Client Alert: Financing the AI Infrastructure Boom: Emerging Litigation Risks in AI Data Centers
Sign Up for PublicationsI. Summary
The AI data center buildout—projected to require $5.2 trillion in infrastructure investment by decade's end—has spawned complex financing structures that are generating significant litigation risk. With AI revenues (roughly $60 billion in 2025) falling far short of capital expenditures (roughly $400 billion), technology companies have turned to corporate bonds, private credit, and off-balance-sheet SPVs to bridge the gap, moving more than $120 billion in data center spending off their balance sheets in under two years.
This note surveys the major financing mechanics—direct loans, SPV structures, securitizations, and GPU-collateralized facilities—and identifies nine categories of emerging litigation risk: (1) defaults and insolvency cascades across interconnected capital stacks; (2) securities fraud claims driven by off-balance-sheet opacity; (3) credit ratings litigation echoing post-2008 RMBS suits; (4) structured finance disputes over credit enhancements that may fail to trigger when needed; (5) valuation and margin call fights over rapidly depreciating GPU collateral; (6) construction and power contract disputes tied to aggressive build timelines; (7) investment treaty arbitrations as the buildout globalizes; (8) take-or-pay contract disputes as anchor customer commitments become unstable; and (9) environmental and community litigation over energy and water demands.
II. Introduction
Data centers form the backbone of the AI revolution. Large language models like OpenAI’s ChatGPT and Anthropic’s Claude consume enormous amounts of computing power, delivered by specialized processors housed in dedicated data centers. Each year, as these models advance, they require exponentially more computing power.[1] The burgeoning demand for compute has spurred a race to build AI data centers at an astounding price tag. The cost of AI infrastructure to meet the global demand for computing power by the end of the decade is estimated at $5.2 trillion, with much of the spending in the United States.[2]
Even for the world’s most profitable companies, internally generated cash is insufficient to fund capital outlays of this magnitude. In 2025, capital expenditures for Alphabet, Amazon, Meta, and Microsoft reached an all-time high at $381 billion. This year, spending is forecasted to increase, by over 60% from those historic levels, to $700 billion,[3] resulting in the dramatic erosion of free cash flow. Amazon, for example, is projected to generate negative free cash flow of up to $28 billion in 2026, while both Alphabet and Meta are expected to see free cash flows drop by roughly 90% in 2026.[4] Cash reserves are also rapidly dwindling.[5]
The AI sector has thus turned to public and private debt markets to finance the data center buildout.[6] New borrowing in 2025 reached at least $200 billion, which analysts admit likely underestimates the aggregate debt issuance as many deals are private.[7] By one estimate, around $1.5 trillion in external financing is needed across the AI ecosystem by 2028 to bridge the gap between cash flows and capital expenditures.[8]
The scale and complexity of the financing structures underpinning the AI data center buildout are already raising concerns among regulators, investors, and policymakers. For example, in January 2026, four U.S. Senators urged in an open letter that regulators investigate the AI sector’s growing reliance on debt.[9] The letter warned that companies are increasingly turning toward “complex and opaque debt markets to borrow staggering sums of cash” and their inability to service the debt “could cause destabilizing losses for an interconnected set of financial institutions, triggering a broader financial crisis that harms the economy.”[10] The Federal Reserve and the Bank for International Settlements have expressed similar concerns.[11]
The foundational risk is straightforward: revenues generated by AI services may prove insufficient to service the massive debt loads incurred to build the infrastructure supporting them. But the most significant legal consequences of the AI buildout may stem not simply from the amount of debt involved, but from how that debt is structured. Much of the capital funding AI data centers flows through layered financing arrangements involving private credit funds, special purpose vehicles (SPVs), securitization vehicles and conditional credit guarantees. These structures frequently separate economic risk from operational control, obscure the true leverage of the underlying technology companies, and distribute exposure across a complex chain of lenders, investors, and institutional intermediaries. When financial stress occurs—whether through declining GPU values, construction delays, or reduced demand for computing capacity—those structural features transform ordinary financial distress into multifaceted legal disputes.
This note examines the litigation risks emerging from the substantial and increasing reliance on debt to finance the AI data center buildout. Part I provides a brief overview of AI data centers, explaining what they are and who builds them. Part II surveys major recent transactions to illustrate the size and complexity of the borrowing. Part III explains how these deals are typically structured, from the sources of capital through the vehicles that deploy it and the collateral that secures it. Part IV identifies current lawsuits and litigation risks associated with the buildout.
III. A Brief Overview of AI Data Centers
The data centers that run AI house densely packed server racks filled with processors, primarily graphics processing units.[12] GPUs provide the computational throughput required to train and deploy large language models. Unlike traditional data centers built around central processing units for general-purpose computing, AI-focused data centers are equipped with high-bandwidth networking cables to connect thousands of GPUs in parallel, advanced cooling systems to dissipate the intense heat generated by GPU clusters, and enormous power capacity to sustain continuous computational workloads, estimated for data centers coming online between 2024 to 2035 at roughly ten times New York City’s summer peak electricity demand.[13]
A complex ecosystem of actors has coalesced to meet this demand for GPU capacity. At the center of the web sits GPU manufacturers, principally NVIDIA and AMD. Surrounding them are technology giants, referred to as “hyperscalers,” that are investing hundreds of billions of dollars in AI infrastructure (e.g., Alphabet, Amazon, Meta, Microsoft, and Oracle). Alongside them are smaller companies called “neoclouds” specializing in providing GPU capacity tailored for AI workloads (e.g., CoreWeave, Fluidstack, and Lambda). Hyperscalers and neoclouds alike contract with developers and operators like TeraWulf, Digital Realty, and Equinix to retrofit existing facilities or construct new ones and to run the data centers once completed. Those developers, in turn, depend on various contractors and equipment manufacturers as well as utility providers to deliver and maintain the physical infrastructure.
Several business models for AI data centers have emerged to meet varying needs. The most capital-intensive are the hyperscale data centers, which are fully integrated facilities where hyperscalers or their affiliates own the building and the servers.[14] A middle ground is the colocation data centers where developers provide the physical shell and tenants lease space to install their own servers.[15] Finally, the GPU-as-a-service model bypasses hardware ownership entirely, by allowing companies to purchase access to computing power from cloud service providers like hyperscalers and neoclouds.
IV. Illustrative Transactions
Several recent transactions illustrate both the scale and complexity of the data center financing. Each deal highlighted below exhibits structural features that are further explained in Part III and give rise to distinct categories of risks as summarized in Part IV.
A. Corporate Bonds and Private Credit: Oracle’s Stargate Data Centers
Oracle’s role as the primary infrastructure provider for the Stargate project—the $500 billion AI initiative led by Oracle, OpenAI, and SoftBank—illustrates the two primary sources of debt capital. The first is bond issuances. In September 2025, Oracle sold $18 billion in publicly traded bonds in a single day to fund its data center commitments to OpenAI.[16] The second is private credit. Oracle has partnered with various financiers to construct numerous data centers through SPVs that Oracle then leases back.[17] These off-balance-sheet arrangements include approximately $13 billion invested by Blue Owl and JPMorgan ($10 billion as debt) into an SPV that owns Oracle’s OpenAI facility in Abilene, Texas; a $38 billion debt package funding two data centers in Texas and Wisconsin; and an $18 billion loan financing a site in New Mexico.[18]
B. SPV Financing: Meta’s Hyperion Data Center ($30 Billion)
Meta’s Hyperion transaction illustrates the use of off-balance-sheet SPV financing. In October 2025, Meta completed the largest private credit data center deal in history: a $30 billion agreement for its proposed Hyperion facility in Louisiana.[19] The transaction created an SPV called Beignet Investor, co-owned by Meta (20%) and Blue Owl Capital (80%). The SPV raised $30 billion, including approximately $27 billion in loans from Pimco, BlackRock, Apollo, and others, plus $3 billion in equity from Blue Owl.[20] The SPV owns the data center and leases it to Meta, which allowed Meta to, in effect, borrow $30 billion without any debt appearing on its balance sheet. Instead, Meta records only its equity stake in the SPV and its lease obligations; the full $27 billion in underlying loans does not appear as a liability on Meta’s financial statements. Meta has also given a “residual value guarantee” to SPV investors, obligating it to compensate them if the data center’s value drops below a specified threshold—a commitment of up to $28 billion that is described only in footnotes to Meta’s most recent annual report and for which no liability has been recorded on Meta’s balance sheet.[21]
Days after closing the SPV transaction, Meta raised an additional $30 billion in the corporate bond market.[22]
C. Credit Backstops: The Google/Fluidstack/TeraWulf Transaction ($3.2 Billion)
The arrangement among Google, Fluidstack, and TeraWulf to build a data center in upstate New York illustrates the use of corporate bonds supplemented by conditional credit enhancements. In October 2025, TeraWulf, a former bitcoin miner, raised $3.2 billion through a high-yield bond offering to build a data center in Barker, New York.[23] The bonds, rated Ba2/BB, carry a coupon of 7.75% and were heavily oversubscribed, with investors placing roughly $10 billion in orders.[24] The deal is structured as follows: Fluidstack will lease TeraWulf’s facility under ten-year agreements to operate GPU clusters for AI customers.[25] As a credit enhancement, Google agreed to backstop Fluidstack’s lease obligations, stepping in to pay a termination fee or assume the leases directly if Fluidstack defaults or becomes insolvent. In exchange, Google received equity warrants for up to 14% of TeraWulf. Google has replicated this structure across multiple transactions, committing more than $5 billion in backstop obligations through Fluidstack to former bitcoin miners pivoting to AI infrastructure.[26]
D. GPU-Collateralized Borrowing: CoreWeave’s GPU-Backed Facility ($7.5 Billion)
CoreWeave’s $7.5 billion debt facility, announced in May 2024 and led by Blackstone’s tactical opportunities group, illustrates another attribute of AI data center financing: GPU-collateralized lending.[27] The facility used the company’s GPUs and customer contracts as security for the loans, divided between investment-grade and speculative-grade tranches. The variable interest rate averaged approximately 11%, with repayments beginning in January 2026—just as the collateral’s market value was declining sharply.
