Datacenter Financing - The Chip Put and Bubble Downside

Investment Analysis Clubs Macro Economic Trends and Risks
1

There’s been a fair amount of discussion about whether there was any broad macroeconomic risk involved in a possible AI bubble - and what might happen if it pops. After all, a literal shirt-ton of money is being invested in the infrastructure for these things - a lot of data centers, and a lot of chips in data centers. This has caused some observers to get a little worried, since on a non-trivial number of prior occasions when you’ve had massive infrastructure buildout for a new technology (railroads, fiber), it turns out that there was a level of exuberance that was a little irrational and things ended very badly when the tide went out. And not just badly for the companies involved, but badly for the macro economy.

At times, though, folks have suggested out that these buildouts are different from past ones, because the companies involved aren’t borrowing to do it. They argued that the hyperscalers aren’t taking on debt to finance these things. If they fail, it will be the investors in those companies that will suffer the consequences, with little contagion to the broader economy.

That’s not true, though. After all, the AI era isn’t just marked by technological innovation - but also financial innovation. The datacenters are largely being financed by debt - there’s just not enough equity financing to come close to the amount of money they need. The debt is being carried on the balance sheet of the Special Purpose Entities (SPE’s) that are created to own the datacenters, and they’re backstopped by obligations of the big blue-chip hyperscalers like Meta or Google or OpenAi or whomever. The SPE’s are able to get the financing because the hperscalers have obligations to the SPE’s, like long-term leases. The SPE can point to the leases as backing the debt, and the hyperscalers avoid having debt on their balance sheet - which is why a number of folks early on thought there was no debt.

Turns out, though, that this isn’t the only way that the “not debt” is actually on the balance sheet of some of the big blue chip companies. Matt Levine writes today about the put options being issued on the chips going into all these data centers, which is also a nice recap of all the other financial innovations involved here:

Meanwhile in debt capital markets, the great financial innovation of the AI age is probably the put option on computer chips. It goes like this:

  1. Building and deploying artificial intelligence is a huge infrastructure project; it requires building giant data centers filled with computer chips, and new electric power plants to run them.

  2. Sure you can raise equity. But the traditional way to fund giant infrastructure projects involves a lot of debt financing: You fund a data center by borrowing money from insurance companies looking for steady cash flows.

  3. Those lenders want steady cash flows.

  4. The projects expect steady cash flows: The big AI companies will happily enter into giant multi-year leases for not-yet-built data centers, because computing power is one of the main constraints on their growth and they’ll take a much as they can get.

  5. So the project can go to lenders and say “we have a 10-year lease with Anthropic that will pay us $10 billion per year, so we can easily cover your $9 billion a year in interest payments.” Good, good.

  6. But this has just shifted the problem: Now the lenders are effectively lending the money to Anthropic, because they are relying on Anthropic to make the lease payments.

  7. Equity investors love Anthropic, but lenders might worry. Anthropic “is still a startup and lacks the earnings to support such a large debt deal”; it doesn’t have the steady cash flows that will soothe the lenders. AI is all fairly new, and who knows if Anthropic will even exist in five years. OpenAI and some other giant AI companies are in the same boat.

  8. “What else you got,” the lenders might reasonably ask the developers of the project.

  9. “Well, look,” say the developers, “Anthropic might not exist in five years, because the AI business is competitive and fast-moving and maybe someone else will be the winner. But the point is that computing power is the constraint on everyone, so even if Anthropic disappears, someone else — whoever the winner is in the AI race — will still want these chips in this data center. So if we lose our current tenant, we’ll easily be able to replace it with another tenant paying the same rate.”

  10. This is not a bad argument, but it leaves the lenders underwriting the credit of an unknown tenant. “Some yet-to-be-founded AI company will provide your cash flows” is not soothing to lenders. “What else you got?”

  11. The developers might say: “Even if we can’t find a tenant, we can sell the chips for a pretty high price, because, again, computing power is the enduring constraint of AI and somebody will always want it. This is not just a loan backed by cash flows; it’s backed by collateral (the chips).”

  12. This again is not a bad argument, but it might still worry the lenders. Who says the chips will be valuable? Equity investors in the AI boom are telling an up-and-to-the-right story in which demand for AI will only increase and the chips will be in demand forever, but what do the lenders know about that? A world in which Anthropic stops making lease payments is very plausibly a world in which the chips have become worthless. “What else you got?”

  13. “Fine,” say the developers. “We actually have a giant existing investment-grade public company involved in this deal, and that company is also very bullish on this whole AI thing. That company will promise to buy the chips at a fixed price, as last resort. So if everything goes wrong, we can put the chips to the guarantor and get money back to repay your loans. Happy?”

