Executive summary. The AI buildout is no longer funded the way Silicon Valley traditionally funds things. Hyperscalers are offloading capital-intensive infrastructure onto special-purpose vehicles (SPVs), long-term leases, and structured debt markets, turning data centre campuses into quasi-infrastructure assets priced like bonds. Semiconductor foundries are doing something adjacent but structurally different: selling minority equity stakes to private capital under joint venture (JV) programmes backstopped by sovereign grants. At the compute layer, the hardware itself, the graphics processing units (GPUs) sitting in racks, has become collateral for a new class of asset-backed lending. Each financing model reflects the nature of what sits at the bottom of the collateral chain. Real estate cash flows back bonds. Fab equity backs strategic co-investments. Silicon backs equipment loans. The common thread is that the era of funding AI entirely from a corporate treasurer's desk is over, and credit markets are now pricing what AI infrastructure is actually worth when you strip the hype and test the contracts.

For most of computing history, a hyperscaler wanted a data centre and built it with money from its own treasury. The balance sheet absorbed the cost, the asset sat on the books, and the capex showed up as depreciation for the next decade. Simple. Boring. Sustainable when individual campuses cost a few hundred million dollars.

A single AI training campus today costs several billion dollars, pulls a gigawatt of power, and takes years to construct. Multiply that across the pipeline of projects at all the hyperscalers and the numbers exceed what any corporate treasurer can absorb without alarming shareholders about leverage ratios. Something had to give.

What gave was the ownership structure. Developers and private equity firms now build these campuses inside SPVs, ring-fenced legal entities that own just the asset and nothing else. The hyperscaler signs a long-term lease, often fifteen to twenty years, structured as triple-net, meaning the tenant pays taxes, insurance, and maintenance on top of rent. The SPV then raises debt against those contracted lease revenues. Because the hyperscaler's creditworthiness drives the repayment story, the debt can price near investment grade even when the developer behind the SPV would never qualify for that rating on its own.

The linguistic tell is in how these deals get described. Oracle does not issue the bonds on the Michigan campus. Oracle is the tenant. The bond prospectus says Oracle-backed, which means contractually committed, not corporately guaranteed. That distinction carries a spread.

The AI capital stack splinters into three distinct financing models depending on what sits at the base of the collateral chain. Each model has its own logic, its own investor base, and its own failure mode.

Data centres run on lease cash flows. Bank of America led the ~$16.3 billion capital stack for Oracle's Saline Township campus in Michigan, with roughly $14 billion structured as bonds in a 144A private placement (explained below). PIMCO anchored the deal before seeking to syndicate portions to other institutions. Blackstone contributed equity. Pricing discussions sat near 7.5%, tied to Oracle's own corporate yields plus a premium for project-level risks. The debt is secured by the project assets and the long-term lease revenues the campus generates, with Oracle's capacity commitment to OpenAI workloads providing the underlying demand anchor. Construction risk, power delivery timelines, and single-tenant concentration are what drove PIMCO's syndication effort: the firm hit its concentration limits and needed other buyers to absorb the exposure.

The Meta Hyperion and Beignet joint venture in Louisiana is the benchmark that defines how large this market can get. Meta sold 80% of the $30+ billion campus project to Blue Owl Capital for approximately $2.5 billion in equity. The SPV then issued $27.3 billion in 144A investment-grade bonds, rated A+ by S&P (Standard and Poor's), one notch below Meta's own corporate rating, with a 6.58% coupon maturing in 2049, co-terminus with the lease. PIMCO bought nearly $18 billion of that paper. The 120 basis points spread over Meta's equivalent corporate bond is the market's precise answer to the question of what single-tenant concentration risk costs when the asset is a gigawatt-scale campus in the Louisiana bayou.