V. Financing the AI Data Center Buildout
The AI data center buildout is financed through multiple debt sources with differing financial structures. This Part moves from the sources of capital to the vehicles that deploy it, to the securitization markets that distribute it, and finally to the collateral that secures it.
One major source of capital is corporate bond issuances. Hyperscalers issued approximately $121 billion in bonds in 2025 alone—more than four times the five-year average—with AI-related investments accounting for roughly 30% of net issuance in the U.S. dollar-denominated investment-grade corporate bond market, and total data center debt issuance nearly doubling to $182 billion.[28]
The main source of external financing is private credit. Private credit funds—principally Blackstone, Blue Owl Capital, Apollo, Pimco, and BlackRock—originate most of the data center debt, both through off-balance-sheet SPV transactions and through direct lending facilities to neoclouds and data center developers.[29] Private credit funds have directed a rapidly growing share of their lending toward AI infrastructure, with outstanding loans to AI-related companies surging from near zero to over $200 billion in just a few years—and Morgan Stanley projects that private credit will provide an additional $800 billion in data center financing over the next two years alone.[30]
Traditional banks and other institutional and retail investors play a smaller but increasingly significant role. A February 2026 study by the Federal Reserve Bank of Chicago found that banks’ outstanding exposure to AI-adjacent industries averaged just 0.8% of total assets—but cautioned that banks “most likely have additional exposure to AI-adjacent industries through lending to nonbank financial institutions,” including the very private credit funds that originate most data center debt.[31] A separate Federal Reserve Board study quantified this indirect channel, finding that up to a quarter of bank loans to non-bank financial institutions are now made to private credit firms, up from just one percent in 2013, and that major life insurance companies have nearly $1 trillion tied up in private credit.[32] The chain of exposure extends further still: New York and Pennsylvania state pension plans have invested in Blue Owl’s $7 billion digital infrastructure fund—the same fund behind Meta’s Beignet Investor SPV and multiple Oracle data center financings—and in August 2025, President Trump signed an executive order instructing federal agencies to loosen regulations so that ordinary 401(k) holders can invest directly in “alternative assets” such as private credit.[33] The result is a transmission chain that runs from the technology companies, through private credit originators, to the regulated banks that lend to them, to the insurers and pension funds that invest alongside them, and potentially to the retirement accounts of ordinary Americans.
The capital described in the preceding section reaches data center projects through several financing mechanics.
The most straightforward financing mechanic is the direct loan. Data center developers typically rely on “mini-perm” construction loans—facilities with terms of two to five years that finance a project’s construction and its first years of operation as it reaches full cash flow.[34] Banks originate these loans and syndicate portions to other banks and institutional investors to distribute the risk. In 2025, the scale of these transactions changed dramatically: of the roughly $950 billion in global project finance loan issuance, approximately $170 billion was for data-center-related transactions—an increase of 57% from the prior year.[35] Oracle’s Stargate-related financings illustrate the new magnitude: a $38 billion syndicated facility led by JPMorgan funds two data centers in Texas and Wisconsin, and a separate $18 billion syndicated loan finances a site in New Mexico—both structured as four-year mini-perm facilities priced at approximately 250 basis points above the risk-free Treasury rate.[36] When the mini-perm matures, developers must refinance into longer-term instruments—typically through the ABS or CMBS markets discussed in Section B.3—which requires demonstrated tenant cash flows. Project finance lawyers have noted, however, that many data center loans are underwritten against tenants’ cash flows on “booked-but-not-billing” terms—meaning the promised revenue need not have actually materialized.[37]
Much of the data center buildout has been financed through off-balance-sheet structures. Technology companies have moved more than $120 billion of data center spending off their balance sheets in approximately eighteen months.[38] The prevailing structure works as follows: a technology company partners with a private credit fund to create an SPV—a separate legal entity designed to be “bankruptcy remote,” meaning its assets and liabilities are ring-fenced from the parent company’s balance sheet. The technology company may then transfer assets to the SPV, such as GPUs or other equipment, in a transaction that must qualify as a “true sale” rather than a disguised financing arrangement. This distinction is critical: if the transfer is a true sale, the assets belong to the SPV and are beyond the reach of the technology company’s creditors in a bankruptcy; if it is later recharacterized as a secured loan, the assets may be pulled back into the bankruptcy estate, collapsing the structural protections on which lenders relied. The SPV raises debt from institutional lenders, uses the proceeds to construct or acquire the data center, and upon completion, leases the facility back to the technology company. Because the SPV is a distinct legal entity, the debt does not appear on the technology company’s balance sheet, allowing the company to appear less leveraged than it is and to preserve capacity for additional borrowing.[39]
Data center debt has increasingly been repackaged and sold to a wider pool of investors through asset-backed securities (“ABS”) and commercial mortgage-backed securities (“CMBS”). Data center ABS were introduced in 2018 as a new instrument class linked to cash flows generated by data center financing and tenant leases.
In the traditional model, data center operators such as Sabey Data Centers, Compass Datacenters, CyrusOne, and STACK Infrastructure transfer their real estate holdings and tenant lease receivables into a bankruptcy-remote SPV, which issues rated notes to investors in tranches of varying seniority and risk.[40] Data centers now account for 61 percent of the $79 billion market for digital infrastructure securitizations across both ABS and CMBS asset classes, and data-center-related CMBS issuance hit a high of $4.5 billion in 2025.[41] S&P Global has identified four principal approaches for rating data center financings: corporate finance, ABS, CMBS, and project finance.[42]
More recently, a secondary securitization channel has emerged. Lenders who originated data center loans—including the private credit facilities described in Section A—have begun pooling those loans and selling tranches to asset managers and pension funds, spreading risk well beyond the original lending institutions.[43] This layering of securitization on top of already complex SPV and private credit structures creates a chain of exposure that extends from the technology company, through the SPV, to the private credit originator, through the securitization vehicle, and ultimately to the pension funds and asset managers holding the rated tranches.
The debt instruments funding the AI buildout are ultimately secured by two categories of collateral, each presenting distinct risks. The first is the data center itself—the real estate, the land, and the facility’s physical infrastructure. In the event of a default on a data center SPV, lenders’ recourse typically runs to these physical assets. But AI-focused data centers are purpose-built facilities with specialized cooling systems, high-density power configurations, and custom layouts designed for GPU server racks. Converting an AI data center to general-purpose cloud computing or other uses is expensive and complex, requiring replacement of GPU-oriented infrastructure with CPU-oriented systems.[44] If demand for AI computing contracts, these facilities may function as stranded assets with limited alternative use and depressed liquidation value.
The second category of collateral—and the one that presents the most concentrated risk—is the GPUs themselves. A rapidly growing class of loans is collateralized directly by GPU hardware, with notable deals including CoreWeave’s $7.5 billion facility, Fluidstack’s $10 billion arrangement, Lambda’s $500 million loan, and Crusoe’s facilities.[45]
VI. Litigation and Arbitration Risks
The financing structures described above present litigation risks along multiple vectors. The deeply interconnected AI ecosystem means that distress at any single node—a construction delay, a tenant default, unhedged energy cost differentials, a collapse in GPU resale values—can propagate across multiple counterparties and financing layers.
