Distressed AI data centers and bankruptcy

In short

The AI industry spent roughly $400 billion on capital expenditures in 2025 while generating only about $60 billion in revenue. News analysis. To bridge the gap, companies borrowed hundreds of billions of dollars and moved much of that debt off balance sheets using special purpose vehicles. For example, Meta structured approximately $27 billion in debt for an AI data center off its balance sheet through an SPV called Beignet. Fortune. When these borrowers fail, federal bankruptcy law gives creditors and trustees powerful tools. They can reverse transfers made while the debtor was insolvent, recharacterize asset sales as secured loans, combine separate legal entities to share losses, and punish lenders who overreached. Actual distress signals are already flashing. CoreWeave carries $17.5 billion in long term debt and lost more than $600 million in the first six months of 2025. CoreWeave Q2 2025 earnings. Oracle sold $18 billion in bonds and was sued after it raised another $38 billion in loans seven weeks later. Reuters. The first AI data center restructurings have begun.

What is driving AI data center distress?

A wide gap between revenue and spending

AI companies are spending far more on infrastructure than they earn in revenue. In 2025, by one estimate, AI companies collectively will generate about $60 billion in revenue against $400 billion in spending. The Atlantic. That is a capex to revenue ratio of roughly 6.7 to 1. J.P. Morgan estimates that AI would need $650 billion in annual revenue by 2030 just to earn a 10 percent return on the cumulative investment. Tom’s Hardware. McKinsey projects that cumulative AI data center capital spending must reach $5.2 trillion by 2030. McKinsey. The numbers are large. They create a structural mismatch. The largest hyperscalers alone, Alphabet, Amazon, Meta, and Microsoft, accounted for $381 billion in combined 2025 capex. Yahoo Finance. Amazon may generate negative free cash flow of up to $28 billion in 2026. Alphabet and Meta could see free cash flow drop around 90 percent. CNBC. When cash flow retreats, paying down debt gets harder.

Record debt, much of it parked off balance sheets

Companies borrowed heavily to keep building. New borrowing for AI data centers likely exceeded $200 billion in 2025, and that figure understates the total because many deals are private. Law firm analysis. U.S. AI data center secured debt issuance hit $25.4 billion, up 112 percent from the year before. The Economic Times. Tech companies issued a record $108.7 billion in corporate bonds in the fourth quarter of 2025 alone. Washington Post via Finance-Commerce.

A growing share of this debt does not sit on the parent company’s balance sheet. Over about eighteen months, more than $120 billion in AI data center spending was moved off corporate books through special purpose vehicles, commonly called SPVs. Financial Post. An SPV is a separate legal entity created to hold a specific set of assets, often real estate and equipment, and to issue debt against those assets. The parent company’s guarantee may be limited or absent. AI data centers now make up 61 percent of the $79 billion market for digital infrastructure securitizations, across both asset backed securities (ABS) and commercial mortgage backed securities (CMBS). Client alert. That means a large amount of the risk is held by bond investors who may not fully understand what they own.

Regulators have begun to react

In January 2026, four U.S. Senators sent an open letter urging regulators to investigate the AI sector’s heavy reliance on debt. Senate Banking Committee press release. The Federal Reserve and the Bank for International Settlements have expressed similar concerns. The BIS General Manager noted that private credit to the sector is now above $200 billion and growing. LinkedIn analysis. The broader bankruptcy environment is also intensifying. Commercial Chapter 11 filings reached 7,940 in 2025, a small rise, but large company filings above $100 million in assets spiked to 117 in the 12 months through mid-2025, well above the historical average. Epiq AACER, Cornerstone Research. Mega bankruptcies over $1 billion in assets hit 32, up from 24 the prior year. The conditions point to distress in a sector that has never been tested in a downturn.

