Oracle Signed the Check Before the Power Arrived

The most revealing sentence in infrastructure finance right now isn’t in any earnings call. It’s in Oracle’s Project Jupiter contract: Oracle owes its data center investors even if the site has no electricity.

That is not a quirk. That is the whole story.

Here is what happened. Oracle committed to Project Jupiter — a roughly 2.45GW data center campus in New Mexico, set to come online in 2028 — through a structure that transfers demand risk almost entirely onto Oracle itself. Oracle issued a force majeure notice to Stack Infrastructure, the site developer, which is owned by Blue Owl Capital. The site has hit a series of setbacks. Oracle’s counterparties don’t care. The payment obligation runs regardless. Oracle signed up for that exposure. The question worth asking is why any company with experienced infrastructure lawyers would accept terms like that, and the answer is that the AI buildout has made data center capacity so scarce — and the customer queue so long — that hyperscalers are outbidding each other not just on price but on risk allocation. Oracle wanted the capacity. It took the terms.

Everyone says the AI infrastructure boom is a supplier’s market for GPUs. The opposite is closer to true for power and land: landlords and grid operators are extracting brutal concessions from the cloud companies desperate to build. Nvidia’s chips are the headline, but dirt with juice running to it is the actual bottleneck — and the entities who control the dirt are writing their own contracts now.

This matters beyond Oracle. The Project Jupiter structure is almost certainly not unique. Every hyperscaler racing to satisfy AI demand — Microsoft, Google, Amazon, Meta — is signing capacity agreements in a market where the power delivery timeline is measured in years and the model training deadlines are measured in months. That mismatch has to be allocated somewhere. Oracle just showed us where it lands: on the balance sheet of the company that needs the compute, not the company that owns the ground.

The force majeure clause is the tell. Oracle invoked it to try to delay its payment obligations — but the FT reports Oracle remains on the hook regardless. The clause didn’t provide the relief Oracle sought. In a normal commercial lease, force majeure relieves the tenant from paying when the landlord can’t deliver. Here, it apparently doesn’t. That is a remarkable inversion of standard property logic, and it happened because Oracle was willing to sign a contract that reflected the counterparty’s leverage, not Oracle’s.

The deeper issue is that “data center” has become a misleading label. What Oracle and its peers are actually contracting for is reliable electrical capacity at a specific location, sized for the thermal envelope of GPU clusters, delivered on a timeline set by model training schedules that are themselves set by competitive pressure from Anthropic and OpenAI. The physical building is almost incidental. What has value — what commands these one-sided terms — is a guaranteed megawatt-hour commitment from a utility that can’t actually guarantee it. Utilities are not set up to provide firm power contracts for loads this large on timescales this short. The grid wasn’t built for it. So the risk has to live somewhere, and it has migrated onto Oracle’s books.

None of this shows up in the standard AI investment thesis, which focuses on model quality and token economics. But the infrastructure layer is where the real money is moving — and where the real losses will be booked when force majeure notices start multiplying. The AI capex cycle is not just about how many H100s you own. It’s about how many contractual commitments you’ve made to investors who will get paid whether or not a New Mexico substation comes online.

The market will eventually price this. Data center development companies with take-or-pay contracts backed by investment-grade counterparties are worth more than anyone thought two years ago. Oracle’s willingness to absorb this risk isn’t charity — it’s a bet that AI revenue will eventually dwarf the exposure. That bet might be right. But the force majeure clause is a reminder that “eventually” is doing a lot of heavy lifting, and the grid doesn’t care about your model release schedule.