The financial relationships among AI ecosystem participants create circular dynamics that inflate apparent demand and obscure underlying risk. For example, OpenAI has agreements to purchase $300 billion in computing power from Oracle, $38 billion from Amazon, and $22 billion from CoreWeave. Those cloud providers, in turn, are major customers for NVIDIA’s chips. NVIDIA, for its part, has invested in CoreWeave (holding a 7% stake) and committed up to $100 billion to OpenAI—investments whose proceeds flow back to NVIDIA through GPU purchases.[46] OpenAI has also invested in smaller AI startups that, in exchange, have agreed to pay for ChatGPT enterprise accounts.[47] These circular dynamics have striking similarities to the “vendor financing” practices that characterized the dot-com bubble of the late 1990s, when telecommunication equipment manufacturers like Nortel and Lucent lent money to their own customers to create the illusion of revenue growth.[48] When Nortel’s customers defaulted, the loans became worthless and the revenues they had generated proved illusory, contributing to one of the largest corporate bankruptcies in Canadian history. AI executives themselves have acknowledged the risk. Meta CEO Mark Zuckerberg has stated that “misspending a couple of hundred billion dollars” would be “very unfortunate”; OpenAI CEO Sam Altman has conceded that “some investors are likely to lose a lot of money.”[49]
Litigation has already begun. On January 14, 2026, a proposed class of bondholders filed suit against Oracle in New York state court, captioned Ohio Carpenters’ Pension Plan v. Oracle Corp., alleging that Oracle’s offering documents for its $18 billion bond issuance were materially false and misleading.[50] On January 12, 2026, a securities class action was filed against CoreWeave in the U.S. District Court for the District of New Jersey, captioned Masaitis v. CoreWeave, Inc., No. 26-cv-00355, alleging that CoreWeave overstated its ability to meet customer demand and understated infrastructure risks.[51] The SEC’s 2024 enforcement actions against Delphia Inc. and Global Predictions Inc. for “AI washing” signal the agency’s willingness to police AI-related misrepresentations more broadly.[52]
These early cases are likely just the beginning. This Part identifies ten categories of risk that, as the financing cycle matures, are likely to generate sustained litigation and arbitration.
A. Distress, Default, and Insolvency Litigation
The most basic litigation risk in AI infrastructure finance is that the revenues generated by the sector may prove insufficient to service the fixed obligations incurred to build it. The industry brought in approximately $60 billion in revenue in 2025 against roughly $400 billion in capital expenditure.[53] If AI revenues do not keep pace with debt service obligations, defaults will not remain isolated at the borrower level. They will migrate up and down the capital stack, producing sustained litigation among issuers, bondholders, lenders, sponsors, SPVs, and, ultimately, bankruptcy estates.
That default risk manifests at every level of the capital structure. At the corporate bond level, default or credit deterioration triggers bondholder litigation—as the Oracle case already illustrates. Both S&P and Moody’s have placed Oracle on negative credit watch;[54] Oracle’s total debt now exceeds $130 billion with $248 billion in new lease commitments;[55] and its five-year credit default swap spread has risen approximately 310%, pushing its perceived credit risk to a sixteen-year high.[56] By January 2026, JPMorgan was struggling to find investors willing to participate in the Stargate debt syndication.[57] In that setting, the predictable claims include disclosure claims by bondholders, covenant-enforcement claims by creditors, and event-of-default litigation once performance falters.
At the private credit level, defaults will trigger lender enforcement actions: debt acceleration, operational control, collateral seizure, and receivership remedies.[58] The recent CMBS enforcement action in U.S. Bank Trust Co. v. Jericho Plaza Portfolio LLC illustrates the template: the CMBS trustee sued the borrower after a default on a $149 million commercial mortgage, obtained the appointment of a receiver to manage the property, and moved toward foreclosure.[59] Data center defaults would follow the same playbook. Developers that miss milestones or fail financial covenants should expect aggressive enforcement by private lenders, including efforts to take control of project companies, collateral, and cash-management accounts.
At the SPV level, defaults will trigger breach-of-representations-and-warranties claims by SPV investors against the sponsors who structured the deals—analogous to the post-2008 RMBS “putback” litigation in which trustees sued sponsors for misrepresenting the quality of assets transferred to securitization vehicles, ultimately recovering more than $36 billion.[60] If sponsors oversold the durability of lease cash flows, understated obsolescence risk, or transferred assets into SPVs at inflated values, investors will argue that the securitized assets were never what they were represented to be.
Cross-default provisions embedded in most data center loan agreements could amplify these losses dramatically. A default on one facility can trigger defaults across the borrower’s other obligations, cascading through the ecosystem: a CoreWeave facility default could trigger cross-defaults across every CoreWeave credit facility, impairing its ability to make lease payments to SPVs, which would undermine the cash flows backing the ABS issued by those SPVs.[61] In other words, what begins as a project level payment default can quickly become a system-wide enforcement event.
If distress leads to a bankruptcy filing (whether voluntary or involuntary), bankruptcy trustees and/or creditors may pursue intentional or constructive fraudulent transfer actions to avoid pre-petition asset transfers to SPVs, recharacterization of “true sales” as disguised loans, substantive consolidation such that the assets and liabilities of the sponsor and SPV are pooled, and equitable subordination of private lenders’ claims.[62] The resulting litigation will focus not only on who gets paid, but on whether the prebankruptcy structure fairly allocated default risks, the source of potential recoveries, whether asset transfers were made for fair value, or with the intent to hinder, delay, or defraud creditors, and whether lenders crossed the line from creditor to controller and engaged in inequitable conduct.
Securities litigation in the AI data center financing ecosystem will be driven principally by opacity. A pervasive risk across the AI data center financing ecosystem is that investors—bondholders, ABS holders, pension funds, 401(k) participants—may not fully understand the obligations, contingencies, and structural dependencies embedded in the instruments they own. This opacity exists at multiple levels.
At the corporate level, companies have used off-balance-sheet SPVs to shift more than $120 billion in data center spending off their balance sheets in approximately eighteen months.[63] Moody’s has warned that “disclosures may not show the full picture” of hyperscaler lease commitments, noting that the accounting treatment of short initial lease terms combined with residual value guarantees may lead reported liabilities to materially understate companies’ true obligations.[64] Meta’s $28 billion residual value guarantee, for example, appears only in footnotes to its annual report and is not recorded as a balance-sheet liability.[65] The resulting claim will sound in classic disclosure theory: investors will argue that public filings and offering documents failed to reveal the issuer’s true leverage, with understated off-balance-sheet commitments. The Oracle bondholder case—alleging that Oracle concealed the full scope of its borrowing plans—is the first example, and the claims will likely sound in Sections 11 and 12 of the Securities Act and Section 10(b) of the Exchange Act.[66]
At the private credit level, the opacity is structural. Unlike regulated banks, private credit lenders are not subject to comparable capital requirements, stress testing regimes, or ongoing supervisory oversight. As Professors Jared Ellias and Elisabeth de Fontenay have observed, credit markets have “gone dark,” migrating from transparent public markets into opaque private channels that complicate systemic risk assessment.[67] When disputes arise between private credit lenders and borrowers, they are typically resolved through bilateral negotiation or confidential arbitration rather than public litigation, meaning that problems may not become visible until they have already metastasized. That creates a second-order litigation risk for downstream investors—pension funds, insurers, and asset managers invested in private credit funds—who may later contend that they were never given a full picture of concentration risk, collateral fragility, customer dependence, or inter-SPV conflicts. If losses materialize and those investors contend that the private credit fund manager failed to disclose material risks, fiduciary duty and misrepresentation claims will follow. Blue Owl Capital, for instance, manages SPVs on behalf of both Meta and Oracle, raising potential conflicts of interest if distress at one SPV affects the firm’s management of the others. Analogous claims were asserted against collateral managers and CDO trustees in the post-2008 litigation.[68]
At the regulatory level, the current trajectory is amplifying rather than mitigating these risks. In August 2025, President Trump signed an executive order instructing federal agencies to loosen regulations so that ordinary 401(k) holders can invest directly in “alternative assets” such as private credit—potentially extending the chain of opaque exposure to a far broader swath of the public.[69] As Professor Natasha Sarin of Yale Law School has warned, “Unfortunately, it usually isn’t until after a crisis that we realize just how interconnected the different parts of the financial system were all along.”[70]
Credit ratings create a distinct litigation risk because they serve as the gatekeeper to ratings-constrained capital. In a potential echo of the 2007-2008 financial crisis, credit ratings agencies themselves may face litigation risk if their methodologies fail to accurately reflect the risk profile for these data center investments and investors rely on those ratings when deciding to invest. Ratings are particularly critical for data center developers because the scale of the buildout has far outpaced the traditional project finance market, forcing developers to seek capital from institutional investors, such as insurance companies, that can only invest in highly rated debt.[71] As a result, an investment-grade stamp of approval from the credit ratings agencies has become essential for many of these projects.