How AI data center debt is structured, and the risks that creates

SPVs and the off balance sheet puzzle

The central structure in AI data center finance is the special purpose vehicle. A parent company, say, a large cloud operator or a neocloud, transfers assets to an SPV. The SPV borrows money, using the assets as collateral. The parent often agrees to a long term lease or revenue contract with the SPV to support the debt service. The debt stays on the SPV’s books, not the parent’s. For the structure to survive a parent bankruptcy, the transfer of assets must qualify as a true sale under bankruptcy law. If the court sees it as merely a disguised secured loan, the assets can be pulled back into the parent’s estate and used to pay all creditors.

The multi factor test for true sale considers whether the buyer assumes the risk of loss, whether the seller retains control over the assets, and whether the parties intended a sale. No single factor controls. Client alert. No court has yet applied this test to an AI data center SPV, so how it would come out in a real bankruptcy remains untested. In a distress scenario, like a parent bankruptcy, creditors of the parent will argue that the SPV was never truly separate. If they win, the SPV’s assets consolidate back, and the SPV’s lenders become general unsecured creditors of the parent instead of holding a priority interest in a separate pool.

GPU backed debt and the collateral problem

Many AI data center SPVs borrow against the value of graphics processing units (GPUs), not just real estate. CoreWeave, for example, secured a $7.5 billion GPU backed debt facility in 2024. Client alert. When GPU values are high, the collateral covers the loan. When they fall, the lender can issue a margin call demanding extra collateral or repayment. Borrowers that cannot meet the call default, and the lender can foreclose on the GPUs under the Uniform Commercial Code. Because GPU values can plunge quickly, this kind of collateral creates sharp, sudden default risks that differ from slow moving commercial real estate.

Rental prices for NVIDIA H100 GPUs, the mainstay of AI training, have fallen roughly 70 to 90 percent from their 2023 peaks. Client alert. Silicon Data. Used H100 units that fetched $50,000 during the mid 2024 shortage now trade much lower. Silicon Data. This depreciation directly translates into loan to value ratios that can trip covenants in private credit agreements. The exact thresholds are specified in loan agreements, but the cascade mechanic is well understood. Client alert.

The cascade risk

A distress chain can start with a neocloud tenant that cannot service its debt because its GPU collateral has cratered and its customers are paying less for compute. That tenant defaults on lease payments to the AI data center SPV. The SPV’s cash flow drops. The ABS investors holding securitized cash flows from that SPV take a loss. If the chain occurs at scale, each layer of the financing stack gets hit. Client alert. The bankruptcy court will then have to decide which creditors bear the loss, using the tools described in the next section.

The bankruptcy legal toolkit

When an AI data center borrower or its parent files for Chapter 11, a set of key Bankruptcy Code provisions comes into play. The table below gives a quick map.

Each of these tools can reshape who gets paid and how much.

Fraudulent transfers

Under 11 U.S.C. § 548, a bankruptcy trustee can undo a transfer of property or an obligation incurred within two years before the filing if the debtor made it with actual intent to hinder, delay, or defraud creditors, or if the debtor received less than reasonably equivalent value and was insolvent or left with unreasonably small capital (constructive fraud). 11 U.S.C. § 548. Courts infer actual intent from badges of fraud like transferring to an insider, hiding the transfer, or making it while insolvent. A transferee who took in good faith and gave value can keep the transferred interest, but only to the extent of the value given. 11 U.S.C. § 548(c). State fraudulent transfer law, pulled into bankruptcy through § 544(b), often has a longer lookback period of four years. 11 U.S.C. § 544, UVTA § 9.

For an AI data center operator, a trustee might challenge a transfer of GPUs to an affiliate made months before the bankruptcy petition, arguing it was designed to keep valuable assets out of creditors’ reach. The lookback period means that every significant transaction in the two years before a filing gets scrutiny.

True sale versus disguised loan

The question of whether an asset transfer to an SPV is a true sale or a secured loan is one of the most consequential questions in structured finance. If it is a true sale, the assets are owned by the SPV and are not property of the parent’s bankruptcy estate. If it is a loan, the parent’s bankruptcy pulls the assets back, and the SPV’s lenders hold only a secured claim that may be undersecured.

