Nine months ago, Meta could not have asked for an easier ride from the bond market. Its $27 billion Hyperion financing in Louisiana, backed by Blue Owl Capital, priced on generous terms and was treated by investors as close to a formality: a hyperscaler’s name on the lease, a private credit consortium behind the debt, done.
This week’s sequel is a different animal. BlackRock is marketing more than $12 billion in bonds through Project Sopaipilla Holdings, an 80/20 joint venture with Meta, to fund a roughly one-gigawatt campus in El Paso, Texas, due online in 2028. The mechanics are near-identical to Louisiana; however, the market’s response is not.
Bond investors are now seeking yields north of 7% on the El Paso paper, a marked jump from the terms Meta secured in October. The repricing is the real story – the $12bn headline is not new information; what it costs to raise that $12bn is.
Why the market has started pricing AI debt differently
The structure behind both deals is now familiar territory for anyone following data centre finance: a special purpose vehicle, majority-owned by a private capital partner, raises debt secured against a long lease to a single, creditworthy tenant. It is project finance dressed for AI, and it has let Meta keep the bulk of tens of billions in infrastructure spend off its own balance sheet while still controlling the campuses it needs.
What has changed is not the structure. It is what investors are prepared to pay for it. UBS strategist Matthew Mish has described the pace of AI-related debt accumulation, running at roughly $100 billion a quarter by some estimates, as a trend that “would make anyone familiar with credit cycles raise an eyebrow.”
That caution is now showing up in pricing rather than commentary. Where the Hyperion bonds priced to reflect confidence in Meta’s credit halo, the Sopaipilla bonds are being asked to stand more on their own economics, and investors are charging for that shift.
Meta’s own creditworthiness has not deteriorated. What has shifted is how much protection the market wants against project-specific risk sitting one step removed from Meta’s balance sheet, at a moment when the volume of similar deals from Meta, Oracle and others is starting to test how much of this paper the market can absorb at once.
The mismatch nobody quite underwrote
Strip away the ownership percentages and the more uncomfortable structural detail is one that gets far less attention than it deserves: a duration mismatch running through nearly every one of these deals.
The bonds behind El Paso and Hyperion alike are long-dated, in some cases running to 2049. The GPUs inside the buildings they finance have a useful life closer to three to five years. The leases underpinning the debt, in several of these structures, run shorter still than the campuses they were raised to build.
That is not a hidden flaw so much as a deliberate trade. Lenders are betting that AI compute demand, and Meta’s willingness to keep paying rent, holds for long enough to service debt that will outlast several hardware refresh cycles. Bank for International Settlements analysis of the wider hyperscaler financing wave makes a related point about disclosure: much of this borrowing sits in vehicles capitalised through private placements, backed by long-term operating leases and offtake commitments, rather than showing up as conventional corporate debt. Investors evaluating the coupon are, in effect, also underwriting a bet about how long AI infrastructure stays economically useful once it is built, a question that has no settled answer yet.
A repriced yield on a single deal is a data point. A structural mismatch between 25-year debt and five-year assets, repeated across dozens of SPVs now embedded in the sector’s capital stack, is a sector-wide exposure. It sits alongside the questions Capacity has already raised around vendor and counterparty concentration in AI-linked financing, most recently in the $35 billion Apollo and Blackstone facility backing Anthropic’s compute build-out.
Two hyperscalers, two different bets on their own balance sheets
The El Paso deal also sharpens a divergence in strategy that is easy to miss when every headline reduces to “hyperscaler raises billions.” Meta continues to route the large majority of its AI infrastructure spend through off-balance-sheet SPVs, keeping formal leverage off its own accounts while a private capital partner carries the debt and the residual value risk.
Oracle has taken the opposite route, raising tens of billions directly, with roughly $43 billion in debt financing in fiscal 2026 alone and further raises planned for fiscal 2027. S&P responded by downgrading Oracle to BBB-, just one notch above junk, citing the strain of AI infrastructure spending combined with customer concentration.
Meta’s approach protects its own credit rating and reported leverage but pushes execution risk, and now visibly higher borrowing costs, into vehicles like Sopaipilla and Beignet. Oracle’s approach keeps the debt visible and accountable on its own books, and the rating agencies are pricing that transparency as risk in its own right. Neither path is free. One prices the risk on the hyperscaler’s own credit statement; the other prices it into a rating agency’s downgrade commentary and, increasingly, into the yield an SPV has to offer to get a deal away.
Moody’s has separately flagged that some of this off-balance-sheet activity understates the scale of hyperscalers’ effective obligations, since short initial lease terms and residual value guarantees do not always show up as balance-sheet liabilities in the way the underlying economic commitment would suggest.
What it means going into the next wave
None of this suggests the AI infrastructure financing boom is stalling. Morgan Stanley has estimated that tech companies and related parties could need to raise up to $800 billion in private credit through asset-specific structures by 2028, and JLL’s most recent outlook puts total data centre investment as high as $3 trillion over five years. The pipeline is not the question. The price of capital inside that pipeline is.
El Paso is the clearest evidence yet that lenders are starting to differentiate between AI infrastructure deals rather than treating them as a uniform asset class carrying a hyperscaler halo. That has direct implications for anyone in the sector currently structuring, negotiating or benchmarking financing: the terms that closed in October are no longer the terms on offer in July, and the gap between them is a live signal about how much scrutiny the next tranche of AI debt is going to face.
RELATED STORIES
SpaceX’s $20bn bond deal: AI ambition meets ESG backlash
Data centres at a crossroads: Can capital keep up with demand?
Data centre investments could reach $3tn in five years: Moody’s report
Edged US closes $2bn in financing to power AI data centre push
Trump renews push for AI giants to foot data centre power bill
Trump’s UK visit: A pledge to support data centre investment and industrial growth