That pressure creates a significant vulnerability as in many AI data center transactions, the ratings may reflect less the intrinsic resilience of the project itself than the creditworthiness of the technology company standing behind it. Yet in many cases, ratings for AI data center projects are driven less by the financial strength of the projects themselves than by the credit profile of the technology companies that will occupy the facilities. Rating agencies have reportedly relied heavily on long-term leases and financial guarantees from major technology tenants, which can substantially bolster a project’s perceived credit quality and effectively cap the project’s rating at the tenant’s own credit rating.[72] S&P, Fitch, and KBRA have issued investment-grade ratings on under-construction projects, including Meta’s $27 billion Hyperion facility, largely based on such guarantees and long-term leases from Big Tech tenants.[73]
The litigation risk arises if those ratings prove to have underweighted the factors that make AI data center assets unusually fragile: construction risk, utility and interconnection delays, tenant concentration, technological obsolescence, and the limited alternative use of purpose-built AI facilities. Given the specialized nature of AI data centers, including remote locations, purpose-built facilities, and limited alternative uses, a rating that proves overly optimistic could mislead investors and trigger litigation similar to that faced by rating agencies after the 2008 financial crisis. Even where direct claims against ratings are difficult to sustain, allegedly flawed ratings will likely become central allegations in investor suits against issuers, arrangers, sponsors and other transaction participants.
Specifically, after the financial crisis, major ratings agencies including Moody’s, S&P, and Fitch faced numerous lawsuits and government enforcement actions alleging that they artificially inflated the ratings of risky mortgage-backed securities, misleading investors about the securities’ true credit risk. For example, Moody’s ultimately reached an $864 million settlement with the Department of Justice and a coalition of states, which alleged that its flawed rating methodology and undisclosed conflicts of interest had misled investors into purchasing high-risk mortgage-backed securities.[74] The U.S. Attorney for the District of New Jersey stated, “People making decisions on how to invest their money thought they could rely on the ratings Moody’s assigned to these products. When securities are not rated openly and honestly, individual investors suffer, as does confidence in all parts of the financial sector.”[75] Moody’s also paid a $130 million settlement with the California Public Employees’ Retirement System, the nation’s largest public pension fund, which had invested $1.3 billion dollars in bonds backed by subprime mortgages that Moody’s had rated AAA, representing the lowest perceived level of risk.[76]
In a role reversal, this year Moody’s has warned that AI data center investments may be riskier than they appear, as current U.S. accounting rules allow companies to conceal tens of billions in potential liabilities.[77] Moody’s noted that reported liabilities may understate likely cash outflows, and the agency plans to adjust its credit ratings to account for these hidden or contingent risks when assessing the financial health of tech companies like Meta and Oracle.
D. Structured Finance Litigation
Structured finance litigation will arise when the credit enhancements used to market data center paper fail to function in the circumstances that matter most. Several of the largest data center transactions have been marketed to investors on the strength of credit enhancements—guarantees, backstops, and residual value commitments designed to shift risk away from bondholders and noteholders. The core risk is that these protections are drafted with conditions, triggers, and valuation mechanisms that cause them to fail-or become contestable—precisely when investors expect them to provide support.
Google’s backstop in the TeraWulf transaction is the most prominent example. The backstop activates only after construction is complete and Fluidstack’s lease commences; during the build-out phase, bondholders bear full project execution risk with no Google support. If construction is delayed by more than 180 days past the target completion date, Fluidstack can terminate its leases entirely—and Google’s backstop obligation would never trigger.[78] The offering memorandum itself contemplates the possibility of a “disagreement” between TeraWulf and Google about triggering the backstop, and acknowledges that Google might not honor its support.[79] Fitch, in rating the deal, acknowledged that “the construction schedule, in our view, is a bit aggressive or accelerated.”[80] If construction delays cause Fluidstack to terminate and Google’s backstop never becomes operative, litigation will focus on who drafted, marketed, and administered that support arrangement, and whether investors were led to believe they had meaningful downside protection when, in reality, that protection could evaporate before it ever attached. Bondholders will likely pursue claims for breach of contract, breach of the implied covenant of good faith and fair dealing, and fraudulent inducement—the theory being that the bonds were marketed with the credit enhancement as a key selling point, but the conditions were structured so that the enhancement was unlikely to trigger when needed.
Meta’s residual value guarantee presents a different variant of the same problem. The Beignet Investor SPV includes a guarantee of up to $28 billion, obligating Meta to compensate investors if the data center’s value falls below a specified threshold at the end of the initial lease term.[81] When that guarantee is tested, the disputes will likely center not simply on whether support exists on paper, but on how value is measured, who controls the appraisal process, and whether discretionary decisions—such as a decision not to renew a lese—were exercised in good faith or instead designed to shift losses to investors. When triggered, the disputes will center on valuation methodology—who appraises the data center, what comparables are used, and whether a purpose-built AI facility has any meaningful comparable at all—and on whether Meta’s decision not to renew the lease was made in good faith or was designed to shift losses to SPV investors.[82]
More broadly, if courts conclude that veil-piercing is warranted—that the parent company controlled the SPV so thoroughly that the corporate separation was a fiction—the protections that the SPV structure was designed to provide could collapse entirely. In the Enron litigation, the bankruptcy examiner concluded that certain of Enron’s SPVs were shams because Enron controlled both sides of the transactions and the SPVs lacked genuine economic substance.[83] While the major AI data center SPVs involve arm’s-length financial sponsors, the parent companies’ operational control and residual financial exposure through guarantees and equity stakes may provide a basis for alter ego claims. In a distress scenario, plaintiffs will test whether the SPV boundary is real, whether independent governance was meaningful, and whether the parent’s operational dominance rendered the structure vulnerable to veil-piercing, substantive consolidation, or alter ego theories.
E. Valuation, Margin Call, and Collateral Litigation
This category of risk is, at bottom, a valuation fight. The collateral securing AI data center debt is vulnerable to rapid deterioration, and the disputes that follow will be among the most contentious in the sector. Once lenders, borrowers and investors disagree about what the collateral is actually worth, litigation becomes almost inevitable.
At the GPU level, the fundamental problem is that no one agrees on what GPUs are actually worth. Companies like CoreWeave depreciate their GPUs over six years; project development lawyers and engineers generally estimate useful lives of three to four years; and some analysts believe the actual useful life may be as short as two to three years.[84] The problem is compounded by NVIDIA’s and AMD’s shift from a two-year to a one-year chip release cycle: each new generation drives down the market value of existing models. Rental rates for NVIDIA’s H100 GPUs have declined by roughly 70–90% since 2023.[85] As one commentator has observed, using a GPU-backed loan to purchase new GPUs is “akin to using one’s credit card to pay off the debt on another credit card.”[86] That mismatch between accounting life, engineering life, and market life means that the same GPU portfolio may appear solvent on one balance sheet and distressed on another.
At the data center level, AI-focused facilities are purpose-built with specialized cooling systems, high-density power configurations, and custom layouts designed for GPU server racks. Converting an AI data center to general-purpose cloud computing or other uses is expensive and complex.[87] If demand for AI computing contracts, these facilities may function as stranded assets with limited alternative use and depressed liquidation value.
When collateral values decline below the thresholds specified in loan agreements, a cascade of enforcement actions follows. Lenders will issue margin calls and demand that borrowers post additional collateral or pay down the loan balance. If borrowers cannot comply, lenders will declare covenant defaults, accelerate the outstanding debt, and move to foreclose—through mortgage foreclosure for the data center real estate, and through UCC Article 9 disposition proceedings for the GPUs and equipment.[88] Borrowers will contest whether the lender’s valuation of the collateral was accurate, whether the lender waived the covenant breach by not acting sooner, and whether the lender’s disposition of seized collateral was commercially reasonable.[89] Investors in securitized pools of data center debt will bring claims analogous to those in the RMBS litigation: that the originator or securitizer misrepresented the quality and value of the underlying collateral. In other words, the litigation will not be limited to pure collection actions; it will also encompass appraisal fights, waiver arguments, sale-process challenges, and misrepresentation claims about the collateral’s original underwriting.
The interaction between the two collateral layers creates a cascading risk. If GPU collateral loses value, the neocloud tenant cannot service its debt, which impairs the tenant’s ability to make lease payments to the data center SPV, which in turn undermines the cash flows backing the data center ABS that investors are relying on. The chain of exposure runs from GPU depreciation through tenant default to data center SPV distress to ABS impairment—a sequence that, if it materializes at scale, would propagate losses across every layer of the financing stack.
