Welcome to the 219th Pari Passu Edition.

In October 2025, Meta closed what was then the largest private capital deal on record: more than $27bn of debt for a single data center campus in rural Louisiana, none of it on Meta’s balance sheet. Within a year, deals like it had become routine, as companies handed a growing share of the bill to bond investors, private credit funds and banks, and developers that once relied on bank loans increasingly turned to the high-yield market. Then, in September 2026, Oracle sent a force majeure notice on a data center project backed by an $18bn construction loan from roughly 20 banks. In October, the bonds of CoreWeave, which rents out AI computing capacity and pioneered borrowing against AI chips, traded at yields near 15%. 

Those headlines sit atop a financing problem that is unusual even by infrastructure standards. None of the tools involved is new; companies have used separate financing vehicles, leases and guarantees for decades. What is new is the scale. Over the last three years, the market has increasingly converted hyperscaler commitments into project-level debt through a web of leases, guarantees, and financing vehicles. Those contracts can work exactly as written and still leave lenders exposed, because the protection they provide often attaches later than the debt, sits elsewhere in the structure, or covers less than the debt it supports. 2026 is the first year the market has seriously questioned those protections, even though none has yet failed to pay. 

This piece starts with a short primer on what goes into a data center and who builds them. We’ll then follow the market through three phases: hyperscalers paying their own way, then renting chips and backstopping leases, and finally the joint venture structures of the past year, using Meta’s El Paso financing as the example. It closes by examining how those structures have held up under stress, from Oracle’s Project Jupiter to CoreWeave, and the questions a restructuring would raise.

Data center financing is taking up a growing share of the credit markets, and no investor can afford to ignore it, so we’ve made this primer completely free for everyone. If you find it useful, we would love for you to forward it to a friend or colleague.

Through 2030, the US expects to fund ~$3T in compute CapEx. 

This build out assumes an inference cost curve which, within the next six months, is poised for complete re-imagination.

In The Geometric Language Model Thesis, Phosphor details:

  1. A novel architecture which solves three of the deepest scaling issues in today’s transformer-based LLM architectures

  2. Precise descriptions of breakthroughs expected to enable the majority of inference to be served from consumer hardware (laptop + smartphone CPUs)

  3. How this shift away from GPUs poised to impact markets & bets

Phosphor solves frontier technology problems to scale the origination, development and financing of physical infrastructure (such as data centers). This thesis is based on three years of soon-to-be-announced research and development.

Data Center Primer

At its simplest, a data center is a building that houses and runs computers for others, keeping servers powered, cooled, and connected to the internet at all times. For most of the last two decades, demand grew steadily alongside internet services, enterprise IT, and cloud computing. Then artificial intelligence changed the scale. Training AI models, and then running them for users (known as inference, which is where AI companies earn most of their revenue), relies on specialized chips, mainly graphics processing units (GPUs) and, increasingly, custom chips such as Google’s tensor processing units (TPUs). These chips consume far more electricity and produce far more heat than traditional servers. As a result, AI facilities are now sized in gigawatts rather than megawatts [1]. 

The capital involved is enormous. The five largest US technology companies are expected to spend $660-690bn on capital expenditures this year, with roughly $450bn of that going directly into AI infrastructure [2]. Even after counting the cash those companies generate themselves, Morgan Stanley estimates that the buildout through 2028 leaves a funding gap of about $1.5tn that outside capital must fill [1]. That gap is why data centers have become a credit market story.

What Goes Into a Data Center?

A modern AI data center combines three different assets into a single project. The first is the real estate, meaning the land and the building shell. The second is the power and cooling infrastructure that keeps the facility running, including substations, transformers, backup generators, and cooling systems. The third is the equipment, the servers and GPUs that actually do the computing.   In practice, the first two are usually financed together as the data center, with debt sized over 15-20 years, while the servers and chips are financed separately over about five years. Lenders on each side typically sign agreements with the other that set out notice periods and rights to fix a default before either can act. According to Pari Passu sources, much of what complicates data center financing comes from how those two pieces, with very different lifespans, depend on each other. 

To put numbers to this, we’ll use a hypothetical developer, Campus Co., which is building a 1-gigawatt (1,000-megawatt) AI campus. Building the shell and its infrastructure costs roughly $10-16mm per megawatt, and the all-in cost rises to roughly $40-60mm per megawatt once servers and GPUs are installed, per Pari Passu discussions with market participants. For Campus Co., that means the building alone costs $10-16bn, and the fully equipped campus costs $40-60bn. In other words, roughly two-thirds of the campus’s total cost sits in equipment that will be obsolete long before the building is. 

Power is the most closely watched part of any project because these facilities must run effectively without interruption. Tenants typically require at least 99.9% uptime, and many push for 99.999%, which drives much of the cost. Campus Co. needs a firm, dedicated supply rather than whatever the local grid can spare. Securing that supply has become much slower. Connecting a large new project to the grid took about five years for projects completed in 2022, compared with fewer than three in 2008, so developers increasingly build their own on-site generation, known as “behind-the-meter” power. These can range from banks of smaller gas engines to full turbine plants, each with its own permits and fuel supply, and they let a campus come online without waiting for the grid [1]. It is not a quick fix, however. Generating units are scarce, with manufacturing slots booked two or more years out and second-hand units selling at steep markups. The approach is also largely untested at this scale, since few, if any, major campuses have yet brought large on-site plants into operation. 

Those permits have become harder to secure as local opposition over water use, electricity bills, and the small number of permanent jobs has turned into moratoriums, permit freezes, and lawsuits. Lenders generally will not fund a project before its key permits are in hand, so the risk falls mainly on later phases and expansions. Where a lease covers several buildings that are not legally separated, a delay in a later phase can trigger a default on phases already delivered.  

Figure 1: Google’s Lenoir, North Carolina Data Center

Hyperscalers vs. Developers vs. Neoclouds:

Three types of companies sit at the center of the buildout, and the differences between them matter for the structuring that we explain later on.

Hyperscalers are the handful of companies that run the world’s largest cloud and AI platforms: Amazon, Microsoft, Google, Meta, and Oracle. They build and operate massive facilities, primarily for their own use. With few exceptions, they are investment-grade borrowers with long histories in the bond market, which lets them borrow cheaply. Oracle is the outlier, having borrowed far more heavily than its peers to build its cloud business. 

Developers, such as Vantage, Stack, and QTS, build data centers and lease them to tenants. They typically deliver one of two products. A turnkey facility is a finished building with power and cooling already in place, and the tenant installs only its servers and chips. A powered shell is closer to an energized warehouse; the developer supplies the building and brings power to the site, and the tenant fits out the electrical and cooling systems itself [1]. QTS and Vantage mainly build turnkey facilities, while Stack is mainly a powered shell developer.  Returning to our example, Campus Co. is a developer. If it leases its campus on either basis, the $30-45bn of equipment belongs to the tenant rather than to Campus Co. As a result, Campus Co.’s lenders are not directly exposed to the value of the chips. Several newer developers, including TeraWulf and Cipher Mining, began as cryptocurrency miners and converted existing sites that already had the large power connections AI facilities need [1]. Because they had already secured grid connections and power contracts for mining, they could offer energized capacity within months rather than the years a new connection can take. 

