Data Center Financing at 1–10 MW: What Changes Below Hyperscale

Summary

• Data center financing is two stacks: construction loans, project finance, securitized bonds, joint ventures and funds for the building; equipment leases or offtake-backed loans for GPUs.
• JLL forecasts shell and core at $11.3 million per MW for 2026; AI fit-out can reach $25 million per MW and refreshes in roughly five years.
• At 1–10 MW, deployments go into existing colocation, so the financing question is mostly the GPU layer, sized to the offtake contract.

Introduction

Most coverage of data center financing is written for hyperscale campuses and gigawatt projects. Most operators reading it host or own 1–10 MW of GPUs, where the open question is how to pay for the servers.

The size of the market explains where the attention goes. Dell'Oro raised its 2026 global data center capex outlook to more than $1 trillion in June 2026. Goldman Sachs Research analysts expect leading tech companies to spend a combined $5.3 trillion from 2025 through 2030 (June 2026 forecast).

This article explains how each layer of data center financing works, what changes at 1–10 MW and which risks lenders price. It leaves the steps of a financing application to a separate guide.

How Data Center Financing Works: Two Layers, Two Capital Stacks

Data center financing is the debt and equity that pay for a data center. It splits into two layers financed separately: the land, shell and power, and the IT equipment installed inside, such as GPUs, servers and networking. The layers have different owners, refresh cycles and repayment sources, so they draw different lenders.

JLL forecasts an average global shell-and-core cost of $11.3 million per MW for 2026, excluding land and IT equipment. Shell and core is the building with its base power and cooling systems, before any tenant equipment goes in. The tech fit-out on top "can cost as much as $25 million per MW" for AI infrastructure, according to JLL. That figure is a ceiling with a different scope, so adding the two doesn't produce an all-in cost.

JLL also finds that tenants typically own the GPUs, CPUs and networking, upgraded on a roughly five-year schedule separate from the shell. The building is long-lived and refinanced once leased, while the IT layer is repaid from the compute it sells.

How the Building and Power Layer Is Financed

Project finance accounts for 58.3% of rated North American data center debt, the largest share, according to S&P Global Ratings in October 2026. Of the $105.3 billion issued by the issuers S&P rates, asset-backed securities made up 26.4% and corporate debt 15.3%.

Project finance is debt held in an entity created for one project and repaid from that project's cash flows. The layer follows the life of a facility: construction debt during the build, then long-dated refinancing once it is leased and stabilized.

Construction loans and project finance

Construction loans for the building layer are led by banks and sized to specific buildings and megawatts. In April 2026, First Citizens Bank was administrative agent, joint lead arranger and joint bookrunner on a $525 million facility for T5 Data Centers. The facility funds a new 36 MW data center in Chicago and the expansion of T5@Charlotte II to 28 MW.

DataBank closed a $650 million upsize to its Red Oak, Texas construction financing in June 2026 ($400 million bank financing, $250 million private notes). The upsize took the facility from $2.0 billion to $2.65 billion for a fourth 60 MW building.

ABS, CMBS and refinancing stabilized sites

Once a facility is leased and stabilized, operators refinance into securities repaid from contracted colocation revenue. Asset-backed securities (ABS) are bonds repaid from a pool of contracts. Commercial mortgage-backed securities (CMBS) are bonds secured by mortgages on the property.

KBRA assigned preliminary ratings to DataBank Series 2026-1 in January 2026. In October 2026, KBRA assigned ratings to DataBank Series 2026-2, 2026-3 and 2026-4. These notes pay investors from colocation contracts in place, so the length of those contracts sets the revenue investors can count on. That question of contract tenor returns in the GPU layer.

Corporate bonds, joint ventures and infrastructure funds

At hyperscale, buildings are financed with corporate bonds, joint ventures and infrastructure funds, tools built around investment-grade tenants on long leases. BIS reported in March 2026 that hyperscalers' gross bond issuance rose markedly, "topping $100 billion in 2025." In a common structure BIS describes, a joint venture or special purpose entity develops the campus. That keeps "most of the associated debt off the hyperscaler's balance sheet."

Meta's Hyperion joint venture, announced in October 2025, has about $27 billion in expected development costs. Funds managed by Blue Owl Capital own 80% and Meta 20%, and Meta leases the campus on a four-year initial term with options to extend. Meta also provides a residual value guarantee for the first 16 years, a capped commitment covering a shortfall in the campus's value.

Goldman Sachs Research reports that private infrastructure funds raised "a record $221 billion in 2025," a figure covering all infrastructure. Alphabet kept its full-year 2026 capex guidance at $195–205 billion in October 2026. A 1–10 MW operator rarely has an investment-grade tenant or a long lease, so these structures set the benchmark at the top of the market.

How the GPU Layer Is Financed Separately

The GPU layer is financed as equipment, against the hardware and the offtake contract that pays for it. An offtake contract is the agreement under which a customer buys the operator's compute. Its length and the customer's credit drive the loan, as our guide to how offtake agreements work explains.

