The largest investment-grade corporate debt issuance in U.S. history, the $27 billion Beignet Investor LLC jointly sponsored by Meta and Blue Owl Capital to build the Hyperion Data Center, is less notable for its size than for what its structure reveals about how risk is being allocated in what technology and financial professionals increasingly regard as a generational U.S. capital investment cycle.
The deal exposes a pattern likely to be replicated at scale as AI hyperscalers' data center demands continue to grow: risk — both credit and physical — that appears contained at the asset level resolves (under stress) to the corporate sponsor's balance sheet through a layered structure of off-balance-sheet accounting and project finance mechanics.
As this financing template scales across an estimated $8.2 trillion in projected U.S. data center investment through 2032, the aggregate exposure across finance, accounting, and insurance is beginning to concern even the most sophisticated market participants — raising fundamental questions about where risk ultimately resides.
Scale and Structural Shift
S&P Global Ratings' digital infrastructure team has evaluated more than $150 billion in data center transactions since the market's formation, with publicly rated debt exceeding $85 billion across corporate, multi-finance, and structured finance formats.
The team estimates $1.5 trillion in data center investment needs over the next five years, with total AI-spectrum requirements in the multi-trillion range.
The shift driving this volume is a deliberate migration of capital off hyperscaler balance sheets. The legacy model — direct corporate ownership financed with senior unsecured debt — is being replaced by build-to-lease joint ventures designed to achieve equity accounting treatment under GAAP.
The financial logic is straightforward: companies including Meta, Microsoft, Google, and Amazon carry low leverage and investment-grade credit ratings, and neither the leverage impact of large-scale direct financing nor the re-rating of their equity multiples toward infrastructure comparables is an acceptable outcome.
The financial motive behind the structural shift is straightforward.
Stijn Van Nieuwerburgh, professor of real estate and finance at Columbia Business School, argues that hyperscalers are managing a dual constraint: the speed demands of the AI buildout and the optics of what large-scale direct financing does to their equity story.
"These companies are slowly transforming from platform and software companies to essentially infrastructure firms," he said. "They don't like the PE multiples associated with infrastructure firms — and they don't like the credit rating impact of having a lot more debt. So the combination of needing access to a lot of projects as soon as possible, plus not wanting to do all of it on balance sheet, naturally leads them to finance a larger share of this off balance sheet."
The result is a deliberate arbitrage between two identities (technology company and infrastructure operator) that the current financing structures are engineered, in part, to sustain.
The scale of this structural migration is significant. Moody's has calculated that the major hyperscalers collectively carry approximately $970 billion in lease commitments, of which roughly $660 billion — approximately 68 percent — are not reflected on the balance sheet, sitting instead in footnotes to annual filings.
$660B of $970B in hyperscaler lease commitments sits off the balance sheet
| $ billion | Share | |
|---|---|---|
| Off balance sheet | $660B | 68% |
| On balance sheet | $310B | 32% |
| Total | $970B |
Source: Moody's via Van Nieuwerburgh, "Financing the AI Buildout" (2026)
The buildout trajectory Van Nieuwerburgh projects — 200 gigawatts of additional U.S. data center capacity between 2026 and 2032, implying roughly $8.2 trillion in total investment at approximately 2.8 percent of GDP annually — would exceed every prior U.S. infrastructure boom in relative terms, including the railroad buildout, rural electrification, the interstate highway system, and the telecom fiber expansion of the late 1990s.
The Beignet Structure: Leverage, Legal Architecture, and Implied Credit Support
The Beignet/Hyperion transaction is the clearest current expression of the new financing template and merits detailed examination.
The asset is a two-gigawatt AI training campus in northeastern Louisiana — remote location, significant property tax abatements, and three gigawatts of power contracted from Entergy Louisiana. At approximately three times the footprint of Central Park and with power consumption equivalent to Philadelphia, it is among the largest single privately financed infrastructure assets ever constructed in the United States.
The $29 billion construction cost was financed with $27 billion in 144A private placement bonds, the first publicly rated 144A issuance in the data center market. PIMCO purchased approximately two-thirds of the issuance. The bond was rated A+ by S&P Global — one notch below Meta's own corporate rating.