Neoclouds are a newer group of companies that buy GPUs in bulk and rent computing capacity to AI labs, startups, and enterprises. CoreWeave and Fluidstack are the best-known examples, alongside Crusoe and Lambda. Most lease their buildings from developers and finance their chips largely with debt, though the largest increasingly build their own sites. Unlike hyperscalers, most are young, heavily indebted companies, and they often rely on a handful of large customers for most of their revenue. 

Importantly, credit quality declines from one group to the next. A 20-year lease signed by Microsoft is a very different asset from the same lease signed by a neocloud that is itself borrowing to buy its chips, and the difference shows up directly in how the debt behind each is priced and traded. Most of the buildout still rests on leases signed directly by hyperscalers, and lenders to those projects increasingly require steep payments if the tenant leaves early. In addition, lenders to weaker tenants often require a parent guarantee or a hyperscaler standing behind the lease [1]. We’ll explain both in Phase II, where the second option has become an important feature of this market.

Put together, data center construction faces one problem. Building at this scale requires far more capital than the hyperscalers want on their own balance sheets, and much of the work is done by developers and neoclouds whose credit is far weaker. The financing structures that have emerged are attempts to bridge that gap by letting hyperscaler commitments support debt that the hyperscaler does not itself owe. How well that works depends on when the hyperscaler’s protection attaches, to whom, and whether it can be withdrawn. The next three sections cover how the market arrived at those structures, starting with the simplest version. 

Phase I. (Through 2022)

Data center financing has largely moved through three stages, and in the next three sections we’ll cover each of them. 

To start, for most of the cloud era, financing a hyperscaler’s data center was not a separate question at all. When Microsoft or Google needed more capacity, it bought the land, built the facility, filled it with servers, and typically paid for all of it out of the cash its businesses generated (e.g., software for Microsoft and search advertising for Google). When building was too slow or too expensive, they rented space in facilities run by colocation providers such as Equinix. Historically, the hyperscalers funded their own expansion. Where they did borrow, it was at the corporate level through ordinary investment-grade bonds [2]. 

From a lender’s perspective, this is about as simple as credit gets. A bondholder lending to Microsoft is not underwriting any particular building, tenant, or power contract. It is underwriting Microsoft, a company with a diverse revenue base, enormous cash flow, and very little reason to default because one campus came online late. The building, the infrastructure, and the chips all sit on the same balance sheet, so the fact that the chips wear out long before the building is the company’s problem rather than the lender’s. If a generation of GPUs becomes obsolete early, Microsoft writes it down and buys the next one. 

Early Securitization Deals:

Alongside the hyperscalers, a separate market grew up around third-party developers that build data centers and lease them out, and its defining feature was caution. The first data center securitization came on February 20, 2018, when Vantage Data Centers issued $1,125mm of notes backed by its leased facilities [4]. In these deals, a developer places finished, operating data centers and their leases into a separate, bankruptcy-remote entity, which issues notes repaid from the rent tenants pay [2]. Noteholders are secured by a pledge of the entity’s equity and assets, and rent flows into a controlled account and is paid out in a fixed order, with debt service ahead of the owner. Early deals also included guarantees from a creditworthy parent covering items such as carrying costs and environmental liabilities, per Pari Passu sources. Investors were underwriting buildings that already existed, already had power, and already collected rent, and even then the deals carried tests that trap cash inside the structure or speed up repayment if the facilities’ cash flow falls relative to debt service. The market has stayed relatively small, at roughly $15-20bn of issuance a year, and serves mainly finished facilities with several smaller tenants. It has played little role in financing new AI campuses and instead serves as a way to refinance them once they are built and leased. That is the baseline against which everything in the following sections should be measured.

Figure 2: Simple Data Center Securitization

Why this Stopped Working:

This model worked well as long as the capital required stayed within what the balance sheets of hyperscalers, developers, and banks could comfortably carry. AI broke that assumption for all three at once.

The first pressure fell on the hyperscalers themselves. As AI spending accelerated, capital expenditure began to outrun the cash these companies generate internally. Across Microsoft, Amazon, Alphabet, Meta, and Oracle, capital spending has been growing about 70% a year against roughly 23% for operating cash flow, and on that trend the group’s combined spending was set to overtake its operating cash flow around the third quarter of 2026. Oracle crossed that line first [31]. Borrowing more at the corporate level is the obvious alternative, and the hyperscalers did plenty of it. The hyperscalers issued about $75bn of investment-grade debt in September and October 2025 alone, more than double the sector’s average annual issuance between 2015 and 2024 [5]. But every dollar borrowed this way lands on the company’s balance sheet, raising leverage and pressuring the credit ratings that make its debt cheap in the first place. Keeping the debt off the balance sheet lets a hyperscaler keep building without that pressure [1]. 

Pari Passu channel checks described a second pressure that fell on the developers. Publicly traded data center REITs (e.g., Equinix and Digital Realty) had to keep leverage within the limits that protect their investment-grade ratings. Growth therefore shifted toward privately owned developers (e.g., QTS and CyrusOne, owned by Blackstone and KKR/GIP, respectively) willing to borrow heavily against a small number of hyperscaler leases. 

Lastly, the third pressure fell on the banks. For several years, they carried much of the load through project finance. A developer such as QTS or Vantage would form a separate entity to hold a single campus and its hyperscaler lease, and four to six banks would commit to the loan, then sell most of it to other lenders. A bank might commit $1bn and keep $50mm. The developer or its private equity owner guaranteed completion. This private market grew to well over $100bn a year and quietly financed most of the early buildout, with the pool of participating lenders growing from a few dozen to roughly 250. 

That model strained once single deals grew past about $10bn. Lenders face limits on how much exposure they can hold to any single issuer. Because so much data center debt ultimately traces back to the same handful of companies, the pool of buyers willing to take more of that risk has narrowed. Additionally, in several recent deals, particularly construction loans, banks ended up holding more of the debt than they had planned, which limited their appetite for the next one [6]. Much of that bank debt was also meant to be refinanced once projects were finished, but at current rates many leases do not support the refinancing, and the securitization market is far too small to absorb single-campus loans that now routinely reach $5bn, per Pari Passu sources. 

Taken together, these pressures pointed to the same solution. If the hyperscalers could not keep the assets on their own balance sheets, and the developers and banks could not absorb them on theirs, the assets had to move into new structures built around the hyperscalers’ promises rather than their ownership. That shift is the subject of Phase II.

Phase II. (2023–2025)

Between 2023 and 2025, a growing share of AI infrastructure investment moved into arrangements where hyperscalers committed to pay for capacity without owning the underlying assets. In practice, this split the data center into two layers. Neoclouds took on ownership of the chips and the computing equipment around them, and developers took on ownership of the buildings, while the hyperscaler stood behind both through contracts rather than sitting on top of them as owner.

The GPU Layer: CoreWeave and Nvidia:

Consider first how a hyperscaler gets access to chips without buying them. Rather than buying GPUs itself, a hyperscaler can sign a long-term take-or-pay contract with a neocloud, committing to pay for reserved computing capacity whether or not it uses all of it. The neocloud buys and installs the chips and finances them largely with debt repaid from those payments. These contracts often include a large upfront payment from the customer, which the neocloud uses as its own equity in the chips. In 2024, that prepayment commonly targeted around half the purchase price, per Pari Passu sources. For the hyperscaler, the rental is treated as a service contract and expensed as it is used,  rather than recorded as capital spending and debt. The remaining commitment appears only in the footnotes listing future obligations, and the risk that the chips lose value quickly sits on someone else’s balance sheet. 

