AI debt is safest where a big tech tenant pays the rent, and riskiest where an AI lab's revenue or a GPU's resale value must repay it. Oracle's notice on Project Jupiter adds a third risk: power arriving late

Edition #006 · 28 September 2026 · Article 1 of 8 · All articles in this edition

Download PDF

>>> Whether an AI loan gets repaid depends on who pays the rent and what backs the loan. Where Microsoft, Google, Meta, Amazon or Oracle has signed a long lease, the loan is relatively safe. Where repayment depends on an AI lab's sales, or the collateral is GPUs that lose value within a few years, it is risky. This week showed a third risk that can hit even the safer loans: the building is ready but the power is not.

What happened

The project. Project Jupiter is a 1,400-acre AI data-centre campus in Doña Ana County, New Mexico, near El Paso, Texas. It is one of the Stargate sites that Oracle, OpenAI and SoftBank are building to give OpenAI computing capacity. Stack Infrastructure, a subsidiary of Blue Owl Capital, owns and develops it. Oracle rents the whole campus on a long lease and runs it for OpenAI.

The notice. On 24 September Bloomberg reported that Oracle had sent Stack a notice citing force majeure. Force majeure is a contract clause that excuses or delays a party's obligations when events outside its control intervene. Oracle is not walking away: it cannot end the lease, and it will owe the full rent for the full term once rent starts.

Why Oracle sent it. Under the lease, Oracle, not the landlord, must bring power to the site. The campus is meant to make its own power from gas fuel cells, and both the gas pipeline and the air permit for those fuel cells are late (details below). If both sides agree the power delay counts as force majeure, Oracle can start paying full rent up to three years later than planned. Until then it keeps paying a lower development-stage rent, which covers the interest on the campus's bank loans. Oracle sent the notice early because courts enforce force majeure notice deadlines strictly; a party that waits can lose the right. Oracle calls such notices "commonplace in developments of this scale". The notice is not a default under the loans.

What a delay would mean. The banks keep receiving interest, but wait longer before their loans can be refinanced into long-term bonds. Blue Owl earns about 9% on its equity for longer, instead of the about 11% it expected once the campus is complete.

The power problem is physical. The campus is designed to make its own electricity: Bloom Energy fuel cells producing up to 2.45GW on site, fed by a new gas pipeline. Two approvals are missing. New Mexico has not issued the air permit, and must decide it by 23 November. The 18-mile Green Chile pipeline was due in service in August 2026. Its date has moved to February 2027. The state land office has twice rejected the pipeline's crossing of a 0.6-mile stretch of state trust land. Under county tax-break agreements, a first phase was due to operate between October and December 2026, with the whole campus and power plant complete by Q3 2028.

Is this structure common? Yes, but the terms vary. A special-purpose company owns the campus, funded by a private-markets manager's equity and bank loans. A big tech tenant signs a long lease, and the rent repays the debt. Ten other large AI campuses and chip deals since 2025 that we checked use this model or a close variant. Examples include Meta's Hyperion (Blue Owl) and El Paso (BlackRock's GIP and HPS) campuses, Oracle's Abilene, Vantage and Michigan sites, and the chip vehicles for xAI and Anthropic. The difference that matters is who carries delay. Meta's leases put construction and power risk on Meta, start rent on a fixed date, and add a residual value guarantee. No such guarantee is disclosed for any Oracle-leased campus, and Oracle's sites start on bank loans, so banks and equity carry the delay until bonds are sold. Stargira has written a background paper comparing the eleven deals. [Background paper: link to follow]

Who holds what. Blue Owl owns the project through Stack and has about $3bn of equity in it; Oracle is responsible for paying the debt costs. No source describes Blue Owl as a lender to the campus. The campus secured about $18bn of loans from a consortium of almost 20 banks, according to earlier press reports. During development Blue Owl earns a 9% yield on its equity, rising to about 11% levered on completion, according to a person familiar with the deal. Before the notice, the FT reported the ~$18bn of loans quoted at 89 to 91 cents on the dollar, with banks struggling to distribute them.

