In short
Building an AI data center is a multibillion dollar decision. The way a project is structured determines who takes the risk, who controls the design, how fast it gets built, and how it is paid for. In the United States, no single federal law sets the rules. The structures come from private contracts, standard industry forms, and the demands of the capital markets.
The three big choices are the construction delivery method, the ownership and lease structure, and the financing stack. This article walks through each one, from EPC and design-build to powered shells, joint ventures, and asset backed securitization. The US AI data center financing market reached $30 billion in 2024 and is expected to reach $60 billion in 2025. Project finance analysis
What federal rules apply to AI data center construction projects?
No single statute dictates how an AI data center project must be organized. The controlling rules are set by contract law, industry agreements like the AIA forms, and the securities and tax laws that govern financing. The one federal action that touches these projects directly is an executive order from July 2025. It defines a Data Center Project as a facility that adds more than 100 megawatts of new load and is dedicated to AI inference, training, simulation, or synthetic data generation. The order directs federal agencies to speed up environmental reviews and find ways to categorically exclude qualifying projects from lengthy NEPA processes. White House This does not change how the construction contract is written, but it can shorten the overall timeline.
What are the main construction delivery methods for an AI data center?
A construction delivery method sets out who designs the building, who builds it, and who carries the risk of cost overruns and delays. The five methods used in AI data center projects are EPC, EPCM, design-build, design bid build, and CMAR. The right choice depends on whether the owner wants speed, cost certainty, or design control.
EPC (Engineering, Procurement, Construction)
Under an EPC contract, one contractor takes full turnkey responsibility. The contractor handles detailed engineering, buys all the equipment and materials, and runs the construction. The price is usually fixed, a lump sum. The contractor bears most of the cost and schedule risk. The owner gets a completed facility that must meet specific performance promises. Design, procurement, and construction happen in parallel, which can shorten the schedule. This hands off approach is why many owners favor EPC for large, fast track AI data centers. Construction Dive
For AI data center projects with a total construction cost under $500 million, such as those for commercial cloud services, a guaranteed maximum price model is more common. Large scale AI and hyperscale AI data center projects tend to use a target price construction model with a painshare/gainshare mechanism. Construction contracts analysis
EPCM (Engineering, Procurement, Construction Management)
With EPCM, the owner contracts directly with the designers, equipment suppliers, and trade contractors. An EPCM firm manages the work, keeps everyone coordinated, but does not guarantee the final cost or schedule. The owner keeps full control and flexibility, but also keeps the risk of any overruns or delays. EPCM is often chosen on large or fast evolving AI data center projects where phased deployment, procurement flexibility or evolving technical requirements are priorities. Construction delivery analysis
Design build
A single entity handles both design and construction under one contract. The phases overlap, which can cut the total time. The owner has more input on design than under a pure EPC but less direct control than under design-bid-build. Because the designer and builder are on the same team, there are often fewer change orders and clearer accountability. Late step design changes, however, can get expensive. Project Control Academy
Design bid build (DBB)
Design-bid-build splits the project into two separate contracts. First, the owner hires an architect or engineer to produce complete design documents. Then the owner puts the construction work out for bid and hires a contractor. The roles are clear and the pricing is transparent because the design is fixed before bids come in. But the sequential order makes DBB slower than design-build, and gaps between the design and the construction team can lead to disputes. AIA Contract Documents, ASD analysis
CMAR (Construction Manager at Risk)
In CMAR, a construction manager joins the project during the design phase, gives advice on cost and constructability, and then acts as the general contractor under a guaranteed maximum price. The owner gets early cost visibility, but the arrangement needs strong coordination among the designer, the CMAR, and the owner. AIA Contract Documents
How the methods compare
The table below lines up the five methods side by side. It is a quick way to see where the risk sits and what trade off each method makes.
| Method | Who holds most cost risk | Speed | Design control by owner | Best for |
|---|---|---|---|---|
| EPC | Contractor | Fast (parallel work) | Low | Large, repeatable builds where speed and price certainty matter most |
| EPCM | Owner | Fast to moderate | High | Projects with evolving technical needs |
| Design-Build | Shared | Fast (overlapping phases) | Moderate | Projects that need both speed and owner design input |
| DBB | Owner (no GMP) | Slower (sequential) | High | Owners who want complete design control before pricing |
| CMAR | Shared (GMP sets cap) | Moderate | Moderate | Owners who want early cost input and a guaranteed ceiling |
How do ownership and lease structures work in AI data centers?