CoreWeave became the template. In August 2023, it raised a $2.3bn facility led by Magnetar and Blackstone, secured by Nvidia H100 chips. Its chief executive described the negotiation as matching how quickly the chips would depreciate against how quickly the loan would be repaid [7].  In practice, these facilities gave CoreWeave an initial period of about 8-12 months to sign qualifying customer contracts, after which each loan was repaid in line with the chips’ depreciation, over roughly three to four years within a five-year term. Lenders were really underwriting the contracts rather than the chips. CoreWeave’s May 2024 facility made that explicit by pricing each loan according to the customer behind it: S+6.00% to S+6.50% for investment-grade customers and S+13.00% for everyone else [8]. Lenders also capped how much could rest on weaker customers. Early facilities required roughly 80-85% of contracted revenue to come from investment-grade customers, a threshold that has since loosened to around 60-70%, per Pari Passu sources.  

Because customer contracts begin only once the chips are installed and running, and the loans are repaid within the contract term, these lenders took little risk on contract renewal. Their exposure was instead to whether CoreWeave delivered the capacity in the first place. 

Each of CoreWeave’s facilities sits in its own subsidiary, which owns a batch of chips and the customer contracts those chips serve. Lenders are secured by the subsidiary’s assets. The key question is recourse: whether they can also look to CoreWeave itself if those assets fall short. In a non-recourse loan, the lenders’ claim stops at the subsidiary, and any shortfall is their loss. In a loan with a parent guarantee, CoreWeave has promised to cover the shortfall, so the lenders can claim against the parent’s own balance sheet. CoreWeave’s earliest facilities all carried full parent guarantees, including one backed by its Microsoft contracts [10]. 

CoreWeave’s 2026 facilities show both versions. In March, it raised $8.5bn through a subsidiary whose chips serve its contract with Meta [11]. The loan is non-recourse to CoreWeave, apart from two narrow exceptions. One is a “bad boy” guarantee, which makes CoreWeave liable for losses caused by fraud or similar misconduct. The other is a completion guarantee, a promise to install the GPUs. Because lenders were relying almost entirely on Meta’s payments, it received investment-grade ratings, and part of it priced at S+2.25%. That compares with the S+13.00% CoreWeave paid two years earlier on loans backed by weaker customers [9]. In other words, lenders priced the loan off Meta’s credit rather than CoreWeave’s. That is why the one facility without a parent guarantee carries the tightest spread in Figure 3. A guarantee from CoreWeave would have added little to a loan already backed by a customer stronger than CoreWeave itself.  In May, by contrast, CoreWeave raised $3.1bn through another subsidiary, which serves two undisclosed customers, and signed a parent guarantee [10].

Figure 3: CoreWeave Project-Level Facilities Spread at Signing [10]

To see the difference, imagine each subsidiary has $1bn of debt, and both run into trouble. Lenders to the non-recourse subsidiary take the chips and remaining contract payments and absorb any shortfall themselves; CoreWeave’s other creditors are unaffected. In the guaranteed subsidiary, the shortfall becomes a claim against CoreWeave, competing with the parent’s own bondholders. One problem stays contained, while the other travels upward. 

Ultimately, renting chips does move the debt off the hyperscaler’s balance sheet. Unless the customer is strong enough to carry the loan alone, though, nearly all of the risk lands on the neocloud standing behind it, since the hyperscaler pays nothing until the chips are installed and running. Even the Meta loan left CoreWeave responsible for delivery, since Meta pays nothing until the chips are installed and running [9].

Nvidia added another layer of support in September 2025, agreeing to buy any capacity CoreWeave could not sell to others through April 2032 under a $6.3bn order. Like the other supports in this section, it is also conditional on CoreWeave delivering the capacity first [12]. 

In November 2025, a delay in delivery of a powered shell by a third-party developer forced CoreWeave to cut its full-year revenue guidance and push revenue expected that quarter into 2026 [13]. The customer contracts remained intact, but the delay pushed back revenue for the neocloud while its lenders had already funded chips that were not yet generating income.

The Building Layer: TeraWulf, Fluidstack and Google:

The same logic was applied to the buildings. Here, the hyperscaler neither owns the facility nor necessarily leases it directly. Instead, a developer builds the facility and leases it to a weaker tenant, usually a neocloud, and the hyperscaler stands behind the tenant’s lease.

The clearest example involves TeraWulf, a former bitcoin miner with an existing campus and large power connection in Barker, New York. TeraWulf agreed to build data center capacity for Fluidstack, a neocloud. On its own, a lease from Fluidstack would have been very difficult to finance, since Fluidstack is a young company without the credit profile of a hyperscaler. Google closed that gap by agreeing to backstop Fluidstack’s lease obligations. If Fluidstack misses a rent payment or becomes insolvent, Google must either pay a termination fee or pay the rent due and take over the lease as tenant. In return, Google received warrants that would give it roughly a 14% stake in TeraWulf [6]. 

With that support in place, TeraWulf issued $3.2bn of senior secured notes due 2030 to fund the build. Rating agencies weighed in a range from BB to CCC, well below investment grade, even with Google’s name attached. Google was also careful to describe the limits of its commitment. It is not a guarantee of TeraWulf’s corporate debt; it is tied only to the lease, through what the filings call a “recognition agreement”. Cipher Mining, another former miner, signed a similar arrangement with Fluidstack the same month, with Google backstopping $1.4bn of obligations [14]. Even these backstops do not cover the full cost of a build, and lenders are starting to lend against only a portion of the backstop rather than all of it, per Pari Passu sources. 

Figure 4: TeraWulf, Fluidstack, and Google backstop structure

Timing Uncertainty:

Lenders to these projects face two different credits: the project while it is being built, and the project once it is leased and operating. That split is standard across data center financing, since a tenant’s rent, and any guarantee of it, usually begins only when the facility is ready for service. The question is how the two periods are covered, and here the deals differ. Google’s backstop on the TeraWulf lease takes effect when Fluidstack’s lease begins, which is when construction is finished.  During the build, bondholders have no Google support and bear the full risk that the project is late or over budget. If completion slips more than 180 days past the target date, Fluidstack can terminate the leases entirely, which extinguishes Google’s obligation before it ever attaches. Analysts noted at the time that the construction schedule looked aggressive, and the construction phase was the main factor holding the rating down [6]. Cipher’s arrangement is different. Google’s agreement there took effect on signing, and Google can generally walk away by paying the termination fee only after the first six years of the lease; otherwise, it must take the lease over once capacity comes online [32].

This is a meaningful departure from Phase I, when data center securitizations financed buildings that already existed and already collected rent. TeraWulf’s bondholders committed capital before the building was finished, but the credit support that made the deal financeable arrives only after it is. In other words, the enhancement covers the period after the riskiest phase has passed. The rating it supports is, in substance, a bet that construction goes smoothly enough that the gap never matters [6]. 

