Everyone knows that the capex going into data center buildouts is eyewatering. Roughly $800 billion spent by Big Tech since the start of 2023, with more than half a trillion expected this year alone, and nearly $3 trillion in global data center spend anticipated by 2028.

What I didn’t know until recently is that hyperscalers and data center developers have been increasingly relying on creative (and misleading) financial engineering to fund all of this. An instrument of choice has become special purpose vehicles (SPVs), with private credit funds as the lenders.

For the uninitiated, private credit is lending done by investment funds rather than banks, with the loans held privately instead of traded on public markets. Basically, it’s a loan negotiated behind closed doors that no regulator or ratings agency has to look at as a whole. An SPV is a standalone company created to own a single project, legally walled off from the parent. The hyperscaler signs a long-term lease with the SPV, and the SPV borrows billions against those future lease payments to fund construction, with the debt sitting on the SPV's books. This essentially allows hyperscalers to take on debt that doesn't appear on their books.

For example, in a deal for its Hyperion campus in Louisiana, Meta created an SPV with Blue Owl (a private credit giant). The SPV raised about $27 billion in debt, the largest private credit deal ever, to build a data center that Meta will occupy under a long-term lease. Meta gets its flagship AI facility, but the $27 billion is on the SPV's books, not Meta's, and gets repaid from the rent Meta pays to an entity it partly owns.

The whole situation is laughable - even though the legal entities are technically separate, Meta is the tenant, the co-owner, and the sole source of the debt's repayment, and yet isn’t considered the borrower. 

So, the financing is full of incestuous circular relationships, and it's all built around the assumption that this stuff can actually get built and (just as importantly) powered on schedule. But, the grid as it exists today cannot handle that.

The most blatant issue is interconnection, the process of getting permission to actually plug new power into the grid. At the end of 2025, there were roughly 2,060 gigawatts of proposed capacity sitting in interconnection queues, more than one and a half times the entire existing US power fleet, just standing in line. For projects that actually made it through and came online in 2025, the median wait from request to delivering power was over five years, and historically only about 13% of what enters the queue gets built.

Everyone knows this, so they're all trying to go behind-the-meter (fancy term for generating your own power on-site and plugging it straight into the facility, skipping the grid entirely). But even if you try and do that, you run into the next issue: data centers have been largely using natural gas turbines for behind-the-meter power, and the three manufacturers that supply more than 75% of US projects are estimating that orders placed today will be delivered between 2028 and 2030.

So, given that the financing is a mess and the grid can’t keep up with the promises tied to it, what happens?

Loans for data centers are basically construction loans, which are modeled around a schedule. While construction is happening, the asset being built can’t make money (nobody will pay rent for something that doesn’t exist yet), so part of the capital is set aside upfront to service the debt until construction is completed. The amount set aside is based on the expected timeline, so if the project is delayed significantly, that capital will be eaten up before construction is finished. If this happens, there are compounding issues which I think are easiest to follow if we trace through a hypothetical example. 

Say a developer breaks ground on a facility that has a lease starting at the end of 2027. They set aside enough capital to service the debt until then, plus some extra because delays are expected, so the full reserve capital can service the debt until Q1 2028. Construction starts, but then the turbine needed to power the facility shows up 8 months later than expected, because everyone is scrambling to get their stuff built so supply chain bottlenecks keep getting worse. The reserve runs out with the facility still under construction, and the debt has to keep getting serviced somehow. So, either the sponsor injects fresh capital or the unpaid interest gets added to the loan balance, meaning the debt grows while the lease still hasn’t started.

On top of this, some of these deals have backstops (guarantees to protect lenders in case of non-payment), which can be misleading. Not all deals have backstops, they're mostly used when a data center is being built for a neocloud (a company that rents out GPU capacity, and is considered way less creditworthy than a hyperscaler), so a hyperscaler steps in and promises lenders that if the neocloud can't make rent, they'll cover it. However, in the most prominent of these deals, Google (hyperscaler) is backstopping bonds for a facility leased to Fluidstack (neocloud), but the backstop only kicks in once the lease actually starts. If construction runs more than 180 days late, Fluidstack can walk from the lease entirely, and the backstop never comes into existence at all.

So the same delays that eat through the reserve capital can also erase the guarantee that made lenders comfortable with the loan in the first place. The longer a project runs behind, the more the people holding the debt are just holding construction risk on an unfinished building, without any of the protection they thought they had.

To be clear, this is not a “Big Short” moment on its own yet. The first response to a breached loan is usually pretty boring: the parties quietly agree to extend the timeline because nobody wants to foreclose on a half-built data center they can’t operate. But since these are private loans valued quarterly (by the same people who hold them), all of those agreements are invisible to the public, and the structural issues will grow behind closed doors until something (AI demand missing projections, GPUs depreciating faster than the debt on them, etc) forces it into the open all at once.

And if it surfaces, it won't be isolated to tech. The SPV structure benefits the hyperscalers: if a project dies, Meta loses its equity and moves on. But the money for the loans comes from insurance companies, pension funds, and mutual funds, which means the exposure has already spread across the economy, held in vehicles most of the people funding them have never heard of.

Honestly, I'm not sure what happens here. Comparing this era to the dot com bubble is the obvious way to go, and people forget that that was actually like 3 different bubbles stacked on each other. The situation here is similar,  so if delays keep piling up and we reach a point where expectations are overinflated to the point where adoption of AI falls meaningfully behind, the hangover from this frenzy will be brutal.

However, there is a lot of interesting work going at startups across that country that could save us - there's multiple new turbine startups (Stone Power and American Turbines), and two weeks ago Valar Atomics powered Nvidia chips on a small nuclear reactor. America has a rich history of rising to the occasion, and I don't think this time will be any different.