Oracle's rent notice and PJM's curtailment plan test who carries grid risk

Oracle's rent notice lands on $18 billion of data-center debt through the force majeure provision that requires both parties to consent before service can be interrupted. That requirement turns a power interruption from a physical event into a negotiation between the parties to a Big Tech lease, with a debt stack underneath them that was sized on the assumption the lights stayed on.

Rent on a data-center lease is the revenue line that services the debt, so a notice that puts that line in question reaches every lender in the stack. PJM, meanwhile, has asked federal regulators to approve a framework that would let it cut power to new data centers that add no capacity to the grid. The filing and the rent notice have separate origins and neither caused the other, but they price the same exposure: the moment when power stops arriving and someone has to absorb the cost.

The data-center capital stack has been underwritten around the interconnection queue—how long a project waits for a grid connection, and what it does while it waits—and both items push the binding constraint past the queue. A queue delays a project; a curtailment framework gives an operator a standing right to interrupt one, and a right is harder to model than a delay because it is exercised at somebody else's discretion, on a timetable the project does not control.

Begin with the lease, because that is where the cash flow starts, and a force majeure clause that one side can invoke unilaterally is an abatement or termination mechanism, which lenders price like any other credit event—trigger, notice period, cure rights, tenant credit. A clause that requires mutual consent is a different instrument: it makes the interruption of service a matter for agreement, which puts the outcome in the hands of relative leverage—the rent at stake, the tenant's alternatives, the landlord's ability to replace a tenant at a site that already has power. Oracle's rent notice is a public test of where that leverage sits, and our records frame it as a repricing event for the $18 billion of debt behind the lease.

If rent can be withheld or abated while service is interrupted, then debt service coverage is only as firm as the supply behind it, and the underwriting question is no longer only whether there is a substation nearby. It is also what happens to the cash flow if the operator holds the right to say no. Lenders have been pricing grid access for a while, but the newer work is pricing the terms on which the grid can take service away and the lease-consent rights that determine whether rent keeps flowing in the meantime.

Which party holds the risk is a function of the documents: if a mutual-consent clause lets the rent stop without a default, the exposure does not disappear; it moves down the stack to whoever is servicing debt against that rent stream. That is why the clause matters to a credit committee, and why it belongs in the same conversation as the curtailment framework.

Whether the mutual-consent structure becomes standard is unknowable from one notice, and I will not pretend otherwise. If it does, the practical effect is that interrupting service at a data center requires agreement between two parties rather than a force majeure finding by one. That change would surface first in the leases signed by the largest tenants, who have the leverage to demand consent rights, and last in the pricing of the sites those tenants occupy.

PJM's request to federal regulators is what gives that second question an answer. A framework that lets the operator cut power to new data centers that add no capacity to the grid separates projects that contribute to the system from projects that only draw on it. That is a design standard, and it lands in the pro forma as a category of risk no amount of land assembly can retire. It also creates a diligence item the last underwriting cycle did not have: for any given project, whether the load is a net addition to the grid or a net claim on it.

The market is already pricing pieces of this at the land level. GI Partners paid $14.7 million for data-center land in south Phoenix, and the coverage prices the trade against the substation next door and the city's data center determination—a grid connection and an entitlement, both of which have to exist before a shovel moves. Land beside a substation is worth more than land that is not, and land a city has designated for data centers is worth more than land it has not. Neither premium answers whether the power will flow in the quantity the site was bought for, or for how long.

Energization risk has two halves, and the sector has spent most of its attention on one: construction-phase risk, where the question is whether the interconnection arrives on schedule and whether the site can be served at the capacity the tenant has contracted for. That is a schedule-and-cost problem, visible in the queue and modeled in the budget. The second is operating-phase risk: once power is flowing, who holds the right to stop it and on what terms, because a queue position says nothing about that and a curtailment framework is where it gets a name.

A delay has a carrying cost and a schedule; a curtailment right has neither, because the decision belongs to a third party whose obligations run to the system rather than to the project. That is why the same asset can look sound in a development model and fragile in a loan model, and why the two numbers have begun to diverge.

Brookings has put the AI build-out at $10.3 trillion, and the same analysis assumes 227 gigawatts of proposed data-center capacity never gets built. Inside a forecast of that scale, the concession is the substantial part: capacity that has been drawn, sited and proposed, and on this accounting, never energized. For a lender, that figure contains the problem: debt behind a data center is sized against a lease, the lease is sized against a tenant's need for compute, and the compute depends on power that a planning assumption says may not arrive.

The pipeline figure gives the curtailment framework its context. If 227 gigawatts of proposed capacity never energizes, the projects that do are the ones that clear both the queue and whatever test the operator applies, which makes the value of a site increasingly the value of its proven path to power. Land-and-power packages that looked interchangeable at site selection stop being interchangeable at the credit committee.

Blue Owl's exclusive talks over Stack's Asia Pacific portfolio capture the same split from the capital side, setting a $25 billion bid for income against a build-out ask and suggesting a preference for capacity that exists over paying to energize capacity that does not. The spread between those two positions is where the market is currently trying to price the risk that a project never energizes, or energizes and is then curtailed.

None of this describes a capital shortage: CRE debt funds are holding a record $56 billion in dry powder, and multifamily starts fell nearly 22% in August, so the money is available and the construction is not happening at the scale the allocations imply. What has changed is what the money is willing to be paid for. Underwriting a data center used to be a matter of land, power and a tenant; it is now a matter of grid access, the terms on which the grid can take service away, and whether the lease lets the rent stop when it does.

The likely consequence, and this is inference rather than reporting, is a widening spread between data-center assets with uncontested power and assets whose supply depends on a curtailment framework that has not yet been approved. Lenders working through that spread will look for protections outside the real estate: reserves sized to an interruption, covenants keyed to energization milestones, and consent rights that give a lender a say before the tenant and the landlord agree to something the loan documents did not anticipate. None of that is glamorous, and all of it shows up in the pricing.

More consequential is the diligence standard, which now reaches past the site: a data-center acquisition requires answers to questions the last cycle did not ask—what curtailment authority the local grid operator holds, whether the tenant's lease gives the landlord any control over service interruptions, and whether the force majeure provision requires one signature or two. The Phoenix trade is the shape of that, $14.7 million for dirt whose value depends on a substation next door and a city determination, with the price of the dirt only the first place the answer gets capitalized.

The filing now sits with federal regulators, and until they rule, every data-center lease that reaches a credit committee carries an unquantified line: the probability that a grid operator, rather than a tenant, decides when the rent stops. Oracle has put a figure next to that line—$18 billion of debt, repriced around a clause that has to be negotiated rather than assumed. The next lease to cross a lender's desk will carry a force majeure provision, and the question is whether it still needs both signatures.