On the record. Oracle: "Project Jupiter remains on our planned schedule. We are fully committed to New Mexico and confident in our path forward." Its spokesman added that such notices "do not, by themselves, establish a project delay or change delivery expectations". Blue Owl: "This notice does not change the financial commitments to this multi-year project." Oracle shares fell about 4% on the report.

The tenant's own numbers. S&P cut Oracle to its lowest investment-grade rating in early July. Oracle's 10-Q shows $664bn of remaining performance obligations (signed contract value not yet booked as revenue) at 31 August, of which it expects to recognise about 13% in the next 12 months. That is roughly $86bn within a year and about $578bn over later years, on our arithmetic. Oracle also has $288bn of lease commitments, substantially all for data centres, that are not yet on its balance sheet. They are expected to start between the second quarter of its fiscal 2027 and fiscal 2029, and run for 15 to 19 years. Its borrowings stand at $125.3bn.

The same week, four more price points

• SoftBank Group priced about $11.1bn of senior notes, all issued at par. Three tranches are in US dollars: $1.0bn due 2030 at 8.625%, $4.5bn due 2032 at 9.25% and $4.5bn due 2034 at 9.75%. Two are in euros: €500m due 2030 at 7.125% and €500m due 2032 at 8.000%. Over US Treasuries of the same maturity, the dollar tranches priced about 365bp, 425bp and 469bp wide, on our calculation from the Treasury curve of 23 September. The average BB-rated US bond paid 159bp that day (ICE BofA BB index), so SoftBank paid about 2.3 to 3 times the spread of a typical bond of its rating. The notes are unsecured and rated BB+, below investment grade, by S&P and Fitch's Japanese units. The money funds the $10bn final tranche, due 1 October, of SoftBank's $30bn follow-on investment in OpenAI. The dollar book topped $30bn and each tranche priced at or below the bottom of its range, at the highest yields SoftBank has ever paid on dollar bonds of those maturities. The FT calls it the largest junk bond sale on record.
• Firmus Technologies, an Nvidia-backed Australian group, is in talks for about $10bn: roughly $7.5bn of senior and mezzanine debt plus $2.5bn of equity. Bloomberg reported indicative pricing of about SOFR+275bp for the senior debt and SOFR+725bp for the mezzanine. SOFR is the US dollar base rate that floating-rate loans are priced over. Mezzanine debt ranks behind the senior loan, so it is paid more. Nvidia would provide revenue sharing and credit support. The money buys GPUs for a 360MW campus in Batam, Indonesia, ahead of a planned Australian IPO in late October of up to $5bn. On our reading, the ~450bp gap between the two is the price of sitting second in a GPU-backed deal. For comparison, ordinary (non-AI) middle-market senior loans paid SOFR+475 to 550bp in Q2 2026, according to Valuation Research Corp. The Federal Reserve puts the 2025 average at about SOFR+500 for private credit and SOFR+400 for syndicated leveraged loans. So the Firmus senior debt, with Nvidia's support, is priced about 200–275bp tighter than a typical mid-market loan, and only the mezzanine pays more. These are terms in talks, not a signed loan.
• Nscale raised $3.36bn of convertible loan notes led by Third Point, with $2.36bn funded at closing and $1bn from Nvidia due mid-November. The notes turn into shares when the IPO completes. Funds managed by Apollo Global Management, Citadel, Hudson Bay, Abu Dhabi Investment Council and Wellington took part. Nscale cites over $103bn of total contracted value.
• SB Energy, a data-centre and power developer, postponed its IPO from September because bankers struggled to find buyers at a valuation above $50bn, the New York Times reported. A person familiar told Investing.com the IPO is "advancing according to plan". Its S-1 shows about $439bn of backlog, which the filing itself calls "a hypothetical estimate based on management's assumptions", and no data centre yet operating. It says it is "substantially dependent on OpenAI as a tenant". Investing.com's source said the offering includes up to $500m for Japanese retail investors.