  14. So the whole complicated structure is, at bottom, guaranteed by a giant investment-grade company. That’s something the lenders can understand.

Money Stuff: Google Needs More Money | NewsletterHunt

So they go out to Broadcom (or whoever) and get that chip put, which Broadcom is happy to issue very cheaply. And the risk of what happens if there’s an AI bubble keeps getting spread around, even though it’s not showing up on the balance sheets where it truly lies…

10 Likes
2

People continue to tell us AI demand is very high. Inadequate capacity is limiting use of AI. Data centers are being constructed as fast as possible. 3 years seem typical.

It will be a few years before we see how this comes out. I saw Oracle CEO on CNBC yesterday confirming demand is high and noting their traditional business relationships give them a competitive advantage.

No question there is a risk. And some players will fall by the wayside. You expect losers to sell their assets. But maybe not if equipment is obsolete.

Stockholders worry; the companies are full speed ahead.

3 Likes
3

Railroad technology and fiber technology for the most part don’t charge very rapidly. Yes, railroads innovate periodically, and things change, but still not very rapidly. Same for fiber technology, there is plenty of innovation, but it is generally slow, probably the biggest jump was going from single frequency to multiple frequencies, and that was a while ago. So when railroads overbuilt and “went bust”, the remaining infrastructure could still become useful even if a decade had gone by. Same for fiber, all that fiber that was laid still remained under the oceans and ground, and became usable even a decade or two later. BUT chips are in a different class. Chip manufacturers come out with new and better, sometimes WAY better, stuff every few years. And whatever infrastructure you overbuild with chips could easily be worthless or close to worthless a decade later. Now that’s not strictly true for data centers as a whole, the structure, the racks, the cooling, the access to energy, etc remain worth something. All you need to do is to replace the cards with old chips to one with new chips and you may be good to go. But still the majority of the value right now is in the chips.

But I don’t think that is the big issue right now. Seems like any computing power that comes on line is instantly utilized 100%. Nothing remains dormant due to overcapacity. I think the big issue is ROI. Companies are paying more and more and more for AI services, and if they don’t see some ROI soon, they’re surely going to pull back some of the spending. That’s when the s**t may hit the fan!

2 Likes
4

Yes, that’s absolutely the issue. The question isn’t whether AI is useful (it is!), but whether it’s so useful that it will generate sufficient economic demand to pay for all the infrastructure being built.

Early in the hyperscaling buildout, there were folks arguing that there was no risk to the broader economy, because these things were being built without debt. So if AI was a bubble and it popped, the economic damage would be limited to just the companies playing in the bubble, not the broader markets. I think that argument has faded as it becomes clear that the datacenter buildout is in fact being financed through debt.

And quite a lot of it, too. More than $300 billion so far, expected to almost double in the next year or so. For comparison, the capex binge of telecomm and internet debt bubble that led up to the 2000 market crash was around $600 billion. The collateralized debt obligations that triggered the Great Recession in 2008 were about $700 billion. We’ll be in that neighborhood by about this time next year. Granted, the economy and the financial markets are bigger than they were back then - but still. There is a massive amount of debt out there that’s depending on this stuff to make a lot of money in order to get paid back.

9 Likes
5

Corrected for inflation, that $600 billion is now $1.2 trillion. In addition, real US GDP has grown by 80% so the current capex spend is about one-seventh relative to the telecom spend.

DB2

5 Likes
6

When they did the post-mortem of the GFC, many facts came to light, like Lehman overexposure to certain concentrated issues, like AIG insuring LOTS and LOTS of debt with the contracts defining default not quite as a prudent insurer would like, etc. But the core was that there were many financial interconnections that could easily cause a domino effect.

Perhaps we should do a pre-mortem on AI and data center funding now? Sure there’s a roundabout funding from A to B and from B to C and from C to D and often from D back to A. But if everything goes to hell, maybe only A, B, C, and D will go under or dramatically shrink? The question is if any major financial connections exist to other institutions outside that circular dealing?

So it’s not just the level of funding, but it’s also the interconnectivity of the funding.

Ah, but how much of the $100s of billions is debt? And where is debt coming from? Is it concentrated at the next Lehman, or is it spread out across 40 huge pension funds and insurance companies?

And I’ve heard that the world moves faster nowadays, and it appears to be true, so maybe it’ll all blow up before the $trillion is spent and the last few hundred billion will be saved from oblivion?

