The Non-Demand Risk to the AI Trade: What Oracle's Force Majeure Notice Tells Us

Summary

The event. Bloomberg reported that Oracle sent a force majeure notice to the developer of Project Jupiter, a unit of Blue Owl, to shield itself from cost overruns and preserve the option to delay payments if the site slips past its planned 2028 opening, while remaining the main tenant. The campus is a 2.45GW part of the Stargate build-out. Oracle's position is that "Project Jupiter remains on our planned schedule." Blue Owl says the notice does not change the financial commitments to the project.

The thesis. We think this is a very real, non-demand risk to the AI trade. Most bear cases ask whether end demand will show up. This one shows the trade can be strained by the physical and bureaucratic step between capital committed and megawatts energized, with no demand shortfall required.

The mechanism.

• Developer economics. Permitting delays push back power-on dates. Developers like Oracle lose revenue in the early years, which are the most valuable for project returns. A late start can push ROI below breakeven and turn a projected winner into a loss-making project.

• The chip layer. Nvidia's Jensen Huang and Broadcom's Hock Tan, the two most important chip suppliers, have both pointed to powered data center capacity as the binding constraint on delivering their sales and earnings targets. No powered shell means no chip deployment and no chip revenue.

• Rip and replace. In a power-constrained world, the scarce input is watts, not chips. That can justify ripping out older accelerators and replacing them with newer ones to maximize tokens per watt. This pulls forward capex, strands existing assets and weakens the economics of the installed base.

• Second-order losers. The data center infrastructure and hardware layer suffers. Neoclouds and clouds that can't turn capacity on fail to earn an adequate ROI. The exceptions are operators who have already secured and energized power. They can price at a premium, so scarcity favors incumbents with power in hand.

• Leverage and the financiers. Much of the new build is debt-funded. The campus reportedly lined up $18 billion in loans, so any hint that tenant payments are less "locked in" gives lenders a reason to scrutinize the cash flow assumptions. The exposure runs through the big banks, private credit players like Blue Owl, and Nvidia itself, which has put its own balance sheet on the line by backstopping data centers and supporting offtake.

The caveat. This is one example, not the norm. Most projects are not encountering this, and we are not forecasting a wave of force majeure declarations. But it belongs on the risk register. Demand doesn't have to be what breaks the AI trade. Power availability and bureaucratic red tape could do it.

1. What happened

Project Jupiter is a 1,400-acre campus in Doña Ana County, New Mexico, that supports Oracle's capacity commitments to OpenAI. OpenAI signed a $400 billion deal with Oracle and SoftBank last year covering five US data centers, including this one. Stack Infrastructure, a Blue Owl portfolio company, is developing it, and Blue Owl funds have committed equity.

The delays are physical and regulatory:

• A natural gas pipeline from Energy Transfer that was meant to start service this month was pushed back nearly six months to Feb. 1, 2027, after repeated permit denials from the New Mexico State Land Office over the proposed route.

• Without the pipeline, the Bloom Energy fuel cells cannot operate, which leaves the campus without a viable power source.

• The campus also faces litigation over water and air permits. On Sept 17, the New Mexico Supreme Court unanimously rejected emergency challenges from environmental groups, which removes one near-term obstacle.

2. Why delay is so costly: the project economics

Data center returns are front-loaded around lease-up and ramp. A six-to-twelve-month slip in power-on means:

• Lost initial-year revenue that can't be recovered later

• Continued interest carry on construction debt

• Potentially higher construction and equipment costs from repricing

• IRR compression, and in the tail case a project that falls below breakeven

3. Power is the binding constraint for the chip vendors

The industry tends to treat chip supply as the scarce input, but Jensen Huang and Hock Tan's commentary points the other way. If powered, cooled and energized data hall space is scarcer, chips become the downstream variable.

Customers place accelerator orders against power-on dates, so a slipped energization date pushes out deliveries or the order itself. Because AI capacity is concentrated in a small number of very large projects, one 2GW site can be a meaningful share of a vendor's annual shipments. A GPU delivered to a site with no power also still starts depreciating while producing nothing. The bottleneck isn't only chips either:

• Long-lead electrical equipment such as turbines, transformers, switchgear and fuel cells

• Networking, optics and memory suppliers, who share the same deployment timing

The implication is that vendor earnings depend on more than demand and their own supply capacity. They also depend on other people's permitting outcomes.

4. Rip and replace: when watts, not chips, are scarce

If a site's power is fixed, its revenue is roughly megawatts times tokens per watt times price per token. When new megawatts can't be added, the only lever left is raising tokens per watt, which means swapping hardware. Each accelerator generation delivers materially more output per watt, so in a power-constrained environment, replacing a working older fleet can beat waiting years for new power.

The consequences are uncomfortable for operators. Capex is pulled forward, since hardware may be replaced before it has earned back its cost. Multi-year useful-life assumptions look aggressive if equipment is retired early for power reasons, which invites impairments. Debt secured against GPUs relies on residual value assumptions, and faster replacement cycles, plus older chips flooding the secondary market, weaken that collateral just when lenders most want it to hold. It is good for vendors' next-generation sales, but it comes out of their customers' balance sheets.

5. Who wins and who loses

• Losers: developers with unenergized pipelines, the data center infrastructure and hardware layer when orders slip, and neoclouds and clouds without secured power, especially the highly leveraged ones.

• Winners: operators with power already secured and turned on, who can price at a premium as energized capacity becomes scarce. Parts of the power supply chain, such as equipment makers and generation developers with permits and grid access, benefit from the same scarcity.

The dividing line is not who has the best model or the most demand. It is who has energized megawatts, and time-to-power becomes the moat.

6. The financing chain: where a project problem becomes a system problem

Much of the recent build is financed with leverage, and that is what turns a permitting delay into a wider financial issue. Banks size construction and project loans on a tenant's committed payments and a delivery schedule, so delay compresses debt service coverage and can trigger covenant and refinancing pressure. Private credit firms like Blue Owl are both lenders and equity sponsors, so they carry exposure at more than one layer, and marked-down loans affect fund performance and investor sentiment.

Nvidia sits in an unusual position. By backstopping data centers and putting its balance sheet behind offtake, it has taken on demand and delivery risk from the customer side. Its earnings then depend on projects it is also helping to finance, so a stumble can hit revenue and balance sheet at once.

The real danger is correlation. These projects share suppliers, counterparties, permitting climates and politics, so a stressed loan on one flagship site can widen spreads across the category and make the next financing harder, even for well-permitted projects. The campus has also become a symbol of the backlash against the data center boom, just ahead of the US midterms, so the political heat is unlikely to fade soon.

7. Keeping it in proportion, and what to watch

This is one project. Oracle says it's on schedule, Blue Owl says its commitments are unchanged, and a force majeure notice can be a precaution rather than a sign of distress. We aren't calling this the norm, and demand for AI compute remains strong on most measures. The point is that a non-demand failure mode exists and is under-discussed.

What we'd watch:

• Further force majeure notices or delay claims on other large campuses

• Pricing of project loans and data-center-linked credit

• Progress on the Project Jupiter pipeline, with Feb. 1, 2027 as the next marker

• Vendor commentary on deployment timing and power availability

• State-level permitting posture and premium pricing for energized capacity