How AI Data Centers Get Financed: Debt, Equity, and Neoclouds

AI data centers are financed mainly with debt raised outside the sponsor’s own balance sheet — through off-balance-sheet special purpose vehicles (SPVs), GPU-collateralized loans, record corporate bond sales, and asset-backed securities (ABS) — rather than through hyperscalers’ cash reserves or straightforward equity. In 2026, the five largest US hyperscalers issued roughly $159 billion in bonds, Morgan Stanley estimates the sector needs about $800 billion in private credit through 2028, and CoreWeave closed the first investment-grade GPU-backed loan. Neoclouds and hyperscalers use structurally different playbooks, and the difference matters for anyone buying, leasing, or investing in AI capacity.

Key takeaways

• Hyperscaler bond issuance hit ~$159 billion in 2026 across Alphabet, Amazon, Meta, Microsoft, and Oracle — up 47% year over year, per bond-market data compiled by Crypto Briefing.
• Off-balance-sheet SPVs are now the default structure for large campuses. Meta’s “Hyperion” joint venture with Blue Owl Capital raised $27 billion in A+-rated debt and $2.5 billion in equity, with Blue Owl owning 80% of the vehicle and Meta 20%.
• GPU-backed debt went investment-grade in 2026. CoreWeave’s $8.5 billion delayed-draw term loan, rated A3 by Moody’s, was the first GPU-collateralized financing to earn an investment-grade rating.
• The ABS market for data centers has grown 15x since 2020 — from about $4 billion outstanding to roughly $61 billion year-to-date in 2026, per Barclays Research.
• Morgan Stanley projects ~$800 billion of private-credit demand from 2025-2028 for AI data centers, power, and fiber combined.
• Neoclouds carry structurally different risk than hyperscalers. CoreWeave’s total debt reached $35 billion by mid-2026, backstopped partly by a $6.3 billion Nvidia commitment to buy unsold capacity through 2032.

For live pricing context on the capacity this capital is building, see the colocation price index and the global data center catalog.

Why data center financing looks different from five years ago

Data centers used to be financed like other commercial real estate: REITs and developers borrowed against stabilized leases at moderate leverage and raised equity on public markets. That model doesn’t scale to what AI campuses now cost. Individual gigawatt-class campuses run into the tens of billions of dollars, and combined data center lease commitments from Microsoft, Meta, and other hyperscalers topped $700 billion by early 2026, with total leasing exceeding $850 billion, according to Bloomberg and Benzinga reporting on hyperscaler capital plans. Putting that much capital directly on a single company’s balance sheet would hurt its credit rating and its return on invested capital — so hyperscalers, neoclouds, and their lenders have built a parallel financing system that keeps most of the debt project-level rather than corporate-level.

The capital stack: debt, equity, and where the money sits

Most large AI data center projects now separate into three or four financing layers rather than a single mortgage.

Layer	Typical role	2026 example
Senior secured debt / ABS	Bulk of project cost, rated by Moody’s, S&P, or DBRS	Meta-Blue Owl “Hyperion” JV: $27B in A+-rated debt
Sponsor / infrastructure-fund equity	Ownership stake, absorbs first losses	Same JV: $2.5B equity, anchored by PIMCO ($18B) and BlackRock ($3B)
Hyperscaler co-investment	Minority ownership, secures the anchor lease	Meta holds 20% of Hyperion; Blue Owl-managed funds hold 80%
GPU-collateralized loan (neoclouds)	Hardware-specific tranche, separate from the building	CoreWeave’s $8.5B investment-grade delayed-draw term loan, A3/Moody’s

The Hyperion structure illustrates the ratio that’s become common on hyperscaler-anchored campuses: roughly 90% debt to 10% equity at the SPV level, with the corporate sponsor holding a minority equity stake rather than funding the project directly.

Off-balance-sheet SPVs: the hyperscaler playbook

An SPV is a legal entity created to own one project. It borrows and raises equity against that project’s own contracted cash flow — typically a long-term lease from the hyperscaler sponsor — rather than against the sponsor’s corporate credit. Morgan Stanley arranged the Hyperion vehicle for Meta’s Louisiana campus; a separate BlackRock-led vehicle, backed by Global Infrastructure Partners (GIP) and HPS, uses the same 80/20 ownership split for another Meta-linked project. BlackRock separately launched a $12 billion-plus debt deal tied to Meta data center capacity. Because the SPV’s debt doesn’t appear on the sponsor’s own balance sheet, hyperscalers can keep expanding capacity without diluting shareholders or triggering the kind of leverage ratios that would concern corporate bond investors.

GPU-backed debt: neoclouds’ new asset class

Neoclouds — GPU-as-a-service providers without a hyperscaler’s balance sheet — have pioneered a different instrument: loans collateralized directly by Nvidia GPUs. CoreWeave first used Nvidia H100 processors as loan security in a facility led by Magnetar Capital and Blackstone; by March 2026 it had closed an $8.5 billion delayed-draw term loan rated A3 by Moody’s and A(low) by DBRS — the first GPU-backed financing to reach investment grade — priced at SOFR plus 2.25% and maturing in 2032. CoreWeave’s total debt reached $35 billion by June 30, 2026, up from $22.7 billion in long-term debt three months earlier. Other neoclouds followed: Fluidstack has raised more than $10 billion using Nvidia GPUs as collateral, Lambda secured a $500 million GPU-backed loan and a separate $1.5 billion GPU leasing deal with Nvidia, and Crusoe has taken on hundreds of millions in GPU-collateralized loans.

