Who Is Financing the AI Buildout?

Everyone can see the spending.

The less obvious question is who is actually supplying the money.

AI infrastructure is expensive in several different ways at once. Companies need land, power, cooling, data-center shells, networking equipment and GPUs. Some of those assets sit on a hyperscaler's balance sheet. Others sit inside a joint venture, a dedicated financing vehicle or a specialist cloud operator. Customers can prepay. Investors can provide equity. Lenders can fund the hardware. Suppliers can guarantee lease or power obligations.

So when somebody says the AI buildout will require hundreds of billions of dollars, that number does not describe one financing market. It describes a chain of balance sheets.

The useful question is not simply how much is being spent?

It is: which balance sheet pays first, what cash flow is supposed to repay it, and who is left holding the risk if demand, construction or technology does not behave as expected?

The Money Map

There are at least six distinct capital channels visible in the primary evidence.

1. Operating cash flow → hyperscaler infrastructure. Cash-rich companies can fund a large share internally.
2. Corporate debt and equity → operating company. Public capital markets add financing capacity, although general corporate debt should not automatically be labelled "AI debt."
3. Infrastructure equity + project/venture debt → data-center shell, power and cooling. A separate legal vehicle can own the physical campus while the technology company leases capacity from it.
4. Asset-level debt → GPUs and related compute infrastructure. Lenders can finance hardware through dedicated facilities tied to specific assets and customer contracts.
5. Customer prepayments or customer-supplied GPUs → operator or asset directly. The customer can carry part of the upfront capital burden.
6. Strategic guarantees and commitments → credit support, not funded cash. A supplier can make a project easier to finance without itself writing the initial check.

Those channels can coexist on the same buildout. Oracle, for example, has raised large amounts of debt and equity while also receiving customer prepayments and using structures in which customers prepay for or supply GPUs. The important point is not to add every headline number together. It is that capital channels stack rather than substitute for one another.

Meet the players

The hyperscaler or AI operator decides what infrastructure it needs and may fund part of it from operating cash.

The infrastructure investor can own the land, shell, power and cooling layer.

Bond investors, banks and private lenders provide debt against different balance sheets and assets.

The customer is more than a source of revenue. A prepayment, take-or-pay contract or customer-supplied GPU can directly change how much outside capital the operator needs.

The chip or infrastructure supplier can become a financing participant when it invests equity, commits to buy cloud capacity or guarantees lease and power obligations.

The same physical AI workload can therefore involve several separate capital providers with different legal claims and very different risk.

How the machine works

Meta's Hyperion data-center campus is a useful example because it separates the physical infrastructure from the operating company.

Meta disclosed a joint venture with an affiliate of funds managed by Blue Owl Capital to co-develop the Louisiana campus. Meta contributed $4.3 billion of land and construction-in-progress assets; the investor contributed $7.0 billion in cash. Blue Owl-managed funds hold 80% of the venture and Meta 20%. Meta said the parties would fund their pro-rata share of roughly $27 billion of development costs covering the buildings and long-lived power, cooling and connectivity infrastructure.

That already answers part of "who is financing it?" Blue Owl-managed funds are providing most of the equity capital.

But the stack goes one level deeper. Meta also said a portion of Blue Owl's capital would be funded by debt issued to PIMCO and select other bond investors through a private securities offering. Meta will lease the completed facilities and provided the venture with a capped residual-value guarantee under specified conditions.

One campus. At least four economically distinct positions: Meta as user and minority owner; Blue Owl-managed funds as majority equity investor; bond investors as debt capital; and Meta again as lessee and contingent guarantor.

That is what a chain of balance sheets looks like.

The compute layer can be financed differently

Now move from the building to the GPUs.

CoreWeave says it primarily finances infrastructure development through asset-level debt supported by take-or-pay customer contracts, supplemented by corporate equity and debt. In March 2026, it closed an $8.5 billion delayed-draw term-loan facility tied to high-performance-computing infrastructure and an associated customer contract. Blackstone Credit & Insurance anchored the facility; MUFG and Morgan Stanley acted as co-structuring agents and bookrunners, with other institutions, asset managers and insurance investors participating.

The financing can therefore sit closer to the asset and its contracted cash flow than a normal unsecured corporate bond.

IREN provides another window into the same idea. It disclosed a $3.6 billion investment-grade GPU financing for a Microsoft contract at 6%, and a separate $2.4 billion financing led by Blue Owl and PIMCO at 9% for non-investment-grade customer deployments. The company said the latter funded 90% of associated GPU capex.

The useful lesson is not that GPUs have a universal borrowing rate. They do not.

The lesson is that the economics of the customer sitting behind the compute can materially affect the economics of the financing.

Follow the Money

The money path can now be followed more precisely.

Infrastructure investor / bond investor → infrastructure venture → land, shell, power and cooling.

AI operator / asset-level lender → GPUs and compute equipment.

Customer → prepayment or take-or-pay commitment → operator / financed vehicle.

Customer usage and contracted payments → operating cash flow → interest, principal, lease payments and investor returns.