Hut 8's River Bend campus takes the structure further by showing how a company with no investment-grade rating of its own can access institutional debt markets at near-investment-grade pricing. Hut 8, historically a Bitcoin miner pivoting into AI and high-performance computing (HPC), signed a 15-year triple-net lease with Fluidstack across 245 megawatts of capacity. Google then provided a full financial backstop on the lease payments, meaning bondholders are effectively underwriting Alphabet's ability to pay rent, rated AA+/Aa2, rather than Hut 8's ability to run a company. Expected net operating income (NOI) sits at approximately $454 million annually, rising 3% each year, with a $7.0 billion base value and up to $17.7 billion across three 5-year renewal options. That contractual certainty allowed Hut 8's subsidiary to prepare at least $3 billion in high-grade bonds with up to 85% loan-to-cost project financing underwritten by J.P. Morgan and Goldman Sachs. A Bitcoin miner's cash flows became infrastructure receivables by one stroke of a lease agreement and a backstop letter from one of the world's safest corporate credits.

Semiconductor foundries run on equity, not leases. A data centre is, at its core, a very expensive shed with power and cooling. If the tenant walks, you can in theory find another tenant. A semiconductor fabrication plant (fab) is something else entirely. The equipment inside costs billions to procure, takes years to calibrate, and produces chips at tolerances measured in nanometres. You cannot replace the operator of a 2-nanometre fab mid-construction the way you might replace a building manager. That operational complexity makes the lease-backed bond model largely unworkable for foundries and pushed the industry toward a different structure: strategic equity co-investment.

Intel pioneered what the industry now calls the Semiconductor Co-Investment Program (SCIP). Rather than borrowing money to build a fab, Intel sold a minority equity stake, typically 49%, in the physical facility itself to a private capital partner. Apollo Global Management invested $11 billion for a 49% stake in Intel's Fab 34 in Ireland, functioning as bridge equity. Intel received the cash to sustain construction during a period of intense capex pressure, and once the asset de-risked and became operational, Intel exercised a repurchase right in 2026 to reclaim full control. The investor got returns commensurate with greenfield industrial equity risk. Intel got the capital without diluting its corporate equity or blowing through its debt covenants.

Unlike data centres, foundries also carry a sovereign financing layer. The CHIPS and Science Act of 2022 (CHIPS Act) provided $7.86 billion in grants to Intel and $6.6 billion to TSMC's Arizona operations. These government grants function as first-loss capital, absorbing initial project risk and making the private equity co-investment more attractive by improving the overall risk-return profile. Revenue in a foundry is spread across many customers on shorter contract cycles, which makes it far less amenable to the long-dated lease securitisation that powers data centre bonds. The financing therefore looks more like traditional heavy industry than infrastructure project finance.

Compute infrastructure runs on the silicon itself. The newest and most structurally interesting financing model treats the hardware as the collateral. CoreWeave's $8.5 billion debt facility is the flagship example: lenders took a lien on the company's inventory of Nvidia H100 and B200 GPUs, along with the associated contracts with AI labs. The logic holds as long as GPUs have a liquid secondary market and the compute contracts underpinning demand remain enforceable. GPU-backed financing is equipment leasing scaled to data-centre density, with the added complexity that GPU generations turn over faster than the buildings housing them.

In the high-bandwidth memory (HBM) segment, the financing model takes a further turn. SK Hynix and Micron are reportedly receiving substantial prepayments from Nvidia to lock in future production capacity. Here the customer functions as the bank, providing upfront capital in exchange for a priority claim on output. The manufacturer receives the funds needed to expand production lines without accessing public debt markets. The customer secures supply in a constrained market. Both parties are hedging against scarcity, and the transaction doubles as working capital financing for the supplier.

Beyond the headline deal types, the AI capital stack has assembled a full arsenal of instruments, each occupying a different position in the risk-return hierarchy and serving a different investor constituency.

SPV project debt, issued as 144A bonds or syndicated bank loans, is the workhorse. Leverage ratios of 70 to 90% debt-to-cost are common, justified by the long-dated, contracted nature of the lease revenues. These deals are typically non-recourse, amortising over the lease term, and structured so that the debt matures alongside the lease rather than forcing a refinancing at an unpredictable future spread.

Private credit and structured equity fills the capital structure between senior debt and the thin equity layer retained by the developer. Firms like Blue Owl, PIMCO, Blackstone, and Apollo frequently take equity stakes in the JV and then layer senior debt on top, capturing the spread between the hyperscaler-quality revenues flowing into the SPV and the private credit cost of the debt sitting above them. This carry trade is the core value proposition for alternative asset managers in this space.