Courts weigh factors such as the degree of recourse to the seller, the seller’s continuing control over the assets, the allocation of risk, and the intent of the parties. No single factor decides the case. Client alert. In an AI data center SPV, a parent that retained too much operational control, or whose lease payments are structured more like debt service, could see the transaction recharacterized. Security opinion letters from law firms will be tested in a real downturn.

Substantive consolidation

A bankruptcy court can, in extraordinary circumstances, disregard the separate legal existence of related entities and pool their assets and liabilities. The Third Circuit’s Owens Corning standard requires a showing that prepetition the entities disregarded separateness so significantly that their creditors relied on the breakdown of entity borders and treated them as one legal entity, or that postpetition their assets and liabilities are so scrambled that separating them is prohibitive and hurts all creditors. Legal analysis. This remedy is rare and a last resort.

In the AI data center context, an overleveraged parent that created dozens of project SPVs could face consolidation arguments if creditors can show they relied on the parent’s credit, not each SPV. A successful motion would eliminate the priority of secured creditors who lent against a specific project.

Equitable subordination

Under 11 U.S.C. § 510(c), a court may subordinate a claim when the claimant engaged in inequitable conduct that injured other creditors or gave the claimant an unfair advantage. 11 U.S.C. § 510(c), In re Mobile Steel Co., 563 F.2d 692 (5th Cir. 1977). A classic trigger is lender control, when a creditor moves from being a passive lender to a de facto controller of the debtor and uses that control to benefit itself at others’ expense. Texas Bankers analysis. In an AI data center bankruptcy, a hyperscaler customer that also provided mezzanine financing and controlled the debtor’s business decisions could see its claim subordinated to those of trade creditors.

Section 363 sales

A debtor can sell assets outside the ordinary course of business with court approval under 11 U.S.C. § 363. The sale can be free and clear of liens, claims, and encumbrances if one of five statutory conditions is met. 11 U.S.C. § 363. Sales to good faith purchasers are protected from reversal on appeal unless a stay is obtained. DailyDAC. The process can be fast, as short as 30 to 90 days. Mayer Brown.

A distressed AI data center operator could use a § 363 sale to offload a partially built facility or a pool of GPUs to a buyer at a steep discount, giving secured creditors whatever the sales price covers and leaving unsecured creditors with little. Because the buyer takes clean title, a § 363 sale can be the fastest way to exit, but it also invites low ball offers in a concentrated market.

Distress signals already visible

Several companies show financial strain and pending litigation. None has yet filed a pure AI data center Chapter 11, but the warning signs are clear.

CoreWeave

CoreWeave is a neocloud provider that rents GPU capacity. It went public in March 2025 at $40 per share, far below its target range. CNBC. By March 31, 2026, long term debt stood at $17.527 billion, up from $14.881 billion at year end 2025 and $4.954 billion a year earlier. Macrotrends. Current liabilities include $3.6 billion payable by June 30, 2026. Total liabilities exceed liquid assets by more than $20 billion. Financial analysis. The weighted average interest rate on its debt is 11 percent, with floating rates between 9 and 15 percent. Fortune. In 2025, the company reported a net loss of $1.17 billion on revenue of $5.13 billion. CoreWeave 10-K. In the first quarter of 2026, it lost another $740 million on $2.078 billion in revenue. CoreWeave IR.

Roughly 62 percent of its 2024 revenue came from a single customer, Microsoft. CNBC. That makes the business brittle. The company also faces a securities class action filed in January 2026 alleging it overstated its ability to meet customer demand. Client alert. One equity analyst publicly predicts that bankruptcy is the likeliest outcome within five years. Fortune. CoreWeave does have a large $99.4 billion revenue backlog, but those contracts depend on continued customer health and will be scrutinized in any restructuring.