The capital structure is a joint venture: 80 percent Blue Owl, 20 percent Meta retained, with Meta serving as day-to-day asset manager subject to material consent rights negotiated by investors. The off-balance-sheet treatment required maintaining Meta's operational control while achieving GAAP deconsolidation — a balance structured through call rights, distribution mechanics (including a dividend pusher tied to Meta Platforms' own dividend payments), and a totality-of-circumstances test for equity accounting purposes.
The leverage implied by the structure is 90 percent — $27 billion in debt against a $30 billion asset. By comparison, Meta's own balance sheet carried approximately 25 percent leverage at year-end 2025. The S&P A+ rating reflects the contractual protections embedded in the deal documents: construction overrun responsibility absorbed by Meta above a defined threshold, ratings reaffirmation requirements, leverage restrictions at the JV level, and a residual value guarantee providing a minimum recovery floor in the event of non-renewal, early termination, or default.
Capital structure
Beignet carries 3.6× Meta's corporate leverage — 90% debt on a $30B asset
| Debt | Equity / other | |
|---|---|---|
| Beignet / Hyperion | 90% | 10% |
| Meta corporate | 25% | 75% |
Source: S&P Global Ratings; Van Nieuwerburgh (2026)
Van Nieuwerburgh's interpretation of the rating is direct: the A+ rating is essentially a restatement of Meta's own credit with 90 percent leverage layered on top. That reading is consistent with S&P's own analytical framework, which is transparent about the role of Meta's credit in the rating. The structural question it raises is what happens to the rating model when the template is applied to tenants with weaker balance sheets.
The lease structure also embeds optionality that is analytically relevant. Rather than a single 20-year master lease, the Hyperion campus uses 20 to 30 individual building-level leases with successive renewal periods. The residual value guarantee addresses the risk of non-renewal — but the guarantee is contingent on Meta's solvency at the time it is called.
"What's the state of the world where Meta says, four or eight years from now, we don't need half of this data center anymore? That's a state of the world where probably nobody else needs these data centers either. And is Meta still solvent? I don't know. Do I think it's possible that Meta is bankrupt eight years from now? I do. I'm not saying it's very likely, but I do think it's a non-trivial probability. Eight years is a lot of time in this new world of AI."
— Stijn Van Nieuwerburgh, Columbia Business School
The bond priced at 6.58 percent — approximately 100 basis points above where Meta could have issued unsecured corporate debt at the time. Over the bond's life, that premium translates to more than $5 billion in additional interest expense passed back to Meta through higher lease payments.
Complexity premium
Beignet priced ~100 bps above Meta unsecured — translating to $5B+ in extra interest over the bond's life
| Yield | |
|---|---|
| Beignet / Hyperion bond | 6.58% |
| Meta unsecured equivalent | ~5.58% |
| Spread | ~100 bps |
Source: S&P Global Ratings; Van Nieuwerburgh (2026)
Balance Sheet Opacity: The Probability Vacuum
The accounting treatment at the core of the Beignet structure — and the broader hyperscaler lease model — creates a disclosure gap that Van Nieuwerburgh terms a "probability vacuum."
Under current GAAP, the joint venture structure allows Meta to recognize neither the future lease payment obligations (which will accrue with certainty over the lease terms) nor the residual value guarantee (which will ultimately materialize either as a liability, if Meta exits, or as a contingent claim, if it does not). Both cannot simultaneously remain off the balance sheet: with certainty, one of the two must eventually appear. Current accounting rules permit both to remain in footnote disclosure rather than headline financials in the interim.
"Meta neither has to account for the future lease obligations nor for this residual value guarantee on balance sheet — even though with probability one, one of those two liabilities will have to come on balance sheet in the future. I call this a probability vacuum. It smells a lot like the regulatory arbitrage we had in the subprime mortgage crisis — where people were playing similar games of issuing loans, re-securitizing, buying the securitized loans back, keeping the same risk but with only half the capital charges."