F. Construction, Power, and Infrastructure Contract Litigation
Construction, power, and interconnection disputes are likely to be among the earliest and most disruptive forms of litigation in the AI data center sector because financing triggers, lease commencements, and credit support mechanisms are all keyed to milestone completion dates. Data center projects face formidable construction and delivery risks that translate directly into litigation exposure. Bain & Company has warned that projects are being delayed by equipment lead times stretching from eight to twenty-four months and by acute skilled-labor shortages.[90] Electrical utility connection delays—which can stretch to five years—present perhaps the most formidable obstacle to delivering data centers by their ready-for-service dates.[91]
When developers miss delivery timelines, the consequences cascade through the contractual chain. Developers face liquidated damages claims from tenants whose lease commencement dates are pushed back. Tenants face liability to their own customers for failing to deliver promised computing capacity. GPUs purchased in advance sit idle and depreciate. Credit enhancements structured around construction completion dates—like Google’s backstop—may never trigger. The resulting disputes will not be confined to a single contract. Developers will pursue contractors, utilities, and equipment suppliers; tenants will pursue developers; lenders and noteholders will examine whether missed milestones trigger defaults or defeat the support mechanisms on which the financing depended.
Force majeure defenses will be litigated extensively, with disputes over whether supply chain disruptions, labor shortages, or utility delays qualify as unforeseeable events excusing performance. The allocation of utility interconnection risk—whether the developer or the tenant bears the cost and delay of connecting to the grid—will generate its own category of disputes.
The CoreWeave securities class action already illustrates this dynamic: the complaint alleges that weather-related delays pushed back the completion date of a Denton, Texas data center cluster intended for OpenAI by several months, that Core Scientific (CoreWeave’s building partner) began flagging these delays nine months before CoreWeave disclosed lowered revenue guidance, and that the scope of the delays was greater than the company acknowledged.[92]
G. Investment Treaty and Domestic Foreign Investment Law Arbitrations
As AI data-center investment globalizes, entirely new class of risk emerges globally arising out of host-state interference. Beyond the United States, major cloud providers are deploying tens of billions of dollars into new facilities, including Amazon’s multibillion-euro expansion of data centers in Spain and large AI campuses emerging in Portugal and the Nordic region, where abundant renewable energy and favorable cooling conditions attract hyperscale infrastructure. In the Middle East, sovereign-backed initiatives are driving large AI infrastructure builds, with Microsoft, Google, Oracle, and Amazon committing billions of dollars to cloud and AI regions in the United Arab Emirates and Saudi Arabia as part of broader national “sovereign AI” strategies. AI data center capacity in the Asia-Pacific region is projected to more than double by 2030,[93] and Cushman & Wakefield forecasts growth of 60% by 2030 for Latin America.[94] As the buildout globalizes, so does exposure to political risk, regulatory volatility, and state action directed at foreign-owned digital infrastructure.
As these AI data center projects depend on state-created or state-controlled inputs—licenses, permits, land rights, grid access, connectivity, data-residency infrastructure, and capital transfer rules—foreign investors face a distinct layer of political and regulatory risk. In that environment, treaty protection and domestic investment law protection should be mapped before capital is committed. UNCTAD currently lists 2,242 bilateral investment treaties (“BITs”) and 417 treaties with investment provisions in force, and UNCTAD also reports that at least 108 countries maintain investment laws as core instruments governing foreign investment, with more than half of those laws providing access to international arbitration.
For data-center sponsors, covered investments can include shares, debt, contractual rights, real property, leases, intellectual property, and rights conferred by law such as licenses and permits; it also explains that ownership or control may be direct or indirect. In the data-center context, that means protection may attach not only to equity in the project company, but also to site-control rights, concession rights, development agreements, and key government-issued approvals—if they are housed in the right entity before a dispute becomes foreseeable.
Accordingly, investors in foreign data-center projects should structure with protection in mind: chose the investment vehicle early, confirm treaty nationality before any dispute is on the horizon, ensure that the investing entity holds the critical permits and contractual rights, preserve evidence of state representations and reliance, and review treat-specific exceptions, carve-outs, and denial-of-benefits issues on a treaty-by-treaty basis. Domestic foreign investment statutes should be analyzed in parallel, because they can provide a second protection track where treaty coverage is unavailable, narrower than expected, or burdened by jurisdictional limits.
In practice, the state actions that should trigger immediate treaty or statutory review in a data-center project are not limited to outright seizure. They include abrupt permit reversals, discriminatory limits on grid access or expansion, localization or data-residency measures targeting foreign operators, restrictions on capital transfers, and nationality-based rules that favor local companies or compel liquidation upon insolvency or limits on restructuring options. Those measures can map onto familiar protections such as fair and equitable treatment, national treatment, most favored-nation treatment, expropriation, transfer protections, and prohibitions on certain performance requirements.
The case law is useful not because data centers are identical to toll roads or airlines, but because it shows that treaty protections can reach indirect and operationally complex investments facing heavy state interference. For example, in LARAH v. Uruguay, private equity firm Tenor Capital alleged that the Uruguayan government hindered its portfolio company LARAH’s access to financing, interfered with its management, and organized a public campaign to discredit its managers.[95] Tenor filed arbitration under the Panama-Uruguay Bilateral Investment Treaty (“Panama-Uruguay BIT”).[96] The tribunal agreed with Tenor and dismissed all of Uruguay’s jurisdictional objections, finding that the investor’s exercise of management activities was sufficient to trigger Uruguay’s treaty obligations, even where it only held indirect ownership.[97]
Other high-profile cases are ongoing. Last year, Canadian investor Brookfield launched a $2.7 billion arbitration against Peru.[98] That suit alleges Peru breached the Canada-Peru Free Trade Agreement by illegally interfering with the collection of tolls by Rutas de Lima, a toll road operator Brookfield controls.[99]
The practical lesson is that once political or regulatory interference begins to impair the economic use of the investment, international investment law may provide a path to neutral adjudication and damages, while domestic investment statutes may provide parallel leverage.
H. AI Compute Take or Pay Contract Litigation
Many AI infrastructure projects depend on long-term commitments from customers to purchase minimum levels of computing capacity, cloud services, or hosting. Those arrangements function much like take-or-pay contracts as they create revenue certainty needed to support project finance, SPV issuance, and large-scale debt underwriting. For example, OpenAI has committed to purchase $22 billion in computing capacity from CoreWeave, alongside $300 billion from Oracle and $38 billion from Amazon.[100]
The litigation risk is that these commitments become unstable precisely when market conditions change. Infrastructure providers, in turn, are likely to respond with breach-of-contract claims seeking to enforce minimum payment obligations, reservation commitments, exclusivity provisions, termination fees, or volume-ramp promises. Customers will respond with the defenses that usually arise when a minimum-commitment agreement becomes painful: service-level failures, delayed delivery, failure to achieve ready-for-service status, technological obsolescence, force majeure, failure of condition precedents, and sometimes frustration or commercial impracticability theories.
These disputes are likely to be among the earliest and most economically significant in the sector because the deals are highly concentrated. The economics of many projects depend on a very small number of counterparties, which means that the loss or downscaling of a single anchor customer can destabilize the whole financing structure. A customer’s refusal to take or pay for committed capacity will not remain a bilateral contract problem for long; it will quickly become a lender, SPV, and disclosure problem as the expected revenue stream begins to unravel.
I. Environmental and Community Litigation
Another major litigation risk arises from the immense energy and water demands of AI data centers, which increasingly expose operators to environmental, public health, and civil rights challenges from host communities. These disputes are no longer hypothetical. Civil rights and environmental groups are already using federal environmental regulations to challenge data center operations. For example, the National Association for the Advancement of Colored People (NAACP) recently filed a notice of intent to sue xAI, Elon Musk’s AI company, under the Clean Air Act for allegedly operating unpermitted methane-powered generators at its Southaven, Mississippi data center.[101] According to the group, the generators emit pollutants such as nitrogen oxides and particulate matter containing chemicals like formaldehyde. These substances are associated with asthma, respiratory disease, and cancer, and could make the facility one of the largest industrial emitters of nitrogen oxides in the region.[102] Mississippi regulators contend that the generators do not require permits, while the Environmental Protection Agency (EPA) maintains that they do, setting up a regulatory dispute over permitting requirements.[103] If the generators ultimately produce pollution linked to adverse health outcomes in the surrounding, largely African-American community, the circumstances could readily give rise to future class action litigation.
A related source of litigation risk stems from the significant water demands associated with cooling large-scale data centers, particularly in drought-prone regions. For example, in 2021 the Mesa, Arizona City Council approved development of an $800 million data center projected to consume up to 1.25 million gallons of water per day to cool its servers.[104] The project drew sharp criticism from local officials concerned about water scarcity in the region. Mesa’s vice mayor warned that “this has been the driest 12 months in 126 years,” adding that “we are on red alert, and I think data centers are an irresponsible use of our water.”[105] In South Carolina, environmental groups challenged Google’s request to draw millions of gallons of groundwater daily from a depleted aquifer to cool its data center in Goose Creek, ultimately resulting in a negotiated settlement restricting the company’s groundwater use.[106] As the data center buildout continues, communities affected by their energy and water demands may increasingly mobilize politically, prompting new regulatory restrictions on siting and resource use, as well as litigation by residents and local groups challenging the environmental and public-health impacts of these facilities.