Ownership structures define who owns the building and the land, who pays for the interior equipment, and what the tenant relationship looks like. The three common arrangements for single tenant facilities are build to suit, powered shell, and turnkey. For shared investment, a joint venture between a developer and a capital partner is the standard model.
Build to suit (BTS)
A developer acquires land, designs and builds the full facility to one tenant’s specifications, and then leases it back under a long term lease. Leases usually run 10 to 20 years and are triple net. That means the tenant pays property taxes, insurance, and maintenance, while the developer retains ownership of the shell and land. Single facility BTS deployments start at 25 megawatts and often exceed 100 megawatts. Dgtl Infra
Powered Shell
A developer delivers only the building envelope, dual feed power from one or two substations, and basic connectivity. All of the critical interior infrastructure, UPS, generators, chillers, cooling systems, is left out. The tenant installs and owns its own IT level systems. Because the developer is not waiting for long lead time equipment, the delivery timeline is 9 to 18 months, much faster than the 18 to 36 plus months for a ground up custom build. Leases are quoted per square foot, typically $10 to $25 per square foot annually, under triple net terms.
Powered shells have historically been concentrated in Northern Virginia, which accounts for roughly 100 of the 130 powered shell deals tracked globally as of early 2025. datacenterHawk
Turnkey
The developer hands over a fully fitted, ready to operate facility, including all power, cooling, and IT infrastructure. The price is quoted per kilowatt of critical load, not per square foot. Turnkey is the opposite of a powered shell. datacenterHawk
Joint ventures for ownership
For large scale development, a developer operator often pairs with an institutional equity partner, an infrastructure fund, a pension fund, or a private equity firm, inside a special purpose vehicle. The operator usually holds 20 to 25 percent of the equity. The financial investor puts up the rest. The operator earns a promote, a share of profits above a preferred return hurdle, which is typically a 10 to 12 percent internal rate of return.
A typical joint venture also includes a development management fee of about 4 to 5 percent of total capital costs, an asset management fee of 3 to 4.5 percent of revenue (not including power), and a sales commission of 3 to 5 percent of the deal value when a project sells. The targeted development yield for a 15 year triple net lease with a major hyperscaler is 7.5 to 8.5 percent. Alantra
A real world example is the Hyperion campus in Richland Parish, Louisiana, the Beignet JV. In October 2025, funds managed by Blue Owl Capital took an 80 percent equity interest in the roughly 2 gigawatt campus and Meta retained 20 percent, in a deal with a total development cost of about $27 billion. The venture was backed by one of the largest data center debt financings on record, rated A+ by S&P. Meta press release, CBRE H2 2025
How are AI data center projects financed?
AI data center construction requires enormous amounts of capital. A single 200 megawatt hyperscale AI training campus costs about $8.2 billion, about $2.2 billion for the shell and core, $0.4 billion for power infrastructure, and roughly $5.6 billion for IT equipment. Columbia Business School Financing can come from equity, debt, or a mix, and the capital stack shifts as a project moves from risky construction to stable operations.
Project finance
Project finance is debt that is structured against a specific project’s assets and future cash flows, not against the sponsor’s corporate balance sheet. The lenders look to the credit of the tenant under the long term lease. For AI data centers, the tenant is often a large hyperscaler with a AA or AA+ credit rating, which makes the deal very safe. Lease terms run 10 to 20 years and are typically hell or high water triple net, meaning the tenant must pay no matter what.