— Stijn Van Nieuwerburgh, Columbia Business School
Dhaval Shah, director on S&P's digital infrastructure team, characterized the broader structural evolution as "blurring of credit lines": "We are seeing corporate transactions which allow for more investment in other data centers with project finance-type coverage and governance — or vice versa, where you have project finance transactions with corporate-like governance." In practice, this means that the analytical frameworks for both corporate and project finance apply imperfectly to these structures — they are hybrid instruments for which standard comparables are difficult to establish.
Physical Risk and the Insurance Capacity Constraint
The insurance market's response to large-scale data center assets provides an independent signal on the structural risk concentration identified in the financial analysis.
At the Beignet scale, full replacement cost coverage is not available from the P&C market. S&P's analytical approach used Probable Maximum Loss scenarios rather than full replacement cost — modeling the most severe plausible loss event rather than total asset value. Even at that threshold, the market could not absorb the required limits. Meta agreed to backstop the insurance shortfall with its own balance sheet, which S&P treated as a credit positive. The practical consequence is that the ring-fence around the project finance structure has a direct opening to Meta's corporate balance sheet — the same sponsor whose credit the A+ rating ultimately reflects.
"You don't get insurance coverage for $27 billion of debt. What we looked at is the probable maximum loss scenarios and whether there will be coverage to cover the probable maximum loss scenario — [it] was very conservative, a one-in-ten-thousand-year case. Insurance companies are not going to write that at all."
— Dhaval Shah, S&P Global Ratings
Swiss Re Institute's March 2026 analysis quantifies the systemic dimension of this capacity constraint. More than 25 percent of U.S. data center capacity sits in locations experiencing three or more large hail days annually. Approximately 40 percent is located in significant-to-very-high tornado exposure zones.
Physical hazard exposure
Share of U.S. data center capacity in high-risk natural hazard zones
| Hazard zone | Share of U.S. capacity |
|---|---|
| Significant-to-very-high tornado | ~40% |
| High hail (3+ days/year) | >25% |
Source: Swiss Re Institute, CatNet analysis, sigma insights (2026)
Fire — historically the highest-severity peril in data center loss history — accounts for 10.9 percent of loss events by frequency but 42.3 percent of total loss costs by value. That ratio is expected to worsen as lithium-ion battery backup units become standard in AI-era facilities, introducing a new ignition source into data center equipment rooms.
Fire risk — frequency vs. severity
Fire causes 11% of data center loss events but 42% of total loss costs
| Fire | All other causes | |
|---|---|---|
| Share of events | 10.9% | 89.1% |
| Share of loss costs | 42.3% | 57.7% |
Source: Swiss Re Institute sigma insights (2026)
Global data center insurance premiums are projected to grow from approximately $10.6 billion to $24.2 billion by 2030, per Swiss Re — a near-doubling in four years. The same analysis notes that the re/insurance industry can only support a fraction of the required limits at competitive rates, with the gap growing as asset values increase.
Insurance capacity
Global data center premiums nearly double by 2030 — capacity doesn't follow
| Global premiums | |
|---|---|
| 2026 | $10.6B |
| 2030 (projected) | $24.2B |
Source: Swiss Re Institute sigma insights (2026)
Joe Macejak , Managing Director, US Property Real Estate Strategy Leader at Marsh framed the pricing dynamic in terms directly relevant to institutional investors: "When risks are unable to be quantified with confidence, they become inefficient, or even difficult to transfer in the global P&C marketplace."
The conclusion from the combined financial and insurance picture is consistent: the project-level ring-fence in these structures is not closed. Under the stress scenarios most relevant to credit analysis — catastrophic physical loss, tenant non-renewal, or hyperscaler credit deterioration — the exposure resolves to the corporate sponsor. The insurance market's refusal to price the tail is the most direct market signal of that resolution.
Risk Distribution: Terminal Holders and Hidden Counterparties
The distribution of Beignet's $27 billion following PIMCO's initial purchase illustrates a broader dynamic in how this paper moves through the system. PIMCO acquired approximately two-thirds at issuance and subsequently sold positions to realize a gain. BlackRock was also an early buyer. The secondary market for this paper leads, in Van Nieuwerburgh's assessment, to a terminal holder base of bond mutual funds and pension funds.