The relationships have also kept shifting. In July 2026, Fluidstack agreed to buy out TeraWulf’s controlling stake in a Texas joint venture the two had formed, part of a broader move by the largest neoclouds to control their own sites [33].

The pattern of this period is that hyperscalers and chipmakers lent their credit to weaker parties, but did so conditionally, and usually only once construction and delivery risk had passed. The risk that did not travel with that credit stayed with the neocloud and the developer, and much of it ultimately reached their parent balance sheets. 

If you are enjoying this edition, subscribe to get new research delivered directly to your inbox.

After years of requests, we finally launched our training programs.

If you are looking to break into restructuring banking, get a special sits/credit investing seat, or get better at your role, this is for you.

Monetizing a claim, a locked digital asset position, or a contingent fee receivable? Contact Turnpage for strategic guidance and turn-key liquidity.

Phase III. (October 2025–Present)

By late 2025, the hyperscalers had spent two years lending their credit to weaker parties through take-or-pay contracts and lease backstops. Next, they applied the same logic to their own largest campuses. In the deals that followed, a hyperscaler builds a campus it will use itself, but a separate vehicle owns the campus and borrows against it. The hyperscaler connects to that vehicle through a lease, a minority equity stake, and a guarantee. Lenders are no longer underwriting the hyperscaler’s balance sheet directly, or even a single contract; they are underwriting the structure that ties those pieces together.

In its simplest form, the template works as follows. A hyperscaler and one or more infrastructure investors form a joint venture that owns the campus. The investors hold most of the equity and the hyperscaler holds a minority stake. The joint venture borrows most of the construction cost from bond or private credit investors. The hyperscaler then leases the finished campus as its sole tenant. Its rent passes up through the joint venture to pay interest and principal, and whatever is left goes to the equity owners. Because the hyperscaler does not control the joint venture, the debt stays off its balance sheet, even though lenders are counting on its rent to be repaid. 

Figure 5: Hyperscaler Joint Venture Financing Template 

The deal that started this era was Meta’s financing of its Hyperion campus in Louisiana in October 2025, known in the market as Project Beignet. Meta and funds managed by Blue Owl Capital formed a joint venture, owned 80% by Blue Owl and 20% by Meta, which raised more than $27bn of debt. The notes were rated A+, mature in 2049, and priced at roughly 225bp over Treasuries [2]. Beignet became the template others followed. 

The clearest view of how that template works, however, comes from the deal Meta did nine months later in El Paso, where the rating agencies published the full structure.

Sopaipilla: 

In July 2026, Meta used the Beignet structure for a roughly one-gigawatt campus in El Paso, Texas, this time partnering with BlackRock’s Global Infrastructure Partners and HPS. The BlackRock funds hold 80% of the project, Meta retains 20%, and Meta serves as developer, operator, and sole tenant. The issuing vehicle, Sopaipilla Investor LLC, placed ~$12.5bn of 7.534% amortizing notes due November 30, 2048 on July 27, 2026, priced at T+287.5 and rated A+ by S&P [15].

Like Beignet, the structure combines construction financing and long-term bonds in one package. Instead of borrowing from banks to build and refinancing later, it raises amortizing bonds up front that cover both construction and roughly 20 years of operations. Because the joint venture owns the campus and leases it to Meta, Meta reports its equity stake and lease obligations rather than the project’s debt. The gap between a 20-year lease and a building expected to last much longer is covered by a residual value guarantee (RVG). If Meta stops renewing, lenders are left relying on what the campus is worth without Meta in it, and a purpose-built AI campus may be worth far less to anyone else. The RVG covers part of that gap via a capped cash payment. Meta promises that the campus will be worth at least an agreed amount if the lease ends. Suppose the agreed amount is $8bn and an appraisal values the campus at $5bn when Meta leaves; Meta pays the $3bn difference, up to its contractual cap. If Meta’s payment is capped at $2bn, that is all it pays, and the remaining $1bn falls on the joint venture and, ultimately, its lenders. If the campus is valued at $9bn, Meta pays nothing. The tenant is therefore guaranteeing that the building will be worth a minimum amount, up to a limit, not guaranteeing the debt itself [2]. If more debt is outstanding than the agreed amount when the lease ends, lenders bear the shortfall.  

Figure 6: Sopaipilla Organizational Structure [15]

S&P’s published organization chart shows how the pieces fit together [15]. Following the numbers in Figure 6: 

  1. Funds affiliated with BlackRock own a pledgor entity, whose ownership of the issuer is pledged to bondholders 

  2. The pledgor owns Sopaipilla Investor LLC, the holding company (HoldCo) that issues the $12.5bn of notes

  3. HoldCo owns 80% of the joint venture (JVCo), Project Sopaipilla Holdings LLC. BlackRock contributed $4.9bn of cash, partly funded by the notes.  

  4. The JVCo owns Wurldwide LLC, the property company that owns the campus itself.

  5. Wurldwide holds the power agreements with El Paso Electric and signs the construction contracts

  6. Then cash flows back up the chain, as Wurldwide pays the campus’s costs and sends the remaining “net amounts” to JVCo, which distributes cash to HoldCo and Meta. HoldCo uses its share to pay the notes and fund reserves [15]

Meta connects to this structure in four roles. It is (i) the tenant, leasing the campus from Wurldwide under a triple-net lease that rating agencies said pushes construction, power, operating, and renewal risk onto Meta. It is (ii) the operator, under a property management agreement. It is (iii) a 20% owner of the joint venture. And it is (iv) the guarantor of the RVG. Meta’s rent flows to Wurldwide, which pays the campus’s operating costs and sends what remains up through the joint venture to HoldCo, where it services the notes. Any RVG payments also flow to HoldCo [15]. The rent is sized to repay the notes over roughly 20 years, and Fitch projects about $1.12 of cash available for every $1.00 of debt service [16].  

Two Levels Above the Buildings:

The detail in the S&P chart that matters most to a credit investor is where the bondholders sit. They lend to HoldCo, two levels above Wurldwide, the property company that owns the campus and collects Meta’s rent. Their collateral is HoldCo’s stake in the joint venture and the pledged ownership of HoldCo itself, not a mortgage on the data centers. In this kind of holding company financing, the lender is secured by equity in the entities below and repaid from cash distributed up the chain. [17] That is likely why S&P rated the notes one notch below Meta despite viewing the credit as close to Meta’s, as there is no direct pledge of the assets. 

The consequence is structural subordination. Here, the bondholders have no claim on the property company at all, so obligations at Wurldwide, including its power and construction contracts, have to be met before cash moves up to HoldCo. How the cash that reaches the joint venture is divided between HoldCo and Meta’s subsidiary depends on the joint venture agreement, which is not public. In our view, HoldCo likely has some form of priority, since BlackRock funded most of the cash and the bonds behind it need a predictable distribution. The financing therefore depends not only on Meta paying rent but on that rent passing through each entity as the documents require. That is why holding company lenders typically insist on cash management controls, limits on cash leaving the structure, and consent rights over key project contracts [2]. 

The same structure shapes recovery in a downside case. Without a mortgage, bondholders cannot foreclose directly on the campus. They could, however, enforce on the pledged equity and take control of HoldCo’s stake in the joint venture. How much that is worth depends on the joint venture agreement, any transfer restrictions, and Meta’s governance rights. It also depends on Meta itself. Whoever owns the stake after enforcement still owns an interest in a campus that Meta leases, operates, co-owns, and backs through the RVG. 