These litigation risks are expected to increase as federal, state, and local governments pursue new legislation and regulations restricting data center development. As of early 2026, moratorium bills targeting data center construction have been introduced in at least eleven state legislatures, while over fifty local moratoriums have been enacted across the United States.[107] In March 2026, responding to bipartisan concerns about electricity price increases tied to data center energy consumption, President Trump brokered the “Ratepayer Protection Pledge,” under which Amazon, Google, Meta, Microsoft, Oracle, OpenAI, and xAI committed to build or buy their own power generation for new data centers and to cover the cost of all associated grid infrastructure upgrades.[108] The administration acknowledged that the pledge merely formalized measures the companies were already beginning to adopt, and conceded that enforcement would depend on state utility regulators rather than the federal government.[109] Still, the pledge reflects the political potency of the issue—Democrats flipped two seats on Georgia’s utility commission in 2025 by campaigning on electricity costs[110]—and may foreshadow binding regulatory requirements that could impose additional costs, permitting delays, and compliance obligations on data center developers and their financing counterparties.
VII. Conclusion
The debt financing of the AI data center buildout presents a litigation landscape of unusual breadth and complexity. The deeply interconnected ecosystem means that distress at any single node can propagate across multiple counterparties. The two major sources of capital carry distinct vulnerabilities: corporate bonds have climbed to historic highs, degrading credit profiles and resulting in bondholder lawsuits, while private credit’s opacity means that losses may not become visible until they have already spread to the banks, insurers, and pension funds that provide upstream financing. The structures through which that capital is deployed—off-balance-sheet SPVs, conditional credit enhancements, and layered securitizations—compound these risks by separating the entities that bear the risk from those that assess it. And the collateral securing the debt—purpose-built data centers with limited alternative use and GPUs whose value erodes on a timeline that may outpace borrowers’ ability to service their debt—introduces valuation disputes with direct parallels to the pre-crisis RMBS market.
The Oracle bondholder lawsuit and the CoreWeave securities class action are the first shots in what is likely to be a sustained wave of litigation. Whether or not the AI boom proves to be a bubble, for litigators, courts, and regulators, the buildout of AI data centers is not merely a story about technological progress. It is the beginning of a new cycle of disputes at the intersection of infrastructure, finance, energy, and technology.
***
For more information, please contact:
Rajat Rana, Partner
Email: rajatrana@quinnemanuel.com
Phone: 212-849-7520
Alec Bahramipour, Associate
Email: alecbahramipour@quinnemanuel.com
Phone: 212-849-7256
To view more memoranda, please visit www.quinnemanuel.com/the-firm/publications/
To update information or unsubscribe, please email updates@quinnemanuel.com
[1] See, e.g., Joshua You & David Owen, Scaling Intelligence: The Exponential Growth of AI’s Power Needs, at 7 (Elec. Power Rsch. Inst. & Epoch AI, White Paper No. 3002033669, Aug. 2025), https://www.epri.com/research/products/000000003002033669.
[2] McKinsey & Company, The Cost of Compute: A $7 Trillion Race to Scale Data Centers (Apr. 28, 2025), https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers. To put that figure in perspective, the more relevant comparison is not traditional public infrastructure projects but the cumulative cost of building the modern internet itself. Over the past three decades, hundreds of billions of dollars were invested in data centers, fiber-optic networks, power infrastructure, and cloud computing platforms operated by companies such as Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle. That investment financed the physical backbone of the digital economy—the servers, networking equipment, electrical systems, and facilities that power the global internet. The current wave of AI infrastructure investment represents a comparable technological transition: the construction of a new layer of computational infrastructure built around GPU clusters, high-bandwidth networking, and enormous energy capacity. In effect, the AI data center buildout may constitute the creation of a new generation of digital infrastructure—what some observers have described as “Internet 2.0”—designed specifically to support large-scale artificial intelligence systems
[3] Laura Bratton, Big Tech Set to Spend $650 Billion in 2026 as AI Investments Soar, Yahoo Fin. (Feb. 6, 2026), https://finance.yahoo.com/news/big-tech-set-to-spend-650-billion-in-2026-as-ai-investments-soar-163907630.html; Jennifer Elias, Tech AI Spending May Approach $700 Billion This Year, But the Blow to Cash Raises Red Flags, CNBC (Feb. 6, 2026) [hereinafter CNBC AI Spending], https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html; Rolfe Winkler, Nate Rattner & Sebastian Herrera, Big Tech's $400 Billion AI Spending Spree Just Got Wall Street's Blessing, Wall St. J. (July 31, 2025), https://www.wsj.com/tech/ai/tech-ai-spending-company-valuations-7b92104b.
[4] CNBC AI Spending, supra note 3.
[5] Id.
[6] See Iñaki Aldasoro et al., Financing the AI Boom: From Cash Flows to Debt, BIS Bulletin No. 120, at 1 (Jan. 7, 2026), https://www.bis.org/publ/bisbull120.htm.
[7] See The $3 Trillion AI Data Center Build-Out Becomes All-Consuming for Debt Markets, Bloomberg (Feb. 3, 2026), https://www.bloomberg.com/news/articles/2026-02-02/the-3-trillion-ai-data-center-build-out-spurs-a-debt-market-boom.
[8] See Paula Seligson et al., How AI Companies Are Keeping Debt Off Their Balance Sheets, Bloomberg (Oct. 31, 2025) [hereinafter Bloomberg Off-Balance Sheet], https://www.bloomberg.com/news/articles/2025-10-31/meta-xai-starting-trend-for-billions-in-off-balance-sheet-debt.
[9] Letter from Senators Elizabeth Warren, Richard Blumenthal, Chris Van Hollen & Tina Smith to Treasury Sec’y Scott Bessent, at 1 (Jan. 22, 2026) [hereinafter Warren Letter], https://www.banking.senate.gov/imo/media/doc/20260122_final_letter_to_secretary_bessent_re_ai_debtpdf.pdf.
[10] Id.
[11] See Speech by Federal Reserve Vice Chair Philip N. Jefferson at the 2025 Federal Reserve Bank of Cleveland Financial Stability Conference (Nov. 21, 2025); Aldasoro et al., supra note 6, https://www.federalreserve.gov/newsevents/speech/jefferson20251121a.htm.
[12] See Charles River Assocs., An Overview of Data Center Financing and Asset-Backed Securities, at 1 (Nov. 2025) [hereinafter CRA ABS Report], https://www.crai.com/insights-events/publications/an-overview-of-data-center-financing-and-asset-backed-securities/.
[13] See EPRI, Scaling Intelligence, supra note 1, at 25–28; see also Colin Smith & Lori Bird, Powering the US Data Center Boom, World Res. Inst. (Dec. 18, 2025), https://www.wri.org/insights/powering-us-data-center-boom.
[14] CRA ABS Report, supra note 12, at 2.
[15] Id., at 1–2.
[16] Ohio Carpenters’ Pension Plan v. Oracle Corp., Index No. 150612/2026 (N.Y. Sup. Ct. Jan. 14, 2026); see also Jonathan Stempel, Oracle Sued by Bondholders Over Losses Tied to AI Buildout, Reuters (Jan. 14, 2026), https://www.reuters.com/legal/oracle-sued-by-bondholders-over-losses-tied-ai-buildout-2026-01-14/.
[17] Tabby Kinder, Tech Groups Shift $120bn of AI Data Centre Debt Off Balance Sheets, Fin. Times (Dec. 24, 2025) [hereinafter FT $120 Billion], https://www.ft.com/content/0ae9d6cd-6b94-4e22-a559-f047734bef83.
[18] Id.; see also Paula Seligson et al., Banks Lend $18 Billion for Oracle-Tied Data Center in New Mexico, Bloomberg (Nov. 7, 2025), https://www.bloomberg.com/news/articles/2025-11-07/banks-lend-18-billion-for-oracle-tied-data-center-in-new-mexico.
[19] Bloomberg Off-Balance Sheet, supra note 8; see also Juliana Liu, Meta Is Building a Massive Data Center. Why It’s Fueling Fears of an AI Bubble, NPR (Dec. 5, 2025), https://www.npr.org/2025/12/05/meta-data-center-ai-bubble.
[20] FT $120 Billion, supra note 17.
[21] Moody’s Ratings, AI Data Centre Leases May Not Tell the Full Story (Feb. 2026) [hereinafter Moody’s AI Leases]; Meta Platforms, Inc., Annual Report (Form 10-K) (2025), https://www.moodys.com/research/ai-data-centre-leases-may-not-tell-the-full-story.