"A lot of this ends up in large bond mutual funds — pension funds, and you and me buying these bond mutual funds. And it has a nice juicy yield. A good yield and a good story — there's a 12-mile line out the door for use cases we currently cannot meet. It's not that hard for investors to get behind that story right now. I don't think they're thinking about 20 years from now."
— Stijn Van Nieuwerburgh, Columbia Business School
The counterparty risk in this market also runs deeper than the rated tenant. Microsoft and Oracle, among others, are building significant data center capacity primarily to serve OpenAI and Anthropic — neither of which carries a credit rating and neither of which has demonstrated the cash flow generation to sustain large, long-duration lease commitments.
Van Nieuwerburgh raises a counterparty risk the public credit analysis of these deals largely ignores: the hyperscalers building out the most aggressively — Microsoft and Oracle in particular — are doing so primarily to serve demand from OpenAI and Anthropic, neither of which carries a credit rating. In his view, the hyperscalers are effectively lending their balance sheets and investment-grade ratings to AI model providers whose own economics remain unproven, and whose revenue is under direct pressure from token price competition. If that demand softens or those companies encounter financial stress, the exposure flows back up the chain to the rated tenant — and in Oracle's case, he argues, that exposure is large enough to matter.
Cycle Dynamics: Current Discipline and the Conditions for Its Erosion
Shah characterized the current market as having evolved from an investment-grade-only base to include speculative-grade issuance, GPU financing as an additional structural layer, and the bundling of power infrastructure financing with data center debt as campus-scale power demand outgrows available grid capacity. NIMBY opposition and permitting risk — previously passive assumptions in underwriting — have become active execution risks that S&P now treats as material.
The current cycle retains one structural safeguard that prior real estate and infrastructure bubbles lacked at equivalent stages: no large data center facility has yet been built without a signed long-term lease commitment in place. That discipline has, to date, constrained the speculative development risk that characterized the late stages of prior cycles.
"Every real estate bubble follows the same stages: rents are strong, that attracts development, and at some point it attracts speculative development with no tenants in place. Then you get overbuilding. Once it's overbuilt, rents fall, vacancies rise, bankruptcies follow. Every real estate boom in world history looks exactly like that. That might change. That usually changes."
— Stijn Van Nieuwerburgh, Columbia Business School
Van Nieuwerburgh identifies two specific warning indicators worth monitoring. The first is compression of the complexity premium — the spread above where simpler transactions clear. So long as investors require incremental compensation for the structural and counterparty opacity of these instruments, market discipline remains functional. The second is the acceptance of lower-credit or unrated tenants as anchor counterparties for large structured financings.
His assessment of where a disruptive event is most likely to originate does not center on the public investment-grade market, where credit rating discipline and equity market scrutiny provide ongoing oversight: "I think it's much more likely to come from the private markets — some private credit fund that blows up because it's overexposed to second-tier data center developers who built on spec with tenant demand that never materializes."
Implications for Life Insurers and Long-Duration Investors
Life insurers and other long-duration private credit investors are natural buyers of data center paper — contracted cash flows, long lease terms, and investment-grade ratings align with liability-matching strategies. The structural features designed to make these assets attractive to that investor base are, however, the same features that concentrate risk in tail scenarios.
A life insurer holding 20- or 30-year paper in this market faces exposure along multiple dimensions that standard credit analysis may not fully capture: contingent corporate balance sheet backstops in the physical risk insurance structure; accounting opacity in the GAAP treatment of lease obligations and residual value guarantees; shadow credit exposure to unrated AI model providers; and the potential for speculative oversupply in a market with no historical precedent at this scale or speed of development.
The complexity premium embedded in current pricing — the approximately 100 basis points above equivalent corporate unsecured rates visible in the Beignet bond — represents compensation for accepting that exposure.
The question for institutional investors is whether that premium accurately prices the tail, and what the market signal would look like if it began to compress.
The insurance market's refusal to absorb the physical tail on even a single $27 billion asset suggests the tail is not fully priced anywhere in the capital structure.