That is a different point from the TeraWulf deal in Phase II. Sopaipilla’s lease places construction and power risk squarely on Meta, a far stronger allocation than a backstop that attaches only after completion. Even with Meta’s credit behind the deal, bondholders hold claims against a separate structure with its own rights and remedies, and their protection is contractual rather than physical.

As a whole, the structure gives each party something it could not easily get on its own. Meta gets a gigawatt campus built and operated to its own specifications without carrying $12.5bn of debt on its balance sheet. It also keeps a 20% stake in the upside, in exchange for long-term rent and a guarantee on the building’s value. BlackRock’s funds get 80% of a large infrastructure asset with a contracted, Meta-backed cash flow, and most of their investment is financed with debt. Bondholders get roughly 20 years of amortizing debt that rates close to Meta while earning a modest premium over Meta’s own bonds. That premium pays them for the risks described above. They sit two levels above the buildings with no mortgage. They rely on a guarantee of the building’s value rather than of the debt. They carry construction risk in the early years. And they hold a large, one-off bond that is harder to sell than Meta’s own. Investors generally price these bonds as a spread over the tenant’s own unsecured bonds of similar maturity, adjusted for how far the campus is from completion. Once a campus is finished and leased to an investment-grade tenant, that gap typically narrows to under 100bp, per Pari Passu sources. 

Testing the Structure (2026)

Phases II and III described how the market moved risk away from the hyperscalers and onto a web of leases, backstops, and guarantees. On paper, those structures assign construction, power, and timing risk to specific parties. This section looks at what happens when that risk actually shows up. The clearest example so far is Oracle’s Project Jupiter in New Mexico, but it is not the only one.

Project Jupiter: 

Project Jupiter is a 2.45-gigawatt campus being built in southern New Mexico as part of the Stargate program, the AI infrastructure initiative announced by OpenAI, Oracle and SoftBank [18]. The structure looks broadly like the developer model from Phase II. Stack Infrastructure, a Blue Owl portfolio company, is developing the campus under a build-to-suit lease, with Blue Owl funds providing roughly $3bn of equity and Oracle signed as the tenant [19]. A group of about 20 banks provided an $18bn construction loan [18].

The project’s problems have centered on power, and under the lease, power is Oracle’s problem. Oracle took responsibility for supplying it, contracting directly with Bloom Energy for fuel cells and arranging the gas pipeline to run them. After the New Mexico State Land Office repeatedly denied the proposed route, which crossed state land, the pipeline’s expected start date slipped by nearly six months to February 1, 2027 [18].  

On September 24, 2026, Oracle sent Stack a notice citing force majeure. A force majeure clause excuses or suspends a party’s obligations when events outside its control prevent it from performing. Oracle did not try to walk away from the lease. Instead, it sought to preserve the right to delay its payments if the campus misses its planned 2028 opening [18]. The timing matters because the lease requires Oracle to start paying rent once the buildings are finished, whether or not power has arrived. 

Both sides played the notice down. Oracle described such notices as routine for projects of this size and said they do not by themselves mean the project is delayed. Blue Owl said the notice did not change the project’s financial commitments. That calmness had some basis. By market standards, Jupiter’s lease is unusually favorable to lenders. It puts the entire power obligation on Oracle and, unlike most data center leases, gives the tenant no right to terminate for a delay of this size. Even if the deferral were granted, rent would eventually start regardless. That turns the question less into whether the campus gets paid for and more into whether Oracle, the most heavily indebted of the hyperscalers, keeps paying. The market reacted more cautiously. The roughly $18bn construction loan was already trading below 90c, partly reflecting higher rates, even though the notice did not constitute a default and no payment had been missed [18]. Additionally, several of the banks in the lending group had reportedly been left holding more of the loan than they had planned, since buyers for the debt were harder to find [28].  

Oracle’s lease obligation is also only part of the picture. Jupiter is one of five US sites tied to OpenAI’s roughly $400bn infrastructure commitment with Oracle and SoftBank, so the demand that ultimately supports the campus comes from a customer that is still funding its own growth through outside capital [18]. The legal obligation to pay rent sits with Oracle, but the economic reason for the rent rests heavily on OpenAI. That is important because many customer contracts across this market roll off well before the debt they support is repaid. If AI revenue takes longer than expected to catch up with spending, the risk shows up first at renewal, when the tenant decides whether the capacity is still worth paying for. 

Force Majeure Explained:

To understand the market’s reaction, it helps to be precise about what Oracle is asking for. If both parties agree that a power-related force majeure event has occurred, Oracle could delay the start of rent by up to three years. It would still pay other project costs in the meantime and would still owe rent for the full lease term once payments begin [18]. If granted, that relief would change when Oracle pays, not how much. Market participants described the relief as a valuable flexibility for a tenant whose financing, lease, and power costs start well before a site earns revenue. 

Timing, however, is exactly what this financing depends on. The roughly $18bn loan and about $3bn of Blue Owl equity pay for the campus itself. Interest accrues on the loan as it is drawn and, with no rent during construction, is typically added to the balance or paid from a reserve sized to last until the planned 2028 opening. Once rent starts, the construction loan is meant to be refinanced with long-term debt sized against that rent, either a new single-campus loan or bond priced close to Oracle’s own bonds, or, for some developers, a securitization alongside their other finished facilities. That expected refinancing is what the construction lenders price their premium, holding period, and prepayment protection around. 

A three-year deferral would break that sequence. The project company could end up with buildings that are largely complete but cannot open without power, so no rent comes in while interest keeps building on a larger-than-planned loan. Oracle would still cover some project costs in the meantime, but it is not clear if those costs include interest [18]. That gap falls first on the sponsor. In deals like this, the developer and its owner customarily guarantee the cost of carrying the project, including interest, until it opens, which would put the burden on Stack and Blue Owl rather than the banks [6]. The banks are exposed only if that guarantee is not honored, which is a real question on a project this size, since developers’ completion and carry guarantees are rarely sized for multi-year delays. In that case, the banks would have to choose between extending the loan at a higher price and declaring a default on a project whose tenant still owes the full lease. 

None of this has happened, and the notice is not a default under the loan [18]. If the deferral were granted, the rent itself would not be lost, since Oracle would still owe it for the full term. The fight would be over who funds the three years in between, whether that is Blue Owl putting in more equity, the banks carrying a larger loan for longer than they underwrote, or both. That uncertainty, alongside higher rates, has weighed on the loan. 

This is where Jupiter connects back to Phase III. Rating agencies took comfort in Sopaipilla’s triple-net lease because it placed construction, power, and timing risk on Meta [20]. Jupiter’s lease does the same with Oracle, and goes further than most by denying it any right to terminate [19].  Jupiter therefore shows that a lease can assign a risk to the tenant on paper while the tenant argues that a particular version of that risk falls outside its control. 