2 Likes
7

Interesting point. Shortage probably means users are paying top dollar for AI service. When AI becomes abundant, will pricing be more competitive? Will profits moderate?

1 Like
8

The $300 billions that I was referring to is all the debt portion. There’s also a ton of equity and direct cash participation, but there’s been more than $300 billion in debt issued.

The buyers of that debt are all through the economy: private equity companies (like Blackstone and Apollo), big institutional bond firms (PIMCO holds a big chunk of it), all the major investment banks and their clients (JPMorgan, Goldman Sachs, BofA), etc. And massive chunk of it gets securitized - cut into tranches, sliced and diced, and sold to pension funds and insurance companies. You know - just like during the Great Recession. This stuff is in accounts all over the economy. It’s all nicely rated because all of these loans are backed by solid collateral in the form of datacenters that can’t possibly materially decline in value. Because of the leases to AI firms, you see - and apparently the chip put options as well. Just like houses!

Yes, in absolute terms and inflation adjusted terms that $300 billion - while large - is still smaller than the debt bubbles that preceded the 2000 crash and the Great Recession. But by end of next year, assuming things keep going (a dangerous assumption!), it will be getting closer to a trillion dollars. Which is enough to leave a nasty mark on the macro economy if it starts to sour.

5 Likes
9

This is more good news than bad news. If it were concentrated among a few firms, then a default that size could kill them. But if it spread out, it’ll hurt them, probably depress their earnings for a few quarters/years, but it won’t kill them.

It may not be “just like” during the GFC. That’s because large bunches of mortgages were packaged into one instrument and then that instrument was slices into tranches. As a result, when individual mortgages defaulted, it was difficult to connect the individual default to a specific piece of a tranche (because it wasn’t known in real-time if the lower tranches had already reached zero or not), and that made apportioning the loss nearly impossible. It was a huge mess. Add to that the fact that defaults take a lot of time, and that sometimes a default is partially remedied and it was a bigger mess. But none of this was really the big problem that broke the camels back. The big problem was that insurance companies (specifically AIG) were issuing policies insuring the bonds that results from the slicing into tranches. And not only did actual holders buy insurance, but others who wanted to “bet” on a bond failure would also buy insurance!

In this case, usually the bonds are based on a single large asset, not 5000 different assets/mortgages. And when that large asset begins to fail, it is known, and even if the bonds are divided by tranches, it is much easier to see when the lower ones start failing and hitting zero, and the next tranches get some percentage of their capital back, and perhaps the top tranches get all their money back. It all depends on how much value is recovered out of the asset at the end. And I suspect that when a data center with $20B in bonds fails, they’ll be plenty of folks (“vultures”) lining up to buy it at $8B or $10B or whatever very quickly. It’s a “spoiling asset” (because new more powerful chips come out periodically) so it has to be resolved fast and put back into service. Usually it never goes out of service, it just gets a new owner being paid for the service.

And I would look closely at the availability of insurance products related to these loans, and if they ever allow duplicate insurance contracts to be written on the same set of bonds.

I agree. It could even tip us into recession. But there is less likelihood that it’ll crash the economy completely.

2 Likes
10

I’m not sure that really makes the current situation better. I don’t think the problem heading into the Great Recession was that the various debt instruments weren’t going bad fast enough.

Ah, history. It doesn’t always repeat, but it rhymes…

Trading in products that pay out when companies default is soaring as investors hunt for ways to protect their portfolios against the risk that the artificial intelligence boom turns into a bust.

Volumes in so-called credit default swaps tied to a handful of US tech groups have climbed 90 per cent since early September, according to data from clearinghouse DTCC.

The expanding use of these strategies underscores how some investors are growing uneasy about a rush of bond deals by tech companies to finance AI infrastructure, which could take years to generate returns.

Investors seek protection from risk of AI debt bust

As Nvidia, Oracle And Amazon Pour Billions Into AI, Wall Street Is Quietly Hedging For Trouble

We are nowhere near the volume of CDS levels we saw before the Great Recession - AIG had about $440 billion in CDS before the crash, much of it on housing. But it’s climbing as the debt levels go up. CDS volumes on the major blue chip hyperscalers are now over ten billion. Again, nowhere near 2007 levels. But that’s just the stuff on the big corporate debt, which has an actively traded market - there’s lots of paper out there that lets you bet on the specific SPE’s (like Blue Owl) debt going bad. As back in the day, you don’t need to own the underlying debt to buy the CDS.

So that’s another thing to keep an eye out for. Again, we’re in early days in this.

3 Likes
11

This is what I never understood. Why insurance policies were ever sold to someone who had no risk of loss.

1 Like