providers are funded, in part, by controlling interests in insurance companies, whose capital they use to fund investments. Finding investments that generate higher yields without higher risk—lending at above-market rates to companies like Meta—is exactly what they want to get higher yield while not running afoul of insurance regulators.

And while it all makes perfect sense as financial engineering, this is where the risks start. Why? Because this system creates a powerful incentive loop between structurally overcapitalized insurers, return-hungry private equity firms, and mega-cap companies trying to avoid looking like they're leveraging up. Everyone gets what they want—until something breaks.

There are three main issues:

  1. Opaque risk. The illusion of off-balance-sheet financing makes risk murky. Yes, the SPV isn’t technically on Meta’s balance sheet, but for all practical purposes, it exists to serve Meta, is controlled by Meta, and its default would materially impact Meta. Ratings agencies are not blind to this, but credit markets often behave as if these are truly arm’s-length deals.
  2. Overbuilding. With capital artificially cheap and risk mismeasured or opaque, the market signal is: Build more. AI data centers, like telecom fiber before them, are highly capital-intensive, and the long-term returns remain speculative. The fact that equity is a small slice of the financing (in this case, reportedly only $3B of $29B) further amplifies the leverage risk. This creates a capital stack that’s thinly cushioned if AI ROI assumptions go awry: a 10% equity buffer is wildly insufficient if projected AI workloads stall or margins compress.
  3. Asset-liability mismatch. Insurers funding these SPVs via their general accounts face what's often called asset-liability mismatch risks. Insurance regulation assumes things that may not hold in a downturn—long-duration liabilities matched with liquid, diversified assets, not with illiquid concentrated bets—or if AI capex is too high. And yet, the rating agencies let it pass because the borrower is Meta, not a pre-revenue startup.

To add to the second point, here is a rough estimate of AI data center now vs telecom spending then during the dot-com bubble. We are already spending more now than then, driven by the same sort of hyperpolic projections of future usage and spending. They may turn out to be correct, but if the returns aren't there quickly, the thin equity cushion above datacenter SPV debtholders will be speedily wiped out.

Conclusion

This epic AI data center spending, partly on the back of financial engineering, will work until it doesn't—and when it doesn't it could be a very big mess. Granted, not a mess on the scale of the global financial crisis after the housing bubble, but that is perhaps only because no one has yet had the bright idea of rolling up cash flows from SPV-controlled data centers and syndicating them. Maybe let's not suggest that.

Meanwhile, there is a new risk regime growing in front of us, and, as usual, it is in the empty spaces between regulations, at the intersection of non-bank finance and AI data centers. It will grow rapidly, and if something breaks, damaging insurance assets, people will wonder why they went along with using home and life insurance to pay for AI data centers.

Is this GFC 2.0? No, not yet. This is not systemic risk in the mortgage-backed sense. But the components are familiar: leverage hidden in plain sight, mispriced risk, and capital chasing yield through increasingly convoluted structures. We’ve seen how that story ends. Collateralized AI Obligations, anyone (CAOs)? I kid ... I hope.

Rougher Notes:

Share Share Share Share Share Email

Read next

Four Ideas (Macro edition): Geoeconomic Shifts, Emerging Markets, Tariffs, Japanese Manufacturing