Wednesday Morning Rant

The AI bubble is a thing to behold, and its scale beggars belief. The question underlying the insane AI build-out is not whether AI is useful - it is, at least for some things - or whether AI is overhyped - it is - but whether AI can pay. Can these investments bring in enough money to payy off the bonds, in time for it to matter? The amounts of money being funneled into AI capacity is truly staggering, and all of it ultimately rests on the answer to the question of "can it pay" being a resounding yes. As the scale increases and the dollars fly, the risks grow in tandem and the odds of a bigger problem increase. While all of AI is bubbly, the behavior of the AI mega-players called hyperscalers is the easiest to examine.

The hyperscalers are the ones building the gigantic datacenters and spending trillions on huge AI capacity and training new models - and promising to spend trillions more. The idea is that they will sell this capacity on through services at a price high enough and at sufficient volume in time to make it pay enough to cover the bonds - and position the company properly before the next major refresh when they have to upgrade. This depends on demand, prices and nobody coming up with a way to do it cheaper. There are very real risks and the Big Five hyperscalers - Alphabet (Google), Amazon, Microsoft, Meta and Oracle - are spending so much on this that it threatens not just the overall enterprise, but the wider economy. The amounts involved are so huge that they're playing accounting games to enable it to continue.

There are a number of examples. The new datacenters being built are interesting in their own right, as it is often difficult to determine who is paying for them. Many of these new datacenters don't have the company's name on the door. Or the paperwork. Or the balance sheet. These data centers are being built through "special-purpose vehicles," or SPVs. Datacenter SPVs are standalone psuedo-corporations created to finance, build and sometimes operate a datacenter on behalf of its actual owner. For example, take a deal called Beignet, the SPV building a new Meta datacenter in Louisiana. Meta owns 20% of the SPV, and put up $1.3 billion. Private equity outfit Blue Owl put up $3 billion and owns 80% of the SPV. Combined, they put up around 14% of the total $30 billion cost of the build-out, with the rest financed by bonds issued by the SPV and purchased in bulk by investment funds like PIMCO. Meta will lease most of the Beignet datacenter from Blue Owl, who will take the lease payments to pay the coupons to bondholders and dividends to itself and its minority partner who is also the primary tenant.

Why the complicated structure? Accounting tricks. Continuing with Beignet as an example, take a worst-case scenario where the SPV fails. If that happens, the majority owner - Blue Owl - will sell off everything it can to pay back bondholders. This will inevitably result in a gap. If Beignet were a normal enterprise, that results in a "haircut" - the bondholders lose the difference. If it were a normal enterprise, it would also have to pay ruinous interest rates to finance a high-risk and high-complexity scheme. But it isn't a normal enterprise, and it gets to issue investment-grade bonds. Why? Because Meta is backstopping it. If Beignet fails and is liquidated and the bondholders are not made whole, Meta puts up the difference. The complicated structure lets the SPV borrow at blue-chip rates and - most importantly - allows the blue chip with the risk exposure to hold that exposure off its balance sheet to maintain that rating while burning through capital at alarming rates. Meta is ultimately on the hook for the bond issue, but Meta also gets to exclude the massive liability from its balance sheet.

If that seems extremely suspicious, it's because it is. It's a vast game of hide-the-risk - but hiding the risk doesn't make it go away, and no amount of wishful thinking or accounting tricks can change that. And it's a very, very big problem. Last year, the hyperscalers issued $121 billion in new debt. That sounds like a lot, but this year, the total issuance is on pace to quadruple that as Morgan-Stanley (via Forbes) estimates $570 billion issuance. And the Big Five hyperscalers are holding, by some estimates, around $1.5 trillion dollars' worth of debt off their balance sheets. And people are finally noticing. From a piece of Bloomberg coverage:

"The accounting treatment itself is in fashion. But what if one of these companies was a house of cards and was propping itself up with this accounting treatment?" said Tom Selling, technical accounting consultant. "To me, that's the risk."

"If we just looked at the financial statements, we would be more or less short sighted: missing liabilities, missing obligations, missing key elements of these structures that really truly represent the economics," Gonzales said.
Investors have to dig through disclosures attached to corporate financial statements to find details on the off-balance sheet structures and how each firm accounts for those arrangements and any related obligations.
The information is there, but finding it requires more work than if Meta, for example, borrowed directly to finance the build out, said Gil Luria, head of technology research for D.A. Davidson & Co.
"Enron's crime wasn't having special purpose vehicles. Enron's crime was hiding them," Luria said.

And that scale is another part of the reason. Nobody can afford this. It is getting into nation-state levels of capital and debt expenditure and even the mighty hyperscalers can't pay for their bets, especially as the market for AI bonds appears to be softening. Between known on-balance-sheet debt and estimated off-balance-sheet debt, the Big Five hyperscalers have racked up almost $2.5 trillion in debt so far. Alphabet (Google) plays the hide-the-risk game comparatively little and most of its debt is on its balance sheet and because of that, its accounting is easier to understand - and it is also is in negative free cash flow for the first time since going public as its capital expenditures exceed its cashflows. How many of its fellow hyperscalers are in the same situation, with the reality hidden off the balance sheet?

The answer to the basic AI question of "can it pay" will have profound consequences not just for the hyperscalers, but for the wider economy.

posted by Joe Mannix at 11:00 AM

| Access Comments