Asset-backed securities (ABS) and commercial mortgage-backed securities (CMBS) aggregate portfolios of stabilised data-centre leases or mortgage loans into rated tranches, allowing insurance companies and pension funds to buy slices calibrated to their own risk appetite. Variable funding notes within master trust structures allow the portfolio to grow over time. Issuance in this format has been doubling as traditional bank lending hits concentration limits and institutional demand for infrastructure-adjacent yield outpaces the supply of corporate bonds.

Synthetic and off-balance-sheet leases allow hyperscalers to retain the economic benefit of computing capacity, guaranteed access to compute resources, without taking on the full balance-sheet liability of owning the assets. This is the mechanism that lets Meta build a $30 billion campus while reporting a much smaller capex figure than the physical footprint implies.

The spread premium in these deals, 120 basis points over corporate equivalents in the Meta Hyperion bonds, pricing near 7.5% on the Oracle Michigan debt, reflects a risk set that institutional investors have itemised carefully.

Construction and power delivery are the first and most concrete exposures. Gigawatt-scale campuses in constrained power nodes take years to build. Grid interconnection queue positions can slip. Transformer lead times now run 128 to 144 weeks in constrained markets. Revenue commencement dates become uncertain, and debt service coverage during the construction period must be supported by equity reserves or delayed-draw structures. A six-month construction delay on a $14 billion debt deal has compounding consequences that the spread must compensate for upfront.

Technology obsolescence is the second and harder to model. The inference and training architectures that justify a 2026 data centre build may look materially different by 2031. A hyperscaler tenant whose unit economics shift substantially due to model efficiency improvements has contractual incentives to renegotiate lease terms at renewal. Lease structures and step-in rights exist precisely to manage this, but the risk cannot be fully contracted away, and bond investors know it.

Single-tenant concentration is the third and most structurally acute. A $14 billion bond secured against one campus in Michigan is a bet on one tenant, one geography, and one technology cycle. If OpenAI's compute needs migrate to a different architecture or a different region, Saline Township becomes expensive real estate looking for a replacement tenant. Covenant packages attempt to address this, but the fundamental concentration risk lives in the spread.

AI demand normalisation is the final and most philosophically interesting risk. The structured credit market is being asked to price 15-year paper on the assumption that AI compute demand remains robust enough to justify these assets for their full term. Rating agencies can model lease covenants. They cannot model whether inference demand distributes across thousands of open-source deployments or consolidates into five hyperscale providers, or whether the next generation of models runs on a fraction of today's compute at comparable performance. That uncertainty is in the price, expressed as basis points above where a simpler, more diversified asset would clear.

The long book sits with developers and private credit funds that secure hard, long-dated hyperscaler contracts and access institutional debt markets at reasonable cost. Tier-1 power, land, and cooling assets with investment-grade lease coverage behave like quasi-infrastructure: yields are infrastructure-adjacent, but entry multiples still reflect project-finance risk rather than stabilised real estate. The carry between hyperscaler-quality revenues flowing into the SPV and private credit costs sitting above them is the core trade for Blue Owl, Apollo, and Blackstone in this cycle.

The short book sits with over-levered pure-play developers whose returns are fragile to construction delays, capital cost increases, or any deterioration in tenant creditworthiness. It also sits with hyperscalers whose balance sheets absorb too much undiversified project exposure, forgoing the structural benefits that off-balance-sheet financing now makes available at scale.

The deeper signal is structural. AI is a physical economy, and the software dominance narrative of the next decade is increasingly underwritten by the credit structures that let capital markets price the difference between a capacity commitment and an enforceable cash flow. The $14 billion Oracle deal, the $27 billion Meta bond, and the Google-backstopped River Bend financing are each individual acts of price discovery, the market finding, one tranche at a time, what AI infrastructure is worth when you strip away the corporate name and test the contracts against a covenant desk.

AI capex moved from corporate treasury to structured credit. The credit market is responding the way it always does: with appetite, discipline, and a spread.