This is a different problem from TeraWulf’s, where the backstop simply does not apply during construction. At Jupiter, the question is how an existing risk allocation will be interpreted. Both cases nonetheless turn on the same issue, which is who bears the cost while a campus is late. If Oracle’s position holds, part of the timing risk the structure was built to push onto the tenant would move back onto the developer’s equity and, if the developer cannot carry it, its lenders. If it does not hold, the risk stays with Oracle, and the project’s protection becomes only as good as Oracle’s credit. Either way, the contract works as written, and the risk ends up somewhere other than the asset the lenders financed. 

As of late October, the buildings at Jupiter are reported to be nearing completion but still have no power. The fuel cell plant meant to supply it cannot be built until New Mexico grants an air quality permit. The hearing on that permit has been pushed to December 7 and could slip again, and opponents have separately asked for the application to be dismissed [29]. Oracle has responded by announcing a redesign it says cuts emissions by about 35%, and by proposing a backup power deal with El Paso Electric. Under that proposal, Oracle would fund up to $55mm a year in customer credits for up to three years while the main supply is resolved. The gas pipeline that would fuel the plant, meanwhile, faces a new federal review [30]. The questions that decide whether the force majeure notice matters are therefore still open: whether the permit is granted and when, whether the pipeline is approved, and whether a temporary power arrangement can get the campus running before 2028.

Not an Isolated Case:

Jupiter drew the headlines, but the bond market had been pricing the same kind of risk across several other projects. Its approach has largely been to sell first and ask questions later. Uncertainty over how serious a permit delay is, how far a lawsuit or new regulation reaches, and at times how little the developer discloses has been enough to move prices. 

Tract Capital (a developer) offers the clearest example. In early 2026, a Tract subsidiary set out to raise $3.65bn for a 200-megawatt campus outside Reno that will be leased to Nvidia. Demand was strong enough that the deal was upsized to $3.8bn and its coupon cut from 6.00% to 5.875%. A few months later, Tract raised another $4.6bn at 6.500% for a second Reno campus also leased to Nvidia. Since then, the $3.8bn notes due March 2031 have fallen about 11 points to ~90, and the $4.6bn notes due May 2031 have fallen 7 points to ~92 [21]. 

Higher interest rates explain part of that decline but not all of it. Tract has not yet received the air permits it needs for gas turbines and backup generators at the two sites. NV Energy, Nevada’s largest utility, sued Tract in August, alleging it was trying to shift the cost of its planned power buildout onto ordinary customers [21]. Some of that pressure has since eased. Nevada’s utility regulator has conditionally approved Tract’s on-site power plan, and concern that the lawsuit would reach the two financed campuses has largely faded. Nevertheless, Nevada’s unpredictable regulation and broader weakness in AI-related credit continue to weigh on the bonds.    

The pattern repeats elsewhere. A Coatue and Fluidstack project in Oklahoma being developed for Jane Street ran into trouble obtaining a floodplain permit. The project’s $2.25bn of bonds, issued in August at 8.875%, now yield more than 11%. In Texas, $3.5bn of high-yield bonds for the Helios campus fell about 4 points in the days after Governor Abbott froze new data center permits [21]. Both concerns have since softened. The Oklahoma project received conditional floodplain approval, and investors concluded that the Helios expansion was largely shielded from the Texas freeze because its application was already in the state’s first review batch. The selloffs still showed how quickly these bonds react to headlines before the facts are clear. 

None of these projects have the same structure as Jupiter, and none of their tenants have issued a force majeure notice, but taken together they show the market pricing the gap between when a project was supposed to be finished and when it will actually receive its permits and power. That gap matters more for these projects than for Jupiter. Most data center leases, unlike Jupiter’s, let the tenant terminate if delivery slips far enough, a far worse outcome for lenders than a deferral [21]. 

The clearest exception shows what avoiding that gap requires. SpaceX’s AI unit, formerly xAI, powered its Colossus campuses in Memphis by running dozens of trailer-mounted gas turbines across the state line in Southaven, Mississippi, before a permanent air permit was in place. It relied on an exemption for temporary mobile units, and it is now replacing them with a permanent 1.2-gigawatt plant permitted in March 2026 [34]. The approach drew lawsuits from environmental groups, but it got capacity running far faster than the projects above, capacity SpaceX now uses for its own models and rents to other AI companies, including Anthropic. Crucially, it was funded largely with the owner’s own capital rather than with project debt. The permitting and power risk did not disappear; it sat with an equity owner willing to carry it, not with bondholders waiting for an approval.

The GPU Layer Under Strain:

The projects above faced pressure from permitting and power delays. CoreWeave faced a different challenge in October 2026: the cost of financing its expansion. The company’s second-quarter revenue more than doubled to $2.6bn, and its contracted backlog reached approximately $104bn [22]. Delivering that backlog, however, required expanding from roughly 1.5 gigawatts of operating capacity to more than 4 gigawatts under contract, with planned capital spending of $35–39bn in 2026 [23]. In September, CoreWeave returned to the capital markets with a share sale program and $4.2bn of convertible notes [24]. 

That reliance on external financing drew greater scrutiny in mid-October. The pressure showed up at the parent, CoreWeave, Inc., rather than in the subsidiary loans tied to individual customer contracts. The yield on CoreWeave’s 9.625% notes due 2032 touched 15% on October 13, its shares fell 15.5% the following day, and the CDS (cost of protection against a CoreWeave default) reached a record high. The selloff followed comments from a co-founder that the company was exploring investment-grade borrowing, raising concerns about how much additional debt its expansion would require [25]. Importantly, the parent’s bondholders are not secured by any particular batch of chips or customer contract. They are repaid from whatever cash the subsidiaries send up after servicing their own loans, and they sit alongside any guarantee claims from those subsidiaries if things go wrong [10]. In effect, the parent’s bonds finance the equity in each subsidiary, the slice that is paid only after that subsidiary’s own lenders, and the selloff reflected investors recognizing as much. 

The repricing did not follow a reported disruption in CoreWeave’s customer contracts. Instead, it highlighted the distinction between contracted demand and the capital required to deliver it. Hyperscaler commitments support individual financing facilities, but CoreWeave remains responsible for funding its broader expansion and provides recourse on many of its loans. Even with a substantial contracted backlog, higher borrowing costs can make the next round of capacity more expensive to finance. CoreWeave has funded itself through subsidiary loans, unsecured bonds, convertible notes, and share sales. At current trading levels, the unsecured bond market is effectively closed to it, leaving secured subsidiary loans as the main route for new buildouts in the near term.

That makes October largely a story for CoreWeave’s shareholders and, at worst, its unsecured creditors, rather than for lenders to the individual chip subsidiaries. No subsidiary is at risk of failing to deliver because of the selloff. Market participants describe the real question as how much dilution, or how much more expensive secured debt, CoreWeave must accept to keep growing at its current pace.

The Restructuring Questions

A restructuring would raise a different set of questions from the delays and repricing described above. For a hyperscaler, AI infrastructure is one large line of spending inside a highly profitable core business. If the economics disappoint, it can slow or stop that spending and still have its advertising, cloud, or software franchise to fall back on. It also has strong reasons not to let a project it stands behind default, since the damage to its credit rating would cost far more than the project itself. A neocloud has no such cushion. Its entire business rests on AI computing being worth more than it costs to finance, so a downturn in AI demand is a threat to the company itself rather than to one division. That difference is why the restructuring questions in this market center on neoclouds, and they look different at each layer. On the building side, they turn on what happens to a developer when a neocloud tenant stops paying rent. On the GPU side, they turn on what happens to lenders when a neocloud’s own loans go bad. 