[22] Davide Scigliuzzo & Gowri Gurumurthy, Meta Sells $30 Billion of Bonds as AI Frenzy Fuels Record Orders, Bloomberg (Oct. 30, 2025), https://www.bloomberg.com/news/articles/2025-10-30/meta-platforms-offers-six-part-bond-amid-ai-spending-rush.
[23] Aaron Weinman & Gowri Gurumurthy, ‘Google Backstop’ Adds New Twist to Data Center Financing Frenzy, Bloomberg (Oct. 16, 2025) [hereinafter Google Backstop], https://www.bloomberg.com/news/articles/2025-10-16/google-backstop-adds-new-twist-to-data-center-financing-frenzy.
[24] Id.
[25] TeraWulf Inc., Exhibit 10.1 to Form 8-K (filed Aug. 14, 2025), https://www.sec.gov/Archives/edgar/data/1083301/000110465925078084/tm2523008d2_ex10-1.htm.
[26] Google Backstop, supra note 23; see also Press Release, Cipher Mining Inc., Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack (Sept. 25, 2025), https://www.ciphermining.com/news/cipher-mining-signs-168-mw-10-year-ai-hosting-agreement-with-fluidstack.
[27] Press Release, CoreWeave, CoreWeave Secures $7.5 Billion Debt Financing Facility Led by Blackstone and Magnetar (May 17, 2024), https://www.coreweave.com/blog/coreweave-secures-7-5-billion-debt-financing.
[28] See Mellon Invs. Corp., Record-Breaking AI-Related Debt Issuance in 2025 (Dec. 2025); Christian Hantel, AI Debt Issuance Is ‘Transforming’ the Corporate Bond Market, Portfolio Adviser (Dec. 16, 2025); Data Center Deals Hit Record Amid AI Funding Concerns, CNBC (Dec. 19, 2025), https://portfolio-adviser.com/ai-debt-issuance-transforming-corporate-bond-market/.
[29] Bloomberg Off-Balance Sheet, supra note 8.
[30] See Aldasoro et al., supra note 6, at 4; Warren Letter, supra note 9, at 2; FT $120 Billion, supra note 17.
[31] Greg Cohen, Cooper Killen & Simon Lau, Tail Risk for Banks Posed by Investments in Generative Artificial Intelligence, Fed. Rsrv. Bank of Chi., Chicago Fed Insights (Feb. 2026) [hereinafter Chicago Fed AI Risk], https://www.chicagofed.org/publications/chicago-fed-insights/2026/tail-risk-banks-generative-ai.
[32] Federal Reserve Board, Bank Lending to Private Credit: Size, Characteristics, and Financial Stability Implications, FEDS Notes (May 23, 2025), https://www.federalreserve.gov/econres/notes/feds-notes/bank-lending-to-private-credit-size-characteristics-and-financial-stability-implications-20250523.html; see also Derek Thompson, Something Ominous Is Happening in the AI Economy, The Atlantic (Dec. 12, 2025) [hereinafter Atlantic AI Economy], https://www.theatlantic.com/economy/2025/12/nvidia-ai-financing-deals/685197/.
[33] Warren Letter, supra note 9, at 5; Exec. Order No. 14330, Democratizing Access to Alternative Assets for 401(k) Investors (Aug. 7, 2025), https://www.whitehouse.gov/presidential-actions/executive-order-on-retirement-security/.
[34] See Advait Arun, Bubble or Nothing: Data Center Project Finance, Ctr. for Pub. Enter., at 11 (Nov. 2025) [hereinafter Bubble or Nothing], https://www.publicenterprise.org/reports/bubble-or-nothing.
[35] See The $3 Trillion AI Data Center Build-Out Becomes All-Consuming for Debt Markets, Bloomberg (Feb. 3, 2026), https://www.bloomberg.com/news/articles/2026-02-03/ai-data-center-build-out-all-consuming-for-debt-markets.
[36] Bubble or Nothing, supra note 34, at 38; Seligson et al., supra note 18.
[37] Bubble or Nothing, supra note 34, at 11–12.
[38] FT $120 Billion, supra note 17.
[39] Bloomberg Off-Balance Sheet, supra note 8.
[40] CRA ABS Report, supra note 12, at 3–5.
[41] Bubble or Nothing, supra note 34, at 10.
[42] S&P Glob. Ratings, supra note 2, at 2–3.
[43] FT $120 Billion, supra note 17.
[44] Bubble or Nothing, supra note 34, at 51.
[45] See Jacob Robbins & Michael Bodley, As Venture Debt Gambles on GPUs, Not All Are Sold on Silicon-Backed Loans, PitchBook (Aug. 27, 2025), https://pitchbook.com/news/articles/venture-debt-gpu-collateral-loans.
[46] Atlantic AI Economy, supra note 32.
[47] Id.
[48] Tomasz Tunguz, Circular Financing: Does Nvidia’s $110B Bet Echo the Telecom Bubble? (Oct. 3, 2025), https://tomtunguz.com/circular-financing; see also David Dayen, The AI Bubble Is Bigger Than You Think, The American Prospect (Nov. 19, 2025), https://prospect.org/2025/11/19/ai-bubble-bigger-than-you-think/.
[49] Warren Letter, supra note 9, at 4.
[50] Ohio Carpenters’ Pension Plan v. Oracle Corp., Index No. 150612/2026 (N.Y. Sup. Ct. Jan. 14, 2026); see also Jonathan Stempel, Oracle Sued by Bondholders Over Losses Tied to AI Buildout, Reuters (Jan. 14, 2026), https://www.reuters.com/legal/oracle-sued-by-bondholders-over-losses-tied-ai-buildout-2026-01-14/.
[51] Masaitis v. CoreWeave, Inc., No. 26-cv-00355 (D.N.J. Jan. 12, 2026); see also Press Release, Kessler Topaz Meltzer & Check, LLP, CoreWeave, Inc. (CRWV) Class Action Lawsuit Seeks Recovery for Investors; March 13, 2026, Deadline (Feb. 23, 2026), https://www.ktmc.com/new-cases/coreweave-inc.
[52] See In re Delphia Inc., SEC Administrative Proceeding File No. 3-21894 (Mar. 18, 2024); In re Global Predictions Inc., SEC Administrative Proceeding File No. 3-21895 (Mar. 18, 2024), https://www.sec.gov/newsroom/press-releases/2024-36.
[53] Atlantic AI Economy, supra note 32 (reporting that AI companies brought in approximately $60 billion in revenue in 2025 against roughly $400 billion in capital expenditure).
[54] S&P Global, Oracle Inc. ‘BBB’ Ratings Affirmed; Outlook Negative; New Debt Rated ‘BBB’ (Sept. 24, 2025), https://www.spglobal.com/ratings/en/research/articles/250924-oracle-inc-bbb-ratings-affirmed-outlook-negative; Moody’s Flags Risk in Oracle’s $300 Billion of Recently Signed AI Contracts, Reuters (Sept. 17, 2025), https://www.reuters.com/business/moodys-flags-risk-oracles-300-billion-recently-signed-ai-contracts-2025-09-17/.
[55] Oracle Corp., Quarterly Report (Form 10-Q) (Dec. 11, 2025), https://www.sec.gov/Archives/edgar/data/1341439/000119312525315925/orcl-20251130.htm; see also David Rovella, Oracle Runs Into More Trouble as Bonds Look Like Junk, Bloomberg (Dec. 12, 2025), https://www.bloomberg.com/news/articles/2025-12-12/oracle-runs-into-more-trouble-as-bonds-look-like-junk.
[56] Barbara Reinhard, The Connection Between Oracle’s Credit Default Swaps and AI, Voya Investment Management (Dec. 18, 2025), https://www.voya.com/insights/connection-between-oracles-credit-default-swaps-and-ai.
[57] See Joe Wilkins, Trump’s Huge AI Project Is Running Into a Major Financial Problem, Futurism (Jan. 24, 2026), https://futurism.com/trumps-huge-ai-project-financial-problem.
[58] See generally U.C.C. §§ 9-601, 9-609, 9-610 (2010) (setting forth secured party’s rights after default, including repossession and commercially reasonable disposition of collateral).
[59] U.S. Bank Tr. Co., N.A. v. Jericho Plaza Portfolio LLC, No. 1:24-cv-00917, Order Appointing Receiver (S.D.N.Y. Feb. 12, 2024) (appointing receiver over 665,592-square-foot office property securing $149.18 million CMBS loan held in the Natixis Commercial Mortgage Securities Trust 2022-JERI).
[60] See Fed. Hous. Fin. Agency v. UBS Americas, Inc., 858 F. Supp. 2d 306 (S.D.N.Y. 2012); U.S. Bank, Nat’l Ass’n v. UBS Real Est. Sec. Inc., 205 F. Supp. 3d 386 (S.D.N.Y. 2016); Press Release, U.S. Dep’t of Justice, RMBS Working Group Recovers Over $36 Billion (2016).