TeraWulf

TeraWulf was a bitcoin miner that pivoted to AI and high performance computing. In October 2025, it raised $3.2 billion through high yield bonds rated Ba2/BB and carrying a 7.75 percent coupon. TeraWulf 8-K, Investor presentation. Google backstopped the lease obligations of TeraWulf’s tenant, Fluidstack, and received warrants for approximately 14 percent of TeraWulf. TeraWulf press release, TeraWulf press release. For full year 2025, TeraWulf reported a net loss of $661.4 million on revenue of $168.5 million. Total liabilities were $6.42 billion. TeraWulf 8-K. The company claims $12.8 billion in contracted lease revenue, but the current cash flow is deeply negative.

Meta’s Beignet SPV, a $30 billion walk away risk

In October 2025, Meta completed a $30 billion SPV, Beignet Investor, to finance an AI data center in Louisiana. Blue Owl Capital owns 80 percent, Meta 20 percent. The SPV raised $27 billion in loans from Pimco, BlackRock, Apollo, and others. Meta gave a residual value guarantee of up to $28 billion, disclosed only in footnotes. Client alert. Starting in 2033, Meta can simply pay the guarantee and walk away from the 20 year lease. Goldman Sachs analysts warn that in single asset single borrower structures, losses can be severe when the anchor tenant exits. LinkedIn analysis. A facility built for 100 kilowatt AI racks cannot easily be repurposed for lower power uses. If AI demand stalls and Meta exercises its option, bondholders could face a large haircut.

Oracle sued by bondholders

Oracle sold $18 billion in bonds in September 2025 to fund Stargate AI data center commitments. Seven weeks later, it announced a $38 billion debt package. Bondholders sued, claiming the offering documents falsely said Oracle may need more borrowing when it was already planning the $38 billion raise. Reuters. The lawsuit, filed in New York State Supreme Court in January 2026, names Oracle, executives, and 16 underwriting banks under the Securities Act of 1933. Reuters, IFR. As of January 2026, a $4 billion segment of the bonds traded at 95.75 cents on the dollar. DCD. Oracle’s total debt reached approximately $108 billion. Reuters. This case signals that bondholders will closely scrutinize disclosure around rapid debt accumulation.

Hyperscale Data subsidiary already restructured

A less known but real example is the Chapter 11 filing of Gresham Worldwide, an affiliated defense subsidiary of data center operator Hyperscale Data (NYSE (GPUS)). The reorganization plan was confirmed in August 2025, and the company was on track to emerge on October 1, 2025. PR Newswire. While not a pure play AI data center bankruptcy, it shows how distress in an affiliated entity can pull an AI data center parent into restructuring.

Why AI data centers create unique bankruptcy challenges

GPU values can fall faster than depreciation schedules assume

Many AI data center operators depreciate GPUs over six years. Analysts say the economic life may be as short as two to three years, while engineers estimate three to four years. Client alert. Stated depreciation that overstates asset life can give a misleading impression of collateral coverage. When lenders relying on that collateral discover the true market value has collapsed, they issue margin calls that the borrower cannot meet. Foreclosure on GPUs under the Uniform Commercial Code then transfers the loss to the lender, who must sell the chips into a falling market.

Power contracts put utilities at risk

An AI data center in bankruptcy may try to keep its power supply contract in place because it is essential to running the business. The power supplier may be forced to continue providing electricity even if the debtor’s ability to pay is uncertain. A bankruptcy court, concerned with preserving the estate, may impose a deposit requirement that does not fully protect the supplier. News analysis. The utility could face a liquidity crunch if the debtor fails to pay and the court’s orders expose it to unrecoverable costs.

The market for distressed assets is locked

North American AI data center vacancy rates hit 2.3 percent in mid 2025, with 73 percent of capacity under construction already preleased. JLL. The buyers for these facilities are a small number of hyperscale customers, Microsoft, Google, Amazon, Meta. Those buyers can afford to wait until a distressed seller has no choice but to accept a fire sale price. A survey of over 400 senior AI data center executives found that 70 percent expect M&A to become more attractive, yet two thirds of REIT and construction players anticipate distress ahead. AlixPartners survey. That imbalance creates a buyer’s market in a downturn, pushing recoveries lower.