Banks will lend 60 to 90 percent of a project’s value during construction, but the higher leverage typically comes later, once the building is up and the lease is in place. More than 50 commercial banks are active in this lending space. The debt taps into commercial bank markets, capital markets, private credit funds, and permanent capital. Project finance analysis A notable example is the $600 million construction financing for a Virginia AI data center in December 2024, which was broadly syndicated across traditional project finance lenders, private credit funds, and regional banks. Project finance analysis
Sale leaseback
A sale leaseback is a tool for an owner operator who has already built a facility. The owner sells the finished AI data center to an investor and immediately leases it back, usually under a long term triple net lease. This converts owned real estate into an operating lease expense, which frees capital for the core business. Sale-leaseback yields are typically 6 to 8 percent. AI infrastructure companies increasingly use this strategy to monetize GPU hardware to fund further expansion. HostingJournalist, Introl
Asset backed securitization (ABS)
Once a developer has a portfolio of stabilized AI data centers with long term leases, the steady cash flows can be bundled and sold as bonds. This type of securitization can reach up to 70 percent loan to value and often provides the lowest long term cost of capital for large standardized portfolios. In 2025, the market for single asset, single borrower commercial mortgage backed securities hit a record $11.2 billion in annual issuance, and AI data center deals made up about 11 percent of that total. AI data center ABS or CMBS bonds typically price about 100 to 150 basis points above similarly rated hyperscale corporate bonds. Project finance analysis
How the capital stack changes over time
A greenfield project that still faces entitlement and construction risk will typically need 50 to 70 percent equity. Once the facility is operational and leased, the debt share can rise to 65 to 75 percent. Preferred equity for the construction phase costs 12 to 15 percent annual return or more. Senior bank debt for a stable, leased asset is priced at SOFR plus a spread. Percepture
A mid sized operator seeking $100 million to expand might offer 20 to 30 percent ownership to equity investors. Those equity investors target a 15 to 20 percent internal rate of return. Debt providers look for 6 to 8 percent annual returns. Global Data Center Hub
Real transactions in the market show the scale. CoreWeave secured a $7.5 billion debt facility in May 2024 led by Blackstone and Magnetar Capital. LinkedIn / Brandon Pfeffer Equinix formed a joint venture with GIC and CPP Investments in October 2024, with plans to raise more than $15 billion for US hyperscale expansion. LinkedIn / Brandon Pfeffer EdgeCore Digital Infrastructure raised $1.9 billion in equity and a matching $1.9 billion in green debt financing, and later completed a $235 million ABS to refinance its construction loans. LinkedIn / Brandon Pfeffer
What role do modular and prefabricated construction play?
Speed is critical in the AI buildout. Modular and prefabricated construction compresses timelines by building portions of the facility in a factory and assembling them on site. The US modular AI data center market was valued at about $11.87 billion in 2025 and is projected to reach $30.61 billion by 2032, a compound annual growth rate of 14.5 percent. MarketsandMarkets
Prefabrication can cut overall deployment time by up to 50 percent compared to traditional construction. For liquid cooled AI modules, deployment can take as little as 8 to 10 months. Introl
Real examples show the approach working at scale. Compass Data Centers was constructing a 400 megawatt campus in Red Oaks, Texas, using 40 megawatt halls that were 70 percent prefabricated, with much of the batch processing done directly on site. AFCOM Vapor IO deployed 36 micro modular AI data centers across 20 cities in 11 months. Each 150 kilowatt module was built in a factory, shipped on a flatbed truck, and became operational within 72 hours of delivery, at a cost about 40 percent lower than traditional construction. Introl Vertiv launched its MegaMod CoolChip in 2024, a prefabricated modular solution with integrated direct to chip liquid cooling that supports multi megawatt configurations and cuts deployment by up to half. MarketsandMarkets
The hybrid approach, a mix of traditional building and prefabricated modules, appears to be the most common approach to new AI data center builds over the next three years, according to the 2024 AFCOM State of the Data Center Report. AFCOM
What key contract provisions protect the parties in an AI data center build?
Beyond the delivery method, the construction contract itself must spell out performance expectations, testing protocols, and what happens when things go wrong. Three areas get a lot of attention.
Performance guarantees and commissioning
AI data center contracts now regularly include explicit performance guarantees. These can require a specific uptime, often 99.999 percent availability, a set power density per rack, a maximum power usage effectiveness number, and compliance with a defined tier certification. If the facility fails to hit those marks, remedies can include performance liquidated damages or termination and rejection rights. Construction contracts analysis
Before a facility is accepted, it goes through commissioning, a step by step testing process. Industry practice defines five or six commissioning levels. Level 0 is design and planning. Level 1 is factory testing. Level 2 covers delivery, installation, and pre startup checks. Level 3 is system startup. Level 4 is functional performance testing. Level 5 is integrated systems testing under failure scenarios, the hardest test, and Level 6 is final handover. Each level uses a color coded tagging system. Red for Level 1, yellow for Level 2, green for Level 3, blue for Level 4, and white for Level 5. constructandcommission.com
Force majeure
Force majeure clauses excuse a party when an unforeseen, outside event makes performance impossible. In AI data center contracts, these clauses should be written specifically, not just as boilerplate. They need to cover utility interconnection delays, permitting holdups, supply chain disruptions for critical equipment, tariff or material price spikes, and third party litigation brought by neighboring landowners. Construction delivery analysis
Incentives and liquidated damages
Contracts manage schedule exposure with milestone based incentives and liquidated damages