[61] See, e.g., Bank of N.Y. v. BearingPoint, Inc., 13 Misc. 3d 1209(A), 824 N.Y.S.2d 752 (Sup. Ct. 2006) (analyzing cross-default provisions and notice requirements in credit agreement).
[62] See 11 U.S.C. § 548 (2018) (fraudulent transfers); 11 U.S.C. § 547(b) (2018) (preferences); 11 U.S.C. § 510(c) (2018) (equitable subordination); see also In re Zohar III, Corp., 639 B.R. 73, 97 (Bankr. D. Del.), aff'd, 620 F. Supp. 3d 147 (D. Del. 2022) (describing the conditions for equitable subordination).
[63] FT $120 Billion, supra note 17.
[64] Moody’s AI Leases, supra note 21.
[65] Meta Platforms, Inc., Annual Report (Form 10-K) (2025) (“As of December 31, 2025, RVG payments are not probable and therefore, no liability has been recorded.”), https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm.
[66] Ohio Carpenters’ Pension Plan v. Oracle Corp., Index No. 150612/2026 (N.Y. Sup. Ct. Jan. 14, 2026), https://www.reuters.com/legal/oracle-sued-by-bondholders-over-losses-tied-ai-buildout-2026-01-14/.
[67] Jared A. Ellias & Elisabeth de Fontenay, The Credit Markets Go Dark, 134 Yale L.J. 779 (2025), https://www.yalelawjournal.org/article/the-credit-markets-go-dark.
[68] Cf. cases cited in Ellias & de Fontenay, supra note 67, at 810–15 (discussing fiduciary duty claims against CLO and CDO managers).
[69] Exec. Order No. 14330, Democratizing Access to Alternative Assets for 401(k) Investors (Aug. 7, 2025), https://www.whitehouse.gov/presidential-actions/executive-order-on-retirement-security/.
[70] Atlantic AI Economy, supra note 32 (quoting Natasha Sarin, Professor of Law, Yale Law School).
[71] Michelle Chan, Data centres seek credit ratings to unlock billions in funding for AI push, Financial Times (Feb. 23, 2026), https://www.ft.com/content/e0d9d5f2-c09d-426e-af03-193b488b7b1e.
[72] Id.
[73] Id.
[74] Press Release, Justice Department and State Partners Secure Nearly $864 Million Settlement With Moody’s Arising From Conduct in the Lead up to the Financial Crisis, U.S. Department of Justice (Jan. 13, 2017), https://www.justice.gov/archives/opa/pr/justice-department-and-state-partners-secure-nearly-864-million-settlement-moody-s-arising.
[75] Id.
[76] James Rufus Koren, CalPERS settles with Moody’s for $130 million in ratings case, Los Angeles Times (Mar. 9, 2016), https://www.latimes.com/business/la-fi-calpers-moodys-settlement-20160309-story.html.
[77] Stephen Foley, Moody’s alert cites gap in data centre accounting for Big Tech companies, Financial Times (Feb. 23, 2026), https://www.ft.com/content/3ff9a481-8be3-4dd4-9a00-ea809f3485fd.
[78] Google Backstop, supra note 23; TeraWulf Inc., Exhibit 10.1 to Form 8-K (filed Oct. 14, 2025).
[79] Id.
[80] Google Backstop, supra note 23 (quoting Anubhav Arora, Senior Director, Fitch Ratings).
[81] Moody’s AI Leases, supra note 21; Meta Platforms, Inc., Annual Report (Form 10-K) (2025).
[82] See Moody’s Ratings, AI Data Centre Leases May Not Tell the Full Story (Feb. 2026) (stating that Moody’s would perform its own probability assessments and that “[a] quantitative debt adjustment would likely be made where we believe the reported lease liability understates the likely cash outflow”), https://www.moodys.com/research/ai-data-centre-leases-may-not-tell-the-full-story.
[83] See In re Enron Corp., Final Report of Neal Batson, Court-Appointed Examiner, No. 01-16034 (AJG) (Bankr. S.D.N.Y. Nov. 4, 2003), https://www.concernedshareholders.com/CCS_ENRON_Report.pdf.
[84] The Question Everyone in AI Is Asking: How Long Before a GPU Depreciates?, CNBC (Nov. 14, 2025) [hereinafter CNBC GPU Depreciation] (quoting Latham & Watkins partner Haim Zaltzman: “Is it three years, is it five, or is it seven? It’s a huge difference in terms of how successful it is for financing purposes.”), https://www.cnbc.com/2025/11/14/gpu-depreciation-question-ai.html.
[85] See Bubble or Nothing, supra note 34, fig. 2.11.
[86] Bubble or Nothing, supra note 34, at 27.
[87] Bubble or Nothing, supra note 34, at 51 (noting that “[c]onverting an AI-focused data center into a cloud-focused one, which has more stable revenue streams, is expensive and complex”).
[88] See U.C.C. §§ 9-609, 9-610 (2010).
[89] See U.C.C. § 9-626 (2010) (providing borrower’s remedies where disposition of collateral was not commercially reasonable).
[90] Bain & Company, Next Phase of Data Center Growth to Be More Disciplined, but Risks of Power Constraints and Construction Delays Remain (Oct. 22, 2025), https://www.bain.com/about/media-center/press-releases/20252/next-phase-of-data-center-growth-to-be-more-disciplined-but-risks-of-power-constraints-and-construction-delays-remain-bain--co-research/.
[91] Id.; Schumpeter, Bottleneck-Busters, The Economist (Jan. 17, 2026), https://www.economist.com/business/2026/01/17/bottleneck-busters.
[92] Masaitis v. CoreWeave, Inc., No. 26-cv-00355 (D.N.J. Jan. 12, 2026); see also Press Release, Kaplan Fox & Kilsheimer LLP, Kaplan Fox Alerts CoreWeave (CRWV) Investors (Feb. 23, 2026) (describing allegations that “Core Scientific began flagging these delays nine months before CoreWeave announced lowered revenue guidance in November 2025”), https://www.kaplanfox.com/cases/coreweave-inc-crwv/.
[93] CBRE, Asia Pacific Data Centre Trends & Opportunities (May 2025), https://www.cbre.com/insights/reports/asia-pacific-data-centre-trends-and-opportunities-2025.
[94] Matheus Cardoso, Data Center Expansion in Latin America, Cushman & Wakefield Insights (Dec. 19, 2025), https://www.cushmanwakefield.com/en/insights/data-center-expansion-latin-america.
[95] Toby Fisher, Uruguay Settles Airline Award, Global Arbitration Review (Jan. 31, 2025) https://globalarbitrationreview.com/article/uruguay-settles-airline-award.
[96] Id.
[97] Javier Ferrero Díaz, Uruguay Found Liable in Treaty Dispute with Airline Company Investor, Kluwer Arbitration Blog (July 26, 2024) https://legalblogs.wolterskluwer.com/arbitration-blog/uruguay-found-liable-in-treaty-dispute-with-airline-company-investor.
[98] Marcelo Rochabrun, Brookfield Files US $2.7 Billion Case Against Peru Over Toll Roads, Financial Post (Mar. 13, 2025) https://financialpost.com/news/brookfield-arbitration-peru-toll-roads; Marcelo Rochabrun & Layan Odeh, Brookfield’s Bet on a Road Becomes a $2.7 Billion Headache in Peru, Financial Post (last updated June 25, 2025) https://financialpost.com/fp-finance/brookfields-road-becomes-a-27-billion-headache-peru.
[99] Juan Martinez, Brookfield’s $2.7 Billion Arbitration Case Against Peru: A Deep Dive, The Rio Times (Mar. 14, 2025) https://www.riotimesonline.com/brookfields-2-7-billion-arbitration-case-against-peru-a-deep-dive.
[100] Atlantic AI Economy, supra note 32.
[101] Dara Kerr, Elon Musk’s xAI faces second lawsuit over toxic pollutants from datacenter, The Guardian (Feb. 13, 2026),
https://www.theguardian.com/technology/2026/feb/13/elon-musk-xai-pollution-naacp
[102] Id.
[103] Id.
[104] Olivia Solon, Drought-stricken communities push back against data centers, NBC News (June 19, 2021),
https://www.nbcnews.com/tech/internet/drought-stricken-communities-push-back-against-data-centers-n1271344
[105] Id.
[106] Id.
[107] See Good Jobs First, Data Center Moratorium Bills Are Spreading in 2026 (Feb. 27, 2026), https://goodjobsfirst.org/data-center-moratorium-bills-are-spreading-in-2026/.
[108] David McCabe & Brad Plumer, Trump Announces A.I. Industry Pledge to Pay for Power, N.Y. Times (Mar. 4, 2026), https://www.nytimes.com/2026/03/04/technology/ai-energy-pledge-white-house-trump.html.
[109] Id.
[110] Id.