The telecom bubble as a caution

The late 1990s parallel is direct. Telecom companies invested over $500 billion in fiber, mostly with debt. By 2001, 95 percent of the fiber was unused. Between 2000 and 2002, global telecom stocks lost $2 trillion in value. Bondholders in some cases recovered only about 20 percent. Wikipedia. The AI data center buildout shares many of those features, including massive debt, a concentrated customer base, and assets that may be worth far less than their construction cost if demand falters.

Key takeaways

• Watch the collateral gap. GPU depreciation is fast. Verify that loan agreements use realistic useful lives and that margin call triggers leave headroom.
• Scrutinize SPV separateness. In a parent bankruptcy, true sale is not automatic. Review transaction documents for recourse, control, and risk allocation that could lead to recharacterization.
• Trace pre-bankruptcy transfers. The two year lookback under § 548 and longer state reach back periods mean that every material transfer to insiders will be examined for fraudulent transfer liability.
• Understand your power contract’s risk. A utility selling power to an AI data center should model what happens if the customer files for Chapter 11 and the court orders continued performance with minimal security.
• Buyer concentration magnifies loss. In a distress sale, the only buyers may be the same hyperscalers that hold the leverage. Recovery projections for secured lenders should assume a thin market and steep discounts.

Frequently asked questions

Q:What is an SPV and why does it matter in AI data center bankruptcies?

A:An SPV is a separate legal entity created to hold assets and issue debt. In AI data centers, sponsors use SPVs to keep debt off their balance sheets. In bankruptcy, whether the SPV’s assets stay separate from the parent’s estate depends on whether the transfer was a true sale. If not, the assets can be pulled back.

Q:What is a fraudulent transfer and can it affect AI data center investors?

A:A fraudulent transfer is a transfer made to cheat creditors or for less than fair value while insolvent. A bankruptcy trustee can undo these transfers. In an AI data center case, a transfer of GPUs or cash to an insider shortly before filing could be clawed back.

Q:How can a lender lose its secured position in an AI data center bankruptcy?

A:Through equitable subordination under § 510(c) or through substantive consolidation. If the lender controlled the debtor and engaged in misconduct, its claim can be demoted. If the court consolidates the parent and its SPVs, project level secured lenders may become general creditors of a larger pool.

Q:What is a Section 363 sale and how could it be used?

A:A § 363 sale lets a debtor sell assets free and clear of liens with court approval. A distressed AI data center operator could use it to sell a facility or GPU fleet quickly, often within 30 to 90 days. The buyer gets clean title, but sellers may get below market prices.

Q:Are there any AI data center companies in bankruptcy right now?

A:As of May 2026, no major pure play AI data center operator has filed for Chapter 11. But CoreWeave and TeraWulf carry massive debt loads and operating losses. Hyperscale Data’s affiliate Gresham Worldwide already restructured under Chapter 11. Securities lawsuits are active against CoreWeave and Oracle.

Q:How does the revenue capex gap lead to bankruptcy risk?

A:When a company spends far more on building infrastructure than it earns, it must borrow. If revenue does not catch up, the company cannot service the debt and defaults. In the AI sector, the gap was $60 billion in revenue versus $400 billion in capex in 2025. That ratio creates a high probability that some borrowers will fail.

Q:What happened in the telecom bubble and does it apply?

A:In the late 1990s, telecom companies borrowed heavily to build fiber networks. Demand did not materialize, and many filed for bankruptcy. Bondholders recovered as little as 20 percent. The AI data center buildout shares the high debt, fast asset depreciation, and concentrated customer risk.

Q:How do I protect against these risks if I am a lender or investor?

A:Insist on realistic GPU depreciation assumptions, review the true sale structure of any SPV, monitor margin call triggers, and stress test recoveries under a thin market scenario. Understand your exposure to a fraudulent transfer challenge if the borrower later goes insolvent.

Subscribe to The Compute Law Brief

The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.

Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.