Your Compute Supplier Is Now a Credit Exposure
Permits, power and borrowed money are the weak links in AI capacity, and the contracts along the chain already decide who pays when one of them breaks.
Oracle says Jupiter “remains on our planned schedule,” Augment reports. Yet the same notice lets it delay payments to Blue Owl. Morning Brew reports Oracle warned it will do exactly that if the site is not online by 2028. Bloomberg describes the notice as a shield against mounting expenses at a site that had already faced local opposition and regulatory setbacks. The tenant stays publicly committed while the cost of waiting moves to the developer and its lenders.
Equity investors barely reacted. Oracle fell about 3.5%, Blue Owl fell 3.6%, and the Nasdaq closed essentially flat, per Morning Brew. The bond market moved first. That is the pattern to expect when risk is being transferred rather than removed.
Nscale shows the same mechanic on a balance sheet
Nscale is targeting a $35B IPO on a $103B backlog. Augment reports it disclosed substantial doubt about its ability to continue as a going concern until Nvidia stepped in. Nvidia committed $1B in convertible notes or non-voting shares, bought $60M of warrants, and guaranteed up to $860.3M of Nscale's lease obligations. Nscale lost $1.02B on $140.6M of first-half revenue. About 85% of its backlog, roughly $87.6B, sits with Microsoft and Anthropic, and much of it is not yet firm. AI Breakfast adds that Nscale's UK site has slipped from 2027 into the 2030s.
When a chip vendor finances the buyer of its chips, part of the demand you see is subsidized demand. The capacity prices behind your compute budget are partly set by a supplier propping up its own customer.
| Link in the chain | Who | How risk moved | Who absorbs it |
|---|---|---|---|
| Anchor tenant | Oracle at Jupiter | Force majeure notice tied to a pipeline permit | Developer Blue Owl and its lenders |
| Project lenders | Holders of $18B of Jupiter debt | Paper repriced to 90 cents | Bondholders |
| GPU cloud provider | Nscale | Going-concern doubt cleared by vendor financing | Nvidia, and customers holding non-firm capacity |
| Enterprise buyer | You | Standard capacity terms | Whatever the links above pass down |
Where the reporting agrees, and where it splits
Morning Brew, Bloomberg and Augment agree that the binding constraint has moved from chips to permits, power and balance sheets. Rates make it worse. With the 10-year Treasury at 5.162%, Morning Brew notes that project finance takes a double hit from higher base rates and wider spreads. That favors hyperscalers that fund themselves over leveraged, project-financed builds, which are the ones most likely to slip.
The three disagree on how much concentration is too much. Augment would let no at-risk provider back more than a quarter of production workloads without tested failover. Morning Brew suggests capping any single provider or unpermitted site at roughly 40% of critical workloads. Bloomberg calls it a concentration problem once more than half of critical capacity depends on one provider's unbuilt sites. None of these is an industry standard. Each is an editorial judgment. Your board should see your own threshold written down.
Announced gigawatts are a press release. Energized, permitted power is the only capacity that belongs in a delivery plan.
The smart move
Underwrite each material compute provider the way a lender would. Ask four questions. Is power delivered to the site? Is your capacity firm? Who is financing the provider? What does the contract let them do if a permit fails? The answers tell you which suppliers can carry critical workloads and which need a tested fallback behind them.
What to do
Direct your CFO and general counsel to map every 2027–28 compute commitment within 30 days by provider, site and power status (energized, permitted or planned), flagging each force majeure and delay clause.
Add termination and step-in rights on going-concern or force majeure events, plus workload portability, to every compute renewal this quarter, and put a board-approved cap on capacity sourced from at-risk providers.
Negotiate a second-source or burst agreement this quarter with a provider that has energized power and a strong balance sheet, so critical AI workloads survive a 6–12 month slip at a primary site.