Authored by Patrick Feeley via Substack,
Private credit is financing the AI buildout against collateral with a short clock, and I think Washington ends up holding the risk…
Karp’s logic
Alex Karp, one of the most contrarian CEOs ever, spent September making an argument most of the market has not yet absorbed, and I am writing this because I think it applies one layer further down the stack than he took it. I admire the way he reasons. He holds a doctorate in social theory from Goethe University Frankfurt, and it shows in a habit I find rare among technology chief executives, which is to treat the state and the tech industry as one system and to say plainly what the people inside it prefer to leave unsaid. His 2025 book The Technological Republic, written with Nicholas Zamiska, argues that the software industry has to work with government, and it asks leaders to preserve space for ideological confrontation and to risk the disapproval of the crowd. As I read him, the frontier labs carry liability too large for any public market to hold, so the only buyer of that risk is the sovereign. He told CNBC the only way to deal with that liability is “to go to the government and say, nationalize us, please”. I think he is right, and his reasoning reaches well past the labs to the debt that built the buildings the models run in.
The AI buildout is being financed by private credit, against collateral that loses most of its value in four to six years, leased to a handful of tenants. When the first large tenant walks, lenders will discover they own stranded GPUs in the desert. Washington has already begun treating compute as strategic infrastructure, and I expect it will not let those facilities go dark. Data center credit then becomes what mortgage credit became after 2008, with private gains and a public backstop.
Karp followed the liability to the place where only the government is large enough to stand behind it, and he said so out loud on national television. That is the discipline I want to borrow, because the same chain of reasoning leads straight to the lenders.
Who pays
Morgan Stanley forecasts about $2.9 trillion of global data center capex through 2028, roughly $1.6 trillion of it chips and servers and the rest buildings and infrastructure. Hyperscaler cash flow covers about $1.4 trillion and leaves a $1.5 trillion external financing gap, of which the bank’s analysts see private credit filling about $800 billion. That is one bank’s forecast of what private credit could supply, and the same figure turns up in the Financial Stability Board’s private credit report with Morgan Stanley as its source. Morgan Stanley was also the sole bookrunner on the Hyperion bond, so it has a commercial interest in the story.
The marginal dollar is moving away from public markets. Demand coverage on hyperscaler bond deals fell from nearly five times in February to below two times in July, which I read as financing shifting toward private credit and off-balance-sheet vehicles.
The template is Meta’s Hyperion campus in Louisiana. Funds managed by Blue Owl own 80% of the joint venture’s equity and Meta owns 20%. The project’s bonds total $27.3 billion and were anchored by PIMCO. Meta leases the campus on an initial four-year term with renewal options, and Meta’s latest 10-Q puts its maximum exposure to loss on its unconsolidated data center ventures at $68.4 billion at September 30, including $41.5 billion of residual value guarantees. An IFR opinion column by an outside author, a former UBS banker, describes the A+ rated bonds and a 16-year residual value guarantee, compares the deal to off-balance-sheet structures from the Enron era, and in its body calls it a corporate-scale version of the accounting that let Fannie Mae and Freddie Mac fund inside their standalone credit, with Meta carrying the economic risk while the debt sits outside its books.
Hyperion is one of many such deals. By the Financial Times’s count, tech firms had moved more than $120 billion of AI data center spending off their books through special purpose vehicles as of December 2025, a figure that includes a $20 billion xAI raise still being pursued at the time. The bondholder is lending to Meta, and Meta’s balance sheet does not say so.
The clock
A building lasts thirty years while the chips inside it do not. Nvidia’s H100 rented for about $8 an hour on demand in summer 2023, against $1.70 in October 2025 and $2.35 in March 2026, which are midpoints of SemiAnalysis’s one-year contract ranges. The comparison mixes an on-demand price with contract prices, and like for like on SemiAnalysis’s spot composite the fall from late 2023 to October 2025 is closer to 57%. CoreWeave depreciates its servers over six years, and Nebius used four until this January and now uses five. When the companies holding the same kind of asset keep revising its useful life, the lender does not know what it owns.
The market has quietly admitted this. CoreWeave’s investment-grade-rated facility from March priced at 225 basis points over SOFR, while its facilities signed since then price at 450 and 550, the newest backed by customers that are unrated or speculative grade, on broadly the same chips. The GPU matters less than the tenant’s promise to pay. CoreWeave owed $35.6 billion of principal at June 30, 2026, of which $13.6 billion was GPU-backed delayed-draw loans, and Moody’s rated its first investment-grade GPU-backed facility A3 in March. By August its newest facility priced at SOFR plus 550, and Octus reported a recent deal pricing at 96 to 97 cents on the dollar, a junk cost of capital.
The guarantees that hold the structure together are circular. Nvidia has residual value guarantees capped at $105 billion for its initial commitment behind OpenAI’s leases at the Ohio PORTS campus, payable to the lessor, SB Energy, if OpenAI defaults, and the total with guarantees for other partners comes to $108.5 billion. The cap is a ceiling and not current exposure, and the guarantees take effect as each lease commences, with the first phase expected in fiscal 2029. OpenAI has agreed to reimburse and indemnify Nvidia for any payout, a promise that is only worth something in the world where the guarantee is never called. What the market sells as diversified asset-backed credit is concentrated exposure to four or five tenants and one supplier’s product cycle, packaged as infrastructure.