Concentration behind the layers. Ares Management traced 26 digital-infrastructure financings, about $573bn of debt, from the past year. Its conclusion: "all of the risk converges on just eight names: Meta, Oracle, Microsoft, Amazon, Google, Nvidia, and, on a look-through basis, OpenAI and Anthropic". Ares puts digital infrastructure at 42% of investment-grade private placements this year. With rare exceptions, each deal was engineered to reach an investment-grade rating.

The house view: where the loss risk sits

Stargira's research paper of 28 September, AI debt and private credit: why the local-LLM risk is bigger than it looks, ranks AI lending by what repays each loan. It then asks which lenders lose if inference (running trained AI models) moves from big cloud campuses to cheaper local and open-weight models. It sizes private credit's exposure from two sides, loans that build AI and loans to software companies that AI may replace, and runs an illustrative stress case. The paper is sold separately. [Buy the paper: link to follow]

Is AI inference moving off the giant campuses? Partly: labs now also rent small sites, but demand for cloud space is still at a record

Anthropic and OpenAI are shopping for 20–30MW data-centre blocks alongside their gigawatt projects, CNBC reported on 18 September; DCD followed on 23 September. OpenAI said "Different workloads need different infrastructure." Jabez Tan of Structure Research: "Many inference workloads can instead serve separate requests across multiple smaller clusters, opening up more locations." JLL expects inference to overtake training as the main data-centre requirement in 2027, rising from 9% of workloads in 2025 to 37% by 2030.

This suggests inference may spread away from giant lab-tied campuses. But it is still cloud capacity rented by labs, not local inference. Cloud demand is still rising: vacancy in North America's primary data-centre markets hit a record low of 1.4% in H1, and Google Cloud carries $513.9bn of backlog. Total demand may grow fast enough that cloud inference keeps rising in absolute terms.

Sources: note 1.

For Wealth Managers:
The product question has changed from "how much AI?" to "which layer of AI?". Two funds that both say "AI infrastructure credit" can hold very different risks. This week the same build-out was priced at SOFR+275 for senior GPU debt (in talks), 9.75% in unsecured BB+ holding-company bonds, and 9% rising to about 11% on equity in an Oracle-leased campus.

Jupiter also shows that an investment-grade tenant protects against default, not against delay. Oracle pays the debt costs, so a rent deferral would fall first on the equity, which lives on the rent. The bank loans were already quoted at 89–91 cents on the dollar before the notice. Before you hold or sell a semi-liquid credit fund, BDC or infrastructure-debt fund, ask the manager four questions:

1. What share of the book is lab-tied chip vehicles, GPU-backed neocloud loans or speculative construction (layers 4–6)?
2. What share is loans to software companies, which a shift to cheaper or local inference would hit at the same time?
3. For data-centre loans, who carries delay risk before power arrives, and does rent start on a fixed date or on completion?
4. Counted through the tenant, how many of Ares's eight names sit behind the book?

A manager who cannot answer by layer is reporting "digital infrastructure" as one line. That is no longer enough.

For Fund Managers:
Capital is available, but it is priced by where you sit in the stack. Senior GPU debt with vendor support is quoted near SOFR+275; mezzanine needs SOFR+725 even with Nvidia behind it. Nscale chose convertible notes, which give lenders equity upside if the IPO lands. SoftBank's $30bn-plus book shows deep demand for AI-linked high yield at the right price. SB Energy shows the limit: bankers struggled to fill an equity book for a pre-revenue, single-tenant backlog.

Two clauses from Jupiter now matter in every build-to-suit loan: the force majeure definition, and who bears power delivery. Meta's Hyperion joint venture with Blue Owl funds is the stronger template for the leased-building layer: a 16-year residual value guarantee over a four-year initial lease. Expect LPs and insurers to ask for exposure by layer and by look-through counterparty. Managers who can show layer mix, rent-start triggers and guarantor coverage will have an easier time with wealth platforms and insurance mandates.

