In New Mexico, the argument over America’s AI expansion has reached the small print of a lease. Oracle issued a force-majeure notice concerning Project Jupiter, citing potential power delays, Reuters reported in September. Oracle said the project remained on schedule; Blue Owl said the notice did not alter its financial commitments.¹

Buildings, processors and power infrastructure must be paid for before computing capacity earns revenue. A delay stretches that interval, leaving the companies involved to decide who keeps paying while the machines wait.

Reuters reported that Oracle was seeking to defer payments if Jupiter missed its planned 2028 opening, rather than leave the project. The underlying lease and notice have not been independently examined for this article.¹ But a project need not be cancelled to strain its financing: receipts arriving a year later may be unavailable for obligations falling due now.

Contracts with a clock

Core Scientific and CoreWeave show how those obligations connect American businesses. They are separate companies: Core Scientific provides data-centre facilities; CoreWeave installs advanced processors and sells specialised cloud computing to customers.

At the end of June, Core Scientific reported 590 megawatts of leased customer power capacity under its CoreWeave arrangements, with 395MW billing and the remaining 195MW in construction and commissioning. That unfinished capacity represented prospective growth, alongside the costs and obligations of completing it.²

An April SEC disclosure set out delivery protections for certain sites awaiting commencement: daily credits of $20,000 for days 31-60 after the target date, rising to $30,000 from day 61. Extensions apply for delays caused by CoreWeave or outside the licensor’s control.

If commencement had not occurred within 120 days of the target, CoreWeave could terminate the relevant order without penalty by giving notice within a specified 15-day window. Accrued credits would then become payable within 30 days. These were disclosed rights, not evidence that a termination had occurred.³

The customer gains protection against waiting indefinitely; the supplier faces the possibility of late revenue, additional credits and a lost order. Core Scientific’s June filing also disclosed that CoreWeave accounted for approximately 77 per cent of its first-half revenue, making the customer’s financial health material to its own.²

When lenders want their money sooner

CoreWeave’s borrowing adds another deadline. In August, it announced a $2.6bn facility with an approximately five-year maturity, supported by customer contracts averaging approximately three years. The company said shorter agreements could command higher prices and capacity could be re-let when initial contracts expired.⁴

That is a credible commercial model. Equipment can serve successive customers, while repayments progressively reduce the outstanding debt. Comparing the two average terms cannot establish an uncovered financing gap.

The executed agreement nevertheless requires action if replacement business falls short. Where an expiring contract is not renewed or replaced on qualifying terms by the contractual deadline, the borrower must repay enough debt to restore specified coverage tests. Payment is due within five business days after that deadline.

The renewal interval is redacted and the amortisation schedule omitted, preventing an outside calculation of debt remaining at individual customer expiries. Cure mechanisms also apply; the agreement does not establish that a repayment has been triggered.⁵

A separate $8.5bn CoreWeave facility provides for model-dependent principal repayment following specified data-centre delays. The amount can be zero, and redacted inputs prevent quantification. It does not make the whole facility payable whenever construction misses a milestone.⁶

Both structures protect lenders by adjusting debt to the business supporting it. For the borrower, that adjustment can mean finding cash sooner when expected earnings are already under pressure.

The route to a spiral

An operator awaiting electricity collects revenue later while committed costs continue. Its customer seeks concessions or another supplier. Where a termination right exists, an order supporting the project’s economics may disappear.

A lender may then require debt reduction under the relevant agreement. Replacement finance could solve the shortfall, but carry higher interest, more collateral or a larger equity contribution. Repeated across the sector, tighter terms could delay further projects and weaken equipment orders.

Operators competing for replacement customers might cut prices. Lower rental income and falling equipment resale values could, in turn, weaken the assumptions supporting further borrowing. Decisions intended to protect individual customers and creditors could collectively deepen the squeeze.

This is a possible transmission mechanism. The separate agreements examined here do not establish that the sequence is occurring, or that one default would activate every provision.

Nor does a project subsidiary always contain the exposure. CoreWeave’s parent guarantee for the $2.6bn facility is broad; the separate $8.5bn facility’s guarantee covers specified conduct and insolvency-related events. The allocation of losses depends on the documents, not the corporate label.^{7, 8}

Useful technology, vulnerable financing

The pessimists must account for functioning investment. Core Scientific reported first-half colocation revenue of $214.2m, against $19.1m a year earlier. Its June filing subsequently disclosed July lease agreements involving AMD and another cloud operator, extending its efforts to diversify.²

Customers can renew, sponsors can contribute equity and lenders can grant extensions. Strong demand may sustain older processors and make a delayed facility profitable. Contractual protections demonstrate that financiers anticipated these risks.

The vulnerability lies in projects whose recovery depends on fresh money repeatedly arriving on favourable terms. Higher financing costs can make a useful operating asset an inadequate investment at the price originally paid.

Distress at specialist operators would not automatically become a banking crisis. That would depend on funded exposures, guarantees, recoveries and creditors’ capacity to absorb losses.

But an AI bubble does not require the technology to fail. It requires enough investments to earn less, or earn later, than their financing can bear. Demand can grow while the companies that financed its infrastructure lose money.

The decisive test will come when a customer deadline and a lender’s repayment demand arrive before the power does.