When we ask a model to generate an answer, the cost seems almost negligible. Behind that instant stand buildings, chips, power lines, cooling systems and financial contracts lasting many years. Oracle’s Project Jupiter, a major AI data centre planned in New Mexico in partnership with OpenAI, makes visible an element frequently absent from technology narratives: debt.
Reports indicate that roughly eighteen billion dollars of financing linked to the project has come under pressure and trades below face value. Banks are struggling to transfer part of the risk while investors and rating agencies watch Oracle’s growing leverage. This does not necessarily mean AI is failing. It means its expansion requires enormous financial bets before future demand can be verified.
A data centre is built on the expectation that models, companies and users will continue purchasing compute. If chips become more efficient, suppliers change or revenue grows too slowly, infrastructure may become oversized. If demand explodes, the first builders control a scarce resource. Debt turns technological uncertainty into concrete economic obligations.
Project Jupiter also reveals the link between finance and territory. Its energy plan involved gas turbines with several gigawatts of capacity, but encountered regulatory obstacles and local opposition over water, air quality and resource use. A delayed environmental permit can change the value of loans; an inadequate grid can become a credit risk. Sustainability is not a chapter separate from the business plan.
The energy scale of AI changes the relationship between technology companies and public utilities. New plants, lines and cooling systems may be financed partly through tariffs, subsidies or guarantees affecting communities. It must be clear who pays for expansion and who retains the benefits. A private project should not automatically transfer the risk of overestimated demand to households and regions.
Concentration adds another vulnerability. Few organisations can raise the capital required for facilities of this size, and only a few clients can absorb their capacity. If one partner reduces orders or encounters problems, the entire financial chain feels the effect. Dependence links banks, funds, utilities, builders, chipmakers and local authorities—not only a cloud provider.
Model efficiency can reduce pressure but creates a paradox. When each operation costs less, total uses often increase. Generative video, permanent agents and personalised applications may consume the savings and require still more compute. Assessing infrastructure only through energy per answer cannot describe the economic system being created.
More complete metrics are needed: power and water, emissions, contract duration, capacity actually used, sources of capital and distribution of risk. Benefits must also be measured: stable employment, access for research and local companies, heat reuse and grid investment. A general promise to become an AI hub is not a substitute for a verifiable balance sheet.
The hidden debt of data centres reminds us that the cloud is not immaterial. Every apparently instant model rests on decisions made years earlier and expectations projected far into the future. AI may become fundamental infrastructure; precisely for that reason it should not grow through invisible costs. Making financing legible is part of the same transparency we demand from models.