About $18 billion of loans tied to Oracle-leased Project Jupiter in New Mexico are being quoted by syndicate banks at roughly 89–91 cents on the dollar. That market signal exposes a financing constraint that AI demand alone cannot solve.
Reuters, citing Financial Times reporting, said banks including Santander and Jefferies have struggled to distribute the project debt as broadly as planned and are consequently retaining more exposure on their own balance sheets. Oracle, Santander and Jefferies did not comment to Reuters on the report.
AI infrastructure has three markets that must clear
The useful finance framework is to separate three markets that are often collapsed into one AI-growth narrative.
- The customer market: Is there contracted demand for compute?
- The physical-infrastructure market: Can land, power, permits, equipment and construction be assembled on schedule?
- The capital market: Will lenders and investors finance those assets at a price that preserves acceptable returns?
A signed compute contract can help clear the first market. It does not automatically clear the other two.
Why loan distribution matters
Large project loans are often originated by a group of banks with the expectation that portions of the exposure will subsequently be distributed to a wider pool of institutional investors. Successful distribution frees bank balance-sheet capacity and establishes a market-clearing price for the risk.
When distribution stalls, the economics change. Banks may hold more exposure than intended, and a below-par secondary-market indication becomes a visible measure of the discount investors require for credit, execution and liquidity risk.
An 89–91 price does not prove that a project will fail. It does say that the market currently values the loans below their face amount. For CFOs, that is important because the cost and availability of the next financing can be influenced by what happens to the last one.
Physical constraints can migrate into financial risk
Project Jupiter illustrates how infrastructure execution and credit can become intertwined. Reuters reported concerns around the project’s power and permitting environment. Oracle has separately said that the data-center campus and proposed microgrid are distinct facilities and that construction of the data-center buildings continues under county permits; a court stay concerns the separate microgrid air-permit proceeding.
Oracle is also seeking proposals for 2 gigawatts of new renewable generation in New Mexico, with projects targeted for delivery between 2027 and 2031. The company says it will prioritize and fund projects capable of accelerated delivery and is targeting 100% carbon-free energy matching for Project Jupiter by 2031.
The distinction matters. A power constraint starts as an engineering or permitting problem. If it delays energized capacity, it can become a construction-schedule problem. If that delays customer delivery or changes financing assumptions, it becomes a financial problem.
Demand is necessary, but financeability determines scale
The AI infrastructure boom has often been analyzed through contracted demand: cloud commitments, GPU orders and long-term compute agreements. Those measures are important, but they do not reveal the full capital requirement.
Every new data-center campus requires a financing stack capable of carrying construction risk, equipment purchases, power infrastructure and the period before the asset reaches steady-state utilization. If the market-clearing yield demanded by lenders rises, project economics can deteriorate even when customer demand remains strong.
This is particularly relevant when multiple hyperscale projects compete simultaneously for the same pools of bank capital, private credit, infrastructure funds and long-duration investors.
The CFO scorecard should extend beyond the customer contract
A better AI-infrastructure scorecard would track at least six separate variables:
- contracted customer demand;
- committed versus still-needed financing;
- loan distribution and secondary-market pricing;
- power availability and interconnection certainty;
- construction and permitting milestones; and
- the project’s weighted cost of capital relative to expected cash returns.
These measures help distinguish a commercially attractive project from a financially executable one.
Capital markets can become the bottleneck
Project Jupiter does not establish that AI infrastructure broadly is becoming unfinanceable. The reported loan pricing is one project-specific market signal, and Oracle has publicly disputed some implications surrounding the project’s permitting situation.
But the episode demonstrates why the next phase of the AI buildout should be evaluated through more than demand forecasts. The sector must continuously convert customer commitments into financed assets, energized capacity and returns sufficient to compensate capital providers.
The ultimate constraint on AI infrastructure may not be demand for compute. It may be the price at which capital markets are willing to finance the assets required to supply it.
Sources
- Reuters, “Oracle’s $18 billion data center debt under pressure, FT reports,” September 18, 2026.
- Oracle, “Oracle Issues RFP for 2 GW of Renewable Energy in New Mexico,” September 8, 2026.
- Oracle, “Project Jupiter Statement on Construction Permitting,” September 14, 2026.
By Robert Young | Numbers & Judgment. Published September 20, 2026.
Numbers & Judgment 
Leave a comment