Nvidia itself underwrites part of the risk. Under a contract running through April 2032, Nvidia is obligated to buy up to $6.3 billion of CoreWeave’s unsold computing capacity if CoreWeave can’t find customers for it — a backstop that has drawn scrutiny for creating circular financing exposure between Nvidia, its chip customers, and their lenders.

Corporate bonds: hyperscalers still borrow directly

SPVs and GPU-backed loans supplement, rather than replace, ordinary corporate bond issuance. Alphabet, Amazon, Meta, Microsoft, and Oracle together sold roughly $159 billion in bonds in 2026, a 47% jump from 2025, largely earmarked for AI infrastructure.

Issuer	2026 bond issuance	Note
Amazon	~$57B	Largest single issuer of the five
Alphabet	~$52B
Meta	$30B corporate bonds, plus a separate $12.5B project bond	Project bond backs a Texas campus at a higher yield than a comparable 2025 deal
Oracle	$18B
Microsoft	Included in the $159B combined total	Company does not break out the figure separately

Meta’s project-specific bond illustrates a trend within the trend: even corporate-style bonds are increasingly tied to a specific campus rather than issued as generic unsecured debt, blurring the line with SPV financing.

Securitizing the leases: the asset-backed securities market

The newest layer is the ABS market, where data center operators bundle long-term hyperscaler leases into rated securities much like commercial mortgage-backed securities. DataBank raised $1.1 billion in September 2025 in what the industry counted as its first data-center ABS dual-rated by both S&P and Moody’s, establishing a hyperscale master trust structure other operators have since copied.

Year	Data center ABS market	Share of esoteric ABS market
2020	~$4B outstanding	~3%
2025	~$27B issued (ABS + CMBS combined)	—
2026 (YTD)	~$61B outstanding	~12%
2028 (forecast)	~$180B outstanding	~20%

JPMorgan projects $30-40 billion of annual data center securitization issuance in both 2026 and 2027 — 7-10% of combined ABS and CMBS issuance, up from about 27 billion dollars in 2025. The SEC has also loosened some securitization disclosure rules specifically to accommodate data center bond structures, which analysts read as regulatory acknowledgment that this asset class is now permanent rather than a one-off boom product.

Equity: infrastructure funds, sovereign capital, and JV structures

On the equity side, the same handful of managers keep showing up across deals: Blackstone, Blue Owl Capital, Apollo, PIMCO, and BlackRock originate most data center debt and equity, both through off-balance-sheet SPV transactions and direct lending to neoclouds and developers, according to J.P. Morgan’s capital markets research. Sovereign capital has entered directly rather than through fund vehicles: BlackRock’s roughly $40 billion agreement to acquire Aligned Data Centers brought in Microsoft, Nvidia, and MGX — the Abu Dhabi sovereign investment platform backed by Mubadala — as co-investors. Blackstone and Google have gone a step further with a joint venture that funds computing capacity itself, not just the buildings that house it, treating GPU fleets as an investable asset class alongside real estate.

Neoclouds vs hyperscalers: two different credit stories

Lenders underwrite these two borrower types very differently. Hyperscaler-anchored deals are easier to price because the tenant carries an investment-grade corporate rating and a multi-decade AI capex commitment. Neocloud-anchored deals are harder: the tenant is often sub-investment-grade, contract terms run shorter, and the main collateral is depreciating GPU hardware rather than a building. Lenders compensate with structural protections — parent guaranties, hyperscaler credit “wrappers” that effectively insure a portion of the lease, or underwriting based on a facility’s ability to be re-leased to a different tenant if the original one defaults — rather than relying purely on the neocloud’s own credit.

Risks: the “WeWork 2.0” concern

Some data center executives have started drawing an explicit comparison to WeWork’s collapse: long-term, capital-intensive leases signed with tenants whose own revenue depends on continued investor enthusiasm for a still-unproven business model. The parallel isn’t exact — GPU-as-a-service has real, metered revenue from day one, unlike WeWork’s subsidized occupancy — but the underlying risk is similar: if AI training and inference demand growth slows, neocloud tenants with thin balance sheets could struggle to service GPU-backed debt or renew data center leases, leaving lenders holding purpose-built capacity with a narrower pool of alternative tenants than a generic office building would have.

What this means for buyers and developers

If you’re evaluating a colocation or wholesale deal with a neocloud tenant, ask who is actually on the hook for the lease — the operating company, a parent guarantor, or an SPV with limited recourse — before treating a long-term contract as bankable. If you’re a developer or operator raising capital, the playbook now runs in parallel tracks: ABS and private credit for stabilized, hyperscaler-leased assets; GPU-backed loans for compute-heavy neocloud deployments; and SPV joint ventures for anything large enough to threaten balance sheet metrics on its own. Track the construction pipeline behind this capital at real estate & development, compare public and private investment vehicles at data center investment, and see how the debt-heavy structures discussed here show up in public company financials in our data center REIT comparison. Operators sourcing quotes on capacity backed by any of these structures can request pricing directly from facilities in the catalog.