What This Changes:
"AI credit" is splitting into priced layers. It was sold as one theme: contracted cash flows from the world's strongest companies. This week put separate, visible prices on the layers. The paper's map is becoming the market's own way of pricing.

Jupiter adds a risk the paper's map did not stress: timing. The paper ranks layers by their exposure to a shift toward local inference. Jupiter shows a second axis. A campus can be sound on credit and still fail to pay on time, because power, permits and pipelines sit outside the lease. Oracle alone has $288bn of leases still to commence, substantially all for data centres. Each rests on a power and permit timetable.

Concentration also weakens the idea that the safer layers are independent of the risky ones. Ares links Oracle's leases to OpenAI's compute commitment. SoftBank's bonds fund OpenAI equity. SB Energy's backlog rests mostly on OpenAI's leases at one Ohio campus. Layers 1–3 are safer against local inference. They are not safe if a hub counterparty weakens. Read the paper's layer map and Ares's counterparty map together.

This partly answers edition 005's watchlist item 4, which asked whether a second hyperscaler would walk from a commitment. Oracle did not walk; it kept the lease and sought a right to delay rent. A delay right would squeeze Blue Owl's equity first. The banks' protection is that Oracle pays the debt costs.

Competitive Landscape

---

Sources

[1] Lead Story — AI debt by layer, and Oracle's Project Jupiter notice. Bloomberg — 2026-09-24; TechCrunch — 2026-09-24; Reuters via Yahoo Finance — 2026-09-24; El Paso Matters — 2026-09-24; Santa Fe New Mexican — 2026-09; Insider Monkey via Yahoo Finance (citing FT) — 2026-09-27; Oracle Form 10-Q — 2026-09; Ares, In the Gaps (Fall 2026) — 2026; SoftBank Group release (JA) — 2026-09-24; Yahoo Finance (SoftBank) — 2026-09-24; Bloomberg via Yahoo Finance SG — 2026-09-23; FT (standfirst) — 2026-09-24; Bloomberg via Seeking Alpha/TradingView (Firmus) — 2026-09-23; SB Energy Form S-1 — 2026-09-01; TheFly citing NYT — 2026-09-22; Investing.com via Yahoo Finance — 2026-09; Nscale release — 2026-09-25; SiliconANGLE — 2026-09-25; Meta release — 2025-10-21; BIS Quarterly Review — 2026-03; Morgan Stanley Research — 2025; Alphabet Form 10-Q — 2026-07; CBRE — 2026; Campus Technology (Broadcom survey) — 2026-07-01; arXiv 2511.07885 — 2025-11; AIMA / ACC — 2025-12-09; CNBC — 2026-09-18; JLL — 2026-01-06; Stargira house research paper (internal) — 2026-09-28; Bloomberg Law (Bloomberg text) — 2026-09-24; Quinn Emanuel, force majeure in AI data-centre contracts — 2026-06-03; Yahoo Finance (Meta El Paso notes) — 2026-07; Yahoo Finance SG (Vantage $38bn loans) — 2025; Related Digital (Michigan) — 2026-04; CoreWeave DDTL 3.0 — 2025-07; US Treasury par yield curve — 2026-09-23; FRED, ICE BofA BB OAS — 2026-09-23; Valuation Research Corp, Private Markets Trends Q2 2026 — 2026; Federal Reserve, FEDS Notes — 2026-08-11; BIS Bulletin 128 — 2026-07-14

Download PDF

---

Disclaimer: Private Credit & Infrastructure Intelligence is published for general information and education only. It is not investment advice, and it is not an offer, solicitation or recommendation to buy, sell or hold any financial instrument, nor an invitation to engage in any investment activity. Nothing in it takes account of any reader's circumstances, objectives or needs. Figures and sources are believed accurate at the dates shown, but completeness and continuing accuracy are not guaranteed. Take your own professional advice before acting on anything here.