Oracle's credit downgrade this month registered as a one-day story: shares gained 2.7% the day of the announcement, as investors focused on the company's $638 billion cloud backlog rather than the balance sheet strain behind it. But buried in S&P Global's rationale for cutting Oracle to BBB-, one notch above junk, was a detail that matters far more than the ticker movement: the agency named OpenAI, a company with no public financial disclosures and a business model still unproven at scale, as a "key credit risk" to one of the most systemically connected infrastructure financiers in the AI economy. Oracle now projects a fiscal 2027 free cash flow deficit of roughly $42 billion, nearly double its prior estimate, while carrying $167 billion in total debt.
That's a company-specific story. The structural story is bigger, and it's happening well outside Oracle's ticker. Across the data center industry, the debt used to build AI infrastructure is being pooled, sliced into risk tranches, rated, and sold off to pension funds, insurers, and asset managers — the same basic securitization machinery that turned a housing correction into a global financial crisis in 2008. The assets underneath this debt are not subprime mortgages. But the mechanics of how that risk gets packaged, rated, and distributed to investors who are several steps removed from the underlying tenant relationships deserve the same scrutiny that nobody applied to mortgage bonds until it was too late.
The scale of what's being financed
AI-related companies tapped debt markets for at least $200 billion in 2025 alone, a figure analysts consider a significant undercount because so many deals are private. Morgan Stanley projects hyperscalers — Microsoft, Meta, Amazon, Alphabet, Oracle, and Apple among them — will issue $250 billion to $300 billion in debt in 2026 to fund computing capacity, pushing the overall investment-grade bond market toward record volumes. Layered on top of straight corporate borrowing is a fast-growing securitization market specifically built around data centers: JPMorgan projects annual data center securitization issuance, across both asset-backed securities and commercial mortgage-backed securities, could reach $30 billion to $40 billion in 2026 and 2027, up from about $27 billion in 2025. By a separate estimate from Charles River Associates, tracking a narrower slice of the same market, data center asset-backed issuance alone has grown from $2.6 billion in 2020 to more than $14 billion annually — illustrating the same trajectory even if the two firms aren't measuring identical baskets of debt. Morgan Stanley separately estimates roughly $130 billion in cumulative U.S. data-center securitized-credit issuance across 2026 through 2028.
The mechanics work like this: a data center operator transfers real estate and tenant lease receivables into a bankruptcy-remote special purpose vehicle, which then issues rated notes to investors in tranches of varying seniority. Because the SPV is a legally distinct entity, the debt doesn't appear on the parent company's balance sheet — letting hyperscalers and cloud providers look less leveraged than they actually are while preserving room for additional borrowing elsewhere. Lenders who originate the underlying loans increasingly pool and resell them too, extending the chain of exposure from the technology company, through the SPV, through the securitization vehicle, and ultimately into the pension fund or insurance portfolio holding the rated tranche.
Data centers now account for the majority of a market approaching $80 billion across both securitization channels combined, and data-center CMBS issuance alone hit a record in 2025. S&P Global Ratings' sector lead for esoteric ABS ratings has said publicly that issuance is likely to keep growing as the construction pipeline expands, and insurers have told trade publications that risk-adjusted returns in the space have been attractive enough to justify deeper allocation.
Why this isn't quite 2008 — and why the differences matter less than they sound
The people building and rating these structures make a consistent argument for why data center securitization is fundamentally safer than mortgage-backed securities were in 2007: the tenants are overwhelmingly investment-grade hyperscalers with enormous cash reserves, the underlying demand for AI compute is real and growing rather than speculative, and the physical assets have genuine ongoing utility that vacant subdivisions never had. All of that is true, and it's the reason rating agencies are comfortable stamping investment-grade ratings on securities backed by an industry that didn't meaningfully exist five years ago.
But the argument skips over the two variables that actually determine whether these structures hold up under stress: tenant concentration and asset lifespan mismatch. Oracle's situation is the clearest illustration of the first problem. Roughly half of Oracle's remaining performance obligations run through a single counterparty, OpenAI, whose ability to pay depends on an eventual IPO that hasn't happened and profitability that hasn't materialized. Moody's has described the underlying Stargate buildout as effectively one of the world's largest project financings — except without the ring-fencing and structural protections that traditional project finance normally provides, because Oracle, not a dedicated project vehicle, is the one carrying the corporate credit risk. Oracle's cost of insuring its own debt against default has climbed steadily over the past year; its five-year credit default swap spread hit roughly 198 basis points in April, the highest level on record for the company, and a Quinn Emanuel legal analysis of the sector separately pegged Oracle's CDS spread increase at around 310% over the prior year, pushing its perceived credit risk to a sixteen-year high. Bondholders led by the Ohio Carpenters' Pension Plan have already sued Oracle in New York, alleging it understated how much additional debt it planned to raise when it sold $18 billion in notes just weeks after announcing the OpenAI contract.
The second problem is more structural and less company-specific: a GPU-collateralized lending market has emerged alongside traditional data center securitization, with CoreWeave pioneering an $8.5 billion investment-grade-rated deal backed by chips that cost $30,000 to $40,000 each. GPUs depreciate on a roughly seven-year technology cycle. The buildings that house them are financed on 20-to-30-year assumptions, the same duration horizon insurers and pension funds want to match against long-dated liabilities. That mismatch — sometimes called the "GPU debt treadmill" — means a facility's most economically important asset can become technologically obsolete years before the financing built around the building itself matures, and rating agencies acknowledge that master trusts expanding to accommodate new deal flow raise questions about asset-quality creep across pooled structures.
Who actually holds the risk
The investors buying into this market are, by design, the same institutions that were buying tranched mortgage risk in 2006: insurance companies and pension funds seeking long-duration, investment-grade-rated paper to match their liabilities. One data center CFO has described the appeal to institutional buyers bluntly, calling the sector the "picks and shovels" of the AI economy — a diversified, structured way to gain AI exposure without picking which hyperscaler wins. That framing isn't wrong, but it also describes exactly the kind of instrument that made 2008 possible: a way for capital to gain exposure to a boom without underwriting the specific, concentrated risks sitting inside any individual deal.
Single data center campus valuations have reached $10 billion to $20 billion — a size that industry insurers say was essentially uninsurable at that scale as recently as three years ago. Private credit funds including Blackstone, Apollo, and Ares have pushed outstanding loans to AI-related firms past $200 billion already, with some projections putting that figure at $300 billion to $600 billion by 2030. None of this means data centers are worthless the way subdivisions built on defaulted subprime loans were worthless. Compute demand is real, and hyperscaler tenants are, for now, genuinely creditworthy. But the risk being underwritten by pension funds and insurers isn't really "data centers." It's the assumption that a small number of AI labs — several of them, like OpenAI, without public financials — will generate enough durable revenue to service leases stretching 15 to 19 years, on hardware that will need replacing three or four times over that span, structured through vehicles specifically designed to keep the exposure off the parent company's own balance sheet.
That's not a bet on infrastructure. It's a bet on a handful of unproven companies' income statements, wrapped in enough structural layers that most of the investors ultimately holding the risk will never see the underlying tenant relationship directly. Wall Street built that exact kind of distance once before, rated it investment grade, and then discovered — all at once — how much the rating had depended on assumptions nobody was stress-testing in real time.