The largest GPU deals show this layer financed apart from any building. In October 2026, SpaceX secured a binding commitment for an approximately $40 billion master finance lease for GPUs and related equipment, not yet closed. The Korea Herald reported in November 2026 that KB Kookmin Bank won a mandate to lead a 10.7 trillion won (about $8 billion) syndicated loan. The borrower, a Brookfield special purpose company, will buy GPUs to supply Naver Cloud under a long-term agreement, within a project of about 13 trillion won.

CoreWeave has financed its GPUs with GPU-backed delayed-draw term loans, which are drawn in stages after closing. Its August 2026 facility carries an "approximate five-year maturity while its underlying customer contracts average approximately three years in length."

Dell'Oro reported that worldwide data center capex grew 92% in 2Q 2026. It found that "Neocloud and AI Model Builder capex grew the fastest among the customer segments." Neoclouds are GPU cloud providers outside the hyperscalers.

The GPU layer is where lenders take hardware and contract risk, and it is where smaller operators need capital. For how a deal moves from purchase order to funding, see the GPU financing process.

What Changes at 1–10 MW

At 1–10 MW, the operator doesn't build a shell: clusters go into existing colocation, and permanent financing is released only once the servers are installed. That removes the speculative construction period of a large campus build, as we set out in our analysis of middle market compute credit. In the middle market deals we review, 3–4 year take-or-pay offtake contracts are common, meaning the customer pays for contracted capacity whether or not it uses it.

These deployments sit in the Compute Middle Market, the GPU cloud operators between startups and hyperscalers. The Compute Middle Market tracker counts GPU-backed financings of $500 million or less, with data as of Nov 3, 2026. That threshold is the tracker's own definition of the segment.

Data center operators meet this financing in one of two roles. As a host, you provide the site and the GPU owner borrows. As an owner, you finance your own GPU deployment like any GPU operator, with the GPUs defined separately from the building and power.

Step-in rights let the lenders take over the GPU owner's position under the colocation agreement if a loan defaults. A lien waiver confirms the data center holds no claim on the hardware ahead of the lenders.

Risks Lenders Price in Data Center Financing

Power, contract tenor and hardware refresh set terms across both layers, and JLL's 1% vacancy in North America (H1 2026) shows how tight powered capacity is. The region had 66 GW under construction, 95% of it pre-committed (August 2026). In January 2026, JLL put the average grid connection wait in primary markets at more than four years, varying by market.

Lawrence Berkeley National Laboratory projects data centers could account for 11.8% of US electricity by 2030, within 9.5–15.3% (June 2026). For a 1–10 MW operator, a colocation slot with power is part of the collateral's value.

Contract tenor sets how fast the debt must pay down. CoreWeave's underlying contracts average about three years (August 2026), so a loan matched to them amortizes inside that window. Where a loan runs longer, as CoreWeave's five-year facility does, lenders carry renewal risk.

Hardware refresh sets the collateral's working life, which JLL puts at roughly five years for GPUs. Straight-line amortization repays principal in equal monthly amounts over that working life. In our view, a GPU-layer loan holds up when the site has power, the offtake contract covers repayment and amortization finishes inside the refresh cycle.

Conclusion

Data center financing splits into a building and power layer funded like infrastructure and a GPU layer funded against hardware and offtake contracts. At 1–10 MW, the financing question is the GPU layer. GPU financing through USD.AI covers eligible GPU deployments; land, buildings and standalone power projects are outside it.

Deployment cost means the eligible hardware on the purchase order: servers, networking, owned racks, PDUs and cabling. Soft costs such as installation, labor, shipping, taxes, software subscriptions and data center build-out aren't financed. A special-purpose vehicle (SPV), a separate company that holds only the deal's assets, holds the GPUs, offtake contracts and revenue accounts.

Capital is held in escrow until the servers are installed, burned in (run under load to catch early faults) and independently verified. The collateral sits in a Tier III or Tier IV data center with a colocation lien waiver. Compute IQ scores every deal, rating the operator, the customer and the offtake contract on a 100-point scale.

As of September 2026, USD.AI had reviewed over 700 GPU financing deals, with no missed payments to date. Hosts and owners can see what each role involves on our page for GPU financing for data center operators. Talk to USD.AI about financing your GPU deployment.

FAQ

How does data center financing work?

Data center financing works in two layers that are funded separately. The land, shell and power are financed through construction loans, project finance, ABS/CMBS, joint ventures and infrastructure funds. The GPUs, servers and networking inside are financed through equipment leases or loans against offtake contracts.

What is the difference between corporate debt and project-level debt in data center financing?

Corporate debt, such as hyperscaler bonds, is raised against the parent company's whole balance sheet. Project-level debt sits in a dedicated entity, is repaid from that asset's leases or contracts, and leads S&P-rated North American issuance. In the off-balance-sheet version BIS describes, a joint venture owns the campus while the hyperscaler holds a minority stake, leases capacity and may give guarantees.

How much of a GPU deployment does USD.AI finance?

Financing through USD.AI covers up to 80% of the deployment cost, with the operator contributing at least 20% as equity. A debt-service reserve covering three months of payments is funded at closing. Financing starts from about $5 million, with larger facilities available.