Strategic supplier guarantee → contingent support to landlord or power counterparty. That last arrow is not cash. It is a promise to absorb specified losses if another party fails.

NVIDIA makes that distinction unusually visible. It disclosed land, power and shell guarantees for select AI-cloud partners and, separately, guarantees tied to a large land, power and shell buildout for OpenAI-related leases at SB Energy's Ohio campus. Those guarantees are contingent obligations that become effective with lease commencement and decline under specified conditions as obligations are performed; they are not the same thing as NVIDIA funding the project upfront.

That is why financing totals can be misleading. A dollar of funded debt, a dollar of equity, a dollar of customer prepayment and a dollar of guarantee exposure are not four interchangeable dollars.

How the price gets set

The financing rate is not simply the price of the hardware.

A GPU has resale value. A data-center shell has physical value. But lenders care about the cash flow expected to repay them.

CoreWeave's model explicitly links asset-level debt to take-or-pay customer contracts. IREN's disclosed facilities also show materially different financing costs across customer-credit profiles.

So the marginal actor setting the price of capital may be looking through the machine to the customer at the other end.

That is a subtle but important point: AI infrastructure finance is credit underwriting wrapped around technology assets.

The money can travel in a circle

The AI financing system is not always a straight line from investor to borrower to data center. Some of the same companies can appear on multiple sides of the transaction.

A chip supplier or strategic investor can provide equity, guarantees, cloud commitments or other support to an AI company or infrastructure provider; that company can then use capital to buy GPUs, lease compute or build capacity that ultimately generates revenue for companies elsewhere in the same ecosystem. A customer can also prepay for capacity, giving an operator capital to build the infrastructure that will later serve that same customer.

Those arrangements are not inherently improper or economically artificial. They can solve a real financing problem: expensive infrastructure has to be funded before all of the revenue arrives.

But they make the headline numbers harder to interpret. A dollar of announced investment, a dollar of customer commitment and a dollar of infrastructure spending may be connected parts of the same economic loop rather than three independent sources of demand.

When evaluating the AI buildout, ask how much genuinely external demand and capital sits behind the system — and how much activity is being reinforced by overlapping commercial, financing and strategic relationships.

Where the risk sits

The financing structure determines where the risk lands.

Construction risk: who absorbs delays and cost overruns before the facility is usable?

Utilization risk: who suffers if the GPUs or data-center capacity are not fully used?

Customer-credit risk: what happens if the customer behind the contracted cash flow cannot pay?

Power risk: who has committed to the power infrastructure or guaranteed the lease and power obligations?

Technology-obsolescence risk: who owns the GPUs if a newer generation makes them less valuable sooner than expected?

Refinancing risk: does the debt mature before the economic life of the asset or contract?

The answer can differ at every layer.

A customer that supplies its own GPUs takes more hardware-obsolescence risk. A data-center investor can own the shell while the technology company remains economically tied through a lease. A supplier guarantee can shift part of tenant-credit risk back to the supplier. A lender can be protected by both assets and contracted customer cash flows.

There is no single place called "AI risk."

What can break the machine

The financing works best when three clocks line up: the time required to build the infrastructure; the contractual life of the customer demand; the maturity and amortization schedule of the financing.

If construction is late, contracted revenue may arrive later than expected. If customer demand weakens, utilization and refinancing assumptions can deteriorate. If equipment becomes obsolete quickly, the asset backing the financing may be worth less than expected. If power infrastructure is delayed, the GPUs can exist without productive capacity.

The risk is not merely that AI demand disappoints. It is that a long-lived financing structure was built around a demand, technology or credit assumption that changes faster than the liability does.

The part nobody explains

The customer contract can be as important to the financing as the physical asset.

The headline asset may be a GPU cluster or a giant data center. But a lender does not get repaid because the building exists. It gets repaid because someone is expected to pay for using it.

That is why take-or-pay contracts, prepayments, leases and guarantees matter so much. They convert an expectation of AI demand into something closer to an underwritable cash-flow claim.

And they can move risk before a single token is processed.

A customer that prepays is financing part of the build. A customer that supplies the GPUs absorbs part of the hardware risk. A customer with stronger credit can support cheaper financing. A supplier that guarantees lease obligations can absorb contingent tenant risk. An infrastructure investor can own the shell while the AI company retains economic exposure through rent and guarantees.

The financing is therefore not just behind the AI buildout.

It is part of the architecture of the AI buildout.

Why it matters

If AI infrastructure were funded only by the largest technology companies' cash balances, the story would mostly be about capex discipline.

It is becoming broader than that.

The capital burden is spreading across public bond markets, infrastructure funds, private-credit investors, insurance capital, customers and strategic suppliers. That can increase the amount of infrastructure the system is capable of building. It also distributes the downside across more balance sheets.

For investors, that changes what to watch.

Do not look only at announced capex. Look at who owns the asset, who funded it, what contract supports the financing, what guarantee sits behind it and which party is still exposed if the economics disappoint.

Because the company with the best model is not necessarily the company supplying the most capital.

And the company using the infrastructure is not necessarily the one ultimately carrying the largest financing risk.