The Building Layer: 

Several recent building financings depend on a neocloud as the tenant. For example, Galaxy Digital (a developer) raised approximately $3.5bn of debt at around 10% for its Helios campus, which is leased to CoreWeave [27]. The developer’s bondholders are therefore relying on CoreWeave’s rent, and CoreWeave’s ability to keep paying it depends on the ability to sell compute and the AI economics described above. 

If a neocloud tenant enters Chapter 11, it may be able to reject leases it no longer wants under Section 365 of the Bankruptcy Code. The developer would generally be left with an unsecured claim for rejection damages, capped by Section 502(b)(6). The cap is the greater of one year’s rent or 15% of the remaining rent, not to exceed three years’ rent, plus certain unpaid rent owed at filing. 

Consider Campus Co. (a developer) and its tenant, Compute Co (a neocloud). Compute Co. rents Campus Co.’s building for $500mm a year, fills it with chips financed by its own lenders, and sells the computing capacity to AI labs under contracts lasting a few years. Its rent and loan payments are fixed, but its revenue is not. Suppose AI revenue grows more slowly than expected. As Compute Co.’s initial contracts expire, customers either walk away or renegotiate at lower prices, and newer chips push down what its older hardware can earn. Compute Co. can no longer cover both its rent and its loan payments, cannot cleanly borrow more to bridge the gap, and files for Chapter 11. 

In bankruptcy, if Compute Co. rejects the lease, the $6bn of remaining rent becomes a capped claim of roughly $900mm (15% of remaining rent), assuming flat rent. Because that claim is unsecured, Campus Co. is likely to recover only a fraction of it, and it is left holding a building with no rent. 

Campus Co.’s best option is to find a new tenant, but the same conditions that sank Compute Co. make that harder. If AI computing has turned out to be less profitable than expected, fewer companies will want to rent a large, power-hungry campus, and those that do will likely have the leverage to demand lower rent. A partially built campus, or one sitting idle while creditors negotiate, is worth even less. For holding company lenders, recovering even that value likely runs through pledged equity rather than foreclosure.

Physical factors could make re-leasing even more difficult.  Many are built in remote locations chosen for available power, and they are designed for power densities and cooling systems that few tenants beyond the largest AI operators can use. A large, energized power connection is itself scarce and valuable, which may partly offset that. For structures relying on a residual value guarantee, the result is uncomfortable, as the guarantee is likely to be called on most heavily exactly when finding a replacement tenant is hardest. The same risk now shapes how the debt is sized. Data center loans were traditionally sized on the assumption that the tenant would renew, which allowed more borrowing. Lenders have since moved away from that assumption. Sponsors therefore face smaller refinancings and, potentially, having to put in more equity to bridge the difference, at returns that may no longer justify it. 

A second building-layer question is whether arrangements documented as leases would keep that characterization in bankruptcy. Courts look at economic substance rather than labels. They have recharacterized some sale-leaseback and lease-leaseback transactions as financings. That does not mean the Meta transactions face the same risk. Debt-sized rent and a residual value guarantee are, however, the kinds of features those courts examined, and the outcome would turn on the specific agreements and applicable law. 

The GPU layer: 

On the GPU side, the first question is where a subsidiary’s losses go. For the Meta-backed facility, lenders’ recovery would largely stop at the subsidiary’s chips and contract. For CoreWeave’s other GPU loans, which carry parent guarantees, any shortfall becomes a claim against CoreWeave itself [10]. In a hypothetical CoreWeave restructuring, those guarantee claims would compete with the parent’s unsecured bondholders for the same pool of value. 

The second question is what the collateral is worth, and lenders do not agree on the answer. Some large private credit investors now view chips as riskier collateral than the buildings that house them. Their concern is whether the chips could actually be put back to work if the neocloud running them failed, even when they sit in a separate subsidiary. A finished building can often be re-leased, but a GPU’s value depends on what the next customer will pay to rent hardware that may be a generation or two old, and estimates of a chip’s useful life run from roughly seven years down to as few as three [1]. 

CoreWeave depreciates its GPUs over six years, which much of the market considers too long. Fascinatingly, current pricing suggests the opposite. Chips two generations old are still renting at high rates, and today a lender that took control of chips whose contracts had been terminated could likely rent them out on the spot market for well above the contracted price. Older chips also remain useful because, within limits on power and space, a given number of them can do the same work as fewer newer ones. What the market doubts is not current prices but how long spot demand will stay that deep. That depends heavily on whether AI labs such as OpenAI and Anthropic grow their revenue into their commitments. It also depends on whether custom chips from Google, Amazon and others, or a breakthrough in chip design, shorten the useful life of today’s Nvidia hardware. 

The scenario in which lenders need to sell chips is also the one in which rental rates have likely fallen, and other neoclouds are trying to sell the same hardware. The support layered on top may not survive a filing either. Nvidia’s agreement to buy CoreWeave’s unsold capacity allows either party to terminate if the other enters bankruptcy proceedings that are not resolved within 90 days [26].

Where the Layers Meet:

Neoclouds that both borrow against their chips and lease their buildings sit on both sides of this map. They are borrowers at the GPU layer and tenants at the building layer. A hypothetical neocloud filing would therefore test both sets of protections at once. Their own lenders would turn to guarantees and used chips, while bondholders behind campuses they lease would turn to capped rejection claims and buildings without a tenant. CoreWeave is the largest of these companies and the least likely to reach that point, but several smaller neoclouds have built the same two-sided exposure on thinner balance sheets and with fewer investment-grade customers, so a filing is more plausible among them. None of this has happened, but it does show how risk that left the hyperscaler’s balance sheet through two separate channels could come back together in a single company.

The supports themselves are also less independent than they appear. Nvidia sells CoreWeave its chips, owns part of its equity and stands behind its unsold capacity, while also committing to invest heavily in OpenAI. OpenAI, in turn, is among the largest customers behind both Oracle’s and CoreWeave’s contracted backlogs [18]. More broadly, the revenue of a few AI labs, OpenAI and Anthropic above all, ultimately pays for much of the capacity that neoclouds and hyperscalers alike are building, since the hyperscalers rent much of their own capacity to those same labs. Each layer of protection was added to make a weaker credit look stronger, but many of them rely on the same few companies and on the same assumption that AI demand keeps growing. A shortfall in that demand would not test them one at a time. It would test them together, at the moment lenders most need them to be independent. 

Conclusion

So far, the contracts have held. Oracle has not walked away from Jupiter, none of the tenants behind Tract, Oklahoma, or Helios have issued a force majeure notice, and CoreWeave’s customers kept paying through the selloff in its bonds. What 2026 has shown is that contracts working exactly as written do not keep lenders whole. Jupiter’s $18bn construction loan trades below 90c, CoreWeave’s bonds have sold off, and new deals across the sector now need wider spreads to clear. While these represent losses in market value and higher funding costs rather than defaults, they are real for anyone who has to mark, sell, or refinance. 

That is the sense in which, as of today, the risk is being priced: unevenly, event by event, as delays reveal where protection actually sits. A hyperscaler’s name in a structure says less than when its obligation begins, which entity it supports, and how much of the debt it covers. For lenders, those questions now matter as much as the identity of the company providing the protection.