The cracks
None of what follows is a default, and all of it is the market repricing who holds the risk.
Oracle invoked force majeure on Project Jupiter, the New Mexico campus it leases from Blue Owl’s Stack Infrastructure. Bloomberg reported the notice on September 23, and if the parties agree power delays qualify, Oracle could defer rent for up to three years, though Oracle says such notices are commonplace and do not by themselves establish a project delay. The project’s financing debt was quoted near 90 cents in September and has since traded near 80. The underwriters, meanwhile, want out. A bank group including JPMorgan and MUFG arranged $38 billion of construction debt for Oracle-leased projects in Texas and Wisconsin, and Bloomberg reported in April that it was taking months to spread across the market.
Issuance has gone junk. Close to half of the roughly $80 billion of data center bonds issued in 2026 are rated below investment grade, which is $37 billion against $8.6 billion in 2025, so the share rose from about 24% to 46% and the dollar volume is more than four times last year’s. Firmus, SB Energy and AirTrunk delayed or shelved IPOs within a month. Hyperscaler credit is at its widest as well, with five-year CDS nearing its widest level of the year in early October and Oracle’s at 257 basis points by October 10, the highest since at least 2009. Behind all of it sits an enormous pipeline, since Microsoft, Amazon, Alphabet, Meta and Oracle report about $1.2 trillion of undiscounted payments on leases that have not yet commenced.
Washington holds it
Nothing on the public record shows that a federal backstop for data center debt is planned or has been requested, so this is my prediction, and the precedent is Fannie Mae and Freddie Mac. The private backstops are already real, with Nvidia’s guarantees of up to $105 billion and Meta’s $41.5 billion of residual value guarantees, and I think that is how a public one starts. The state has already claimed the asset in a looser sense. DOE opened 16 federal sites to data centers in April 2025, the White House’s March 2026 Ratepayer Protection Pledge says data center infrastructure supports economic and national security, and FERC has a large-load rulemaking open. A policy group is pushing for a formal critical infrastructure designation, which tells me none exists yet, and its paper is about security of supply and has nothing to say about financing. Once an asset is called a matter of national security, letting it go dark becomes a policy failure and stops being a credit event.
The industry has floated the idea once already. In November 2025, OpenAI CFO Sarah Friar said at a Wall Street Journal event that a government backstop, the guarantee that allows the financing to happen, could drop the cost of financing and raise the loan to value. The interviewer had asked about a federal backstop and she answered “Exactly,” and within hours she walked it back, saying she had used the word backstop and it “muddied the point”. OpenAI’s October 27, 2025 letter to the White House science office asked for grants, loans or loan guarantees to build industrial capacity in areas such as transformers, rare earths and semiconductors, and for a tax credit expansion that would reach data centers and servers, though not for guarantees on OpenAI’s own data center debt. White House AI adviser David Sacks posted on November 6 that there would be no federal bailout for AI, which is an adviser’s post and not a formal response to the letter, and Treasury Secretary Scott Bessent said on September 21 that AI developers need to take responsibility for themselves. Those answers came before financing tightened, and I expect the question to come back when it has.
The risk is already inside regulated balance sheets. FSOC counts about $445 billion of bank loan commitments to private credit funds as of the second quarter of 2025, which includes undrawn lines. Bloomberg calls insurers the predominant lenders in the private placement part of this market and cites a JPMorgan projection that data center securitization heads toward 7% to 10% of combined ABS and CMBS issuance, which is the profile of a systemic asset class.
Fannie Mae and Freddie Mac borrowed inside their standalone credit because the market assumed Washington stood behind them, and in 2008 the assumption was made explicit. Hyperion paper already prices off Meta’s guarantee, and in my view the next step is pricing it off Washington’s. It will not be called a bailout. It will arrive as a compute continuity facility, a DOE loan program extended from power plants to the data halls, or a conservatorship of a distressed SPV justified on national security grounds.
The mark
Karp used a poker line in September, and it was about AI liability: “If you think the other person’s the mark, you’re the mark. You are the mark.” I borrow it here for the credit market. In this trade the mark is the allocator who believes he bought senior, diversified, asset-backed credit and actually owns correlated exposure to Nvidia’s roadmap and four tenants’ capex budgets, dressed as infrastructure. The asset managers originating it carry the equity risk, and Blue Owl’s shares are down roughly 43% over the past year on price alone, measured from the October 23, 2025 close to the October 23, 2026 close. The taxpayer is the last holder nobody has asked.
What would tell me which way this breaks is mostly visible in the filings and the tape. The first hyperscaler that declines to renew at the end of a four-year initial lease term would matter, as would the first residual value guarantee actually called, and so would the outcome at Project Jupiter, whether the force majeure is resolved or becomes a template for other tenants. I would also watch for further bank sales of data center construction debt at a discount or significant risk transfer trades on that paper, and for any federal program language that extends loan guarantees from fabs and power plants to the data centers themselves. The last is Moody’s following through on its February statement that it may make a nonstandard adjustment to hyperscaler adjusted debt for likely lease cash outflows.
Karp saw where the labs are headed by following his own logic wherever it led, and I expect the lenders who financed the buildout to end up on the same road and to make the request under a different name.
Tyler Durden
Sat, 10/24/2026 – 16:30
