When Data Centers Lose Power, Guarantees Decide Who Pays

The revealing detail in Oracle’s latest data-center warning is the bill it reportedly cannot escape. On September 24, Reuters reported that Oracle had issued a force-majeure notice over potential power delays at Project Jupiter, a New Mexico campus developed by Blue Owl’s Stack Infrastructure. Such notices seek contractual relief for events beyond a company’s control. Yet Reuters’ source said Oracle remained responsible for debt costs and could not terminate its lease. Oracle said the project remained on schedule for its planned 2028 opening. A warning about electricity, in other words, had become a test of who must keep paying before computing begins. Reuters’ reporting¹ makes that distinction essential. (uk.marketscreener.com¹)

I read this as a warning against two easy conclusions: that an unpowered building automatically means unpaid lenders, and that a technology company’s signature makes the financing safe. The better diagnosis is contractual risk with a concentration problem. Protection must work on the date cash is needed, and the company providing it must survive demands from all the other projects it supports.

The structure explains why. In a typical arrangement described by the Bank for International Settlements², sponsors put equity into a dedicated company that develops the facility and borrows against future lease payments. This is the core project-finance mechanism: repayment depends heavily on a particular asset’s cash flow and contractual support. Banks may supply funding lines, while funds and insurers hold the debt. My inference is that delayed rent first squeezes the project’s cash and its owners’ returns; lenders suffer losses when available resources, enforceable support and collateral fail to cover their claims. (bis.org²)

Applied Digital’s Ellendale financing shows the clock running. Its February 2026 quarterly filing³ disclosed $2.35 billion of secured notes carrying 9.25% interest, with principal payments beginning December 15, 2027, and final maturity on December 15, 2030. Proceeds included debt-service reserves. An interest reserve is money set aside to pay interest before operating cash arrives; when funded by borrowing, it buys time rather than supplying fresh equity. Applied Digital also promised additional funding needed to achieve lease commencement, but explicitly stopped short of general parent-company repayment support. (ir.applieddigital.com³)

Using that initial principal and coupon, I calculate that hypothetical six-, twelve- and twenty-four-month postponements would require approximately $108.7 million, $217.4 million and $434.8 million of interest funding, respectively, assuming unchanged principal. These are sensitivities, not forecasts or lender-loss estimates. Available reserves and collectible support must be deducted; other carrying costs must be added. The actual principal-payment schedule would also matter. The disclosed terms³ illustrate financing mechanics, not an identified power-delay default.

Customer backing can leave precisely this construction-period gap. In its May 8 filing⁴, developer TeraWulf disclosed that Google’s backstop for tenant Fluidstack became effective only when the corresponding lease commenced. Following specified tenant defaults, Google could pay a termination fee instead of assuming the lease. I would therefore value that protection against its activation date and termination payment, rather than treating all future rent as immediately available cash. These were disclosed contractual conditions, not evidence that TeraWulf had suffered a power-delay default. (investors.terawulf.com⁴)

The same filing’s revolving-credit agreement makes the timing problem sharper. It specified $250 million of commitments but capped availability before first lease commencement at $75 million. It also required at least $100 million of defined quarterly liquidity during that period, with a limited right to cure a shortfall through equity issuance. A borrower can therefore confront financing restrictions while still holding cash. My test would measure usable borrowing capacity and compliance with those contractual requirements⁴, not simply count down to a zero bank balance. (investors.terawulf.com⁴)

There is substantial reason to resist a crisis narrative. The International Monetary Fund’s April financial-stability assessment⁵ found that major cloud companies’ strong balance sheets and free cash flow kept associated financial-stability risks contained. That materially narrows my concern: a creditworthy customer paying from an established business can bridge a construction setback. But this was an April assessment, not a September stress test, and the IMF also warned that faster equipment obsolescence could increase financing needs. (imf.org⁵)

I would test that vulnerability jointly with delays. Alongside six-, twelve- and twenty-four-month postponements, an illustrative downside case could assume a 20% decline in replacement-lease prices and a 30% decline in relevant equipment collateral values. Those are deliberately hypothetical shocks. Lower market prices should affect contracted revenue only where cancellation, repricing or replacement leasing allows it; equipment depreciation should hit the entity that actually owns or finances the equipment. The IMF’s obsolescence analysis⁵ supports testing asset life, not applying one indiscriminate haircut to buildings, chips and leases.

Financing stress is already visible in Jupiter, although its cause cannot be reduced to electricity. On September 18, Reuters, citing the Financial Times⁶, reported that roughly $18 billion of project loans were quoted at 89–91 cents on the dollar. Distribution to investors had stalled amid concerns about Oracle’s borrowing and creditworthiness, leaving banks holding more debt than intended. These were quoted prices, not a demonstrated loss on every loan. Still, they show how risk intended for a wider investor base can remain concentrated with arranging banks. (aol.com⁶)

The wider transmission mechanism is credible. The Financial Stability Board’s May report⁷, from the international financial-risk coordinating body, explicitly identified electricity shortages as a potential cause of data-center delays, cancellations and credit losses. It also documented connections among banks, private-credit funds, insurers and securities backed by loans. Banks can lend both to a fund and to companies that fund finances. Several differently labelled investments may therefore depend on the same borrower or guarantor. The report does not measure the full overlap for delayed data centers. (fsb.org⁷)

Nor does interconnectedness automatically produce contagion. The FSB found measured direct bank lending to private-credit funds relatively small and recognized that funds locking up investor capital reduce withdrawal pressure. My concern becomes broader financial stress only when simultaneous funding demands exceed available support and creditor buffers, prompting forced sales or curtailed lending elsewhere. That is a conditional mechanism consistent with the FSB’s findings⁷, not evidence that such spillovers are already occurring.

The evidence establishes real protections, conditional guarantees and financing strain. It does not supply a representative, matched set of utility-confirmed original and revised connection dates, construction spending, revenue starts, remaining reserves and ultimate creditor holdings. I cannot responsibly turn these examples into a sector-wide loss estimate. Regulators should demand that joined-up view and count each guarantor’s resources once across all supported projects. To declare the risk contained without it, we would have to believe that separate contracts make shared sources of repayment independent. I would not make that assumption.