The larger test may come from refinancing rather than from any single project. Much of the bank debt that built the last wave of campuses was meant to be refinanced once those campuses were finished. The markets that would take it out are either too small, in the case of securitization, or only partly open, in the case of investment-grade project bonds. If that gap is not closed over the next year or two, the cost will likely be pushed back onto the hyperscalers, through higher rents or more direct support, or the pace of building will slow. Either way, the question for lenders will be the same one this piece has followed throughout: whose promise is actually standing behind the debt when it comes due.

Pari Passu is an independent publication covering leveraged finance, restructuring, and credit markets. This write-up is general commentary distributed on a regular basis to all subscribers. It is provided for informational and educational purposes only, is not tailored to any reader, and is not investment advice or a recommendation to buy, sell, or hold any security. Pari Passu is not a registered investment adviser or broker-dealer. Readers should not rely on this material as a basis for any investment or business decision and should conduct their own diligence.

This write-up draws on publicly available information, conversations with market participants, and Pari Passu's own analysis. Unless a figure is identified as coming from a filing or other public disclosure, it is an estimate prepared by the Pari Passu Research Team. Estimates have not been verified by, and should not be attributed to, the company or any other party discussed. The views expressed are the author's own as of the publication date and may change without notice.

Sources (Visible to Research subscribers only)

Data Center Financing FAQ

Who is paying for AI data centers?

The five largest US technology companies are expected to spend $660-690bn on capital expenditures this year, with roughly $450bn going directly into AI infrastructure. Even after counting the cash those companies generate themselves, Morgan Stanley estimates a funding gap of about $1.5tn through 2028 that outside capital must fill. A growing share of the financing now comes from bond investors, private credit funds, and banks.

How are data centers financed?

Data center financing has moved through three stages. Through 2022, hyperscalers paid their own way from cash flow and investment-grade bonds. From 2023 to 2025, they rented chips from neoclouds and backstopped leases. Since October 2025, joint ventures have owned the campuses and borrowed against them, with the hyperscaler as tenant. Each stage lets hyperscaler commitments support debt that the hyperscaler does not itself owe.

How much does it cost to build an AI data center?

Building the shell and its infrastructure costs roughly $10-16mm per megawatt, and the all-in cost rises to roughly $40-60mm per megawatt once servers and GPUs are installed. For a 1-gigawatt campus, the building alone costs $10-16bn and the fully equipped campus costs $40-60bn. Roughly two-thirds of the total cost sits in equipment that will be obsolete long before the building is.

What is a hyperscaler?

Hyperscalers are the handful of companies that run the world’s largest cloud and AI platforms: Amazon, Microsoft, Google, Meta, and Oracle. They build and operate massive facilities, primarily for their own use. With few exceptions, they are investment-grade borrowers with long histories in the bond market, which lets them borrow cheaply. Oracle is the outlier, having borrowed far more heavily than its peers.

What is a neocloud?

Neoclouds are a newer group of companies that buy GPUs in bulk and rent computing capacity to AI labs, startups, and enterprises. CoreWeave and Fluidstack are the best-known examples, alongside Crusoe and Lambda. Most lease their buildings from developers and finance their chips largely with debt. Unlike hyperscalers, most are young, heavily indebted companies that rely on a handful of large customers.

How long do GPUs last?

Estimates of a chip’s useful life run from roughly seven years down to as few as three. CoreWeave depreciates its GPUs over six years, which much of the market considers too long. Current pricing suggests the opposite, since chips two generations old are still renting at high rates. What the market doubts is not current prices but how long spot demand will stay that deep.

Is AI data center debt a bubble?

So far, the contracts have held, and none has failed to pay. The risk is concentration. Many layers of protection rely on the same few companies and on the same assumption that AI demand keeps growing. The revenue of a few AI labs, OpenAI and Anthropic above all, ultimately pays for much of the capacity. A shortfall in that demand would not test those protections one at a time. It would test them together.

Have any AI data center loans defaulted?

Not so far. Oracle has not walked away from Jupiter, and CoreWeave’s customers kept paying through the selloff in its bonds. But contracts working exactly as written do not keep lenders whole. Jupiter’s $18bn construction loan trades below 90c, and CoreWeave’s 2032 notes touched a 15% yield. These are losses in market value and higher funding costs rather than defaults.

Why do hyperscalers finance data centers off balance sheet?

Capital spending has been growing about 70% a year against roughly 23% for operating cash flow. Every dollar borrowed at the corporate level lands on the company’s balance sheet, raising leverage and pressuring the credit ratings that make its debt cheap. Keeping the debt off the balance sheet lets a hyperscaler keep building without that pressure.

How did Meta finance its Hyperion data center?

In October 2025, Meta and funds managed by Blue Owl Capital formed a joint venture, owned 80% by Blue Owl and 20% by Meta, which raised more than $27bn of debt. The deal is known in the market as Project Beignet. The notes were rated A+, mature in 2049, and priced at roughly 225bp over Treasuries. Because Meta does not control the joint venture, the debt stays off its balance sheet.

What is a residual value guarantee (RVG)?

The tenant promises that the campus will be worth at least an agreed amount if the lease ends, and pays any shortfall up to a cap. Suppose the agreed amount is $8bn, the campus is appraised at $5bn, and the cap is $2bn. The tenant pays $2bn, and the remaining $1bn falls on the joint venture and, ultimately, its lenders. The tenant is guaranteeing the building’s value, not the debt itself.

What is Oracle’s Project Jupiter force majeure notice?

On September 24, 2026, Oracle sent developer Stack Infrastructure a notice citing force majeure on its 2.45-gigawatt New Mexico campus. Oracle did not try to walk away from the lease. It sought to preserve the right to delay rent by up to three years if power delays push the campus past its planned 2028 opening. The notice is not a default, but the roughly $18bn construction loan trades below 90c.

Who is paying for OpenAI’s Stargate data centers?

Take Project Jupiter, a 2.45-gigawatt Stargate campus in New Mexico. Stack Infrastructure is developing it, Blue Owl funds provided roughly $3bn of equity, about 20 banks provided an $18bn construction loan, and Oracle signed as the tenant. The legal obligation to pay rent sits with Oracle, but the economic reason for the rent rests heavily on OpenAI, which is still funding its own growth through outside capital.

How does GPU-backed financing work?

A neocloud buys and installs chips and finances them largely with debt repaid from long-term take-or-pay customer contracts. CoreWeave became the template in August 2023 with a $2.3bn facility secured by Nvidia H100 chips. Lenders are really underwriting the contracts rather than the chips. CoreWeave’s $8.5bn Meta-backed loan priced at S+2.25%, compared with S+13.00% on loans backed by weaker customers.

What happens if a neocloud data center tenant files for bankruptcy?

In Chapter 11, a neocloud tenant may be able to reject leases under Section 365 of the Bankruptcy Code. The developer is left with an unsecured claim capped by Section 502(b)(6): the greater of one year’s rent or 15% of the remaining rent, not to exceed three years’ rent. On $6bn of remaining rent, that is roughly $900mm, and the developer is left holding a building with no rent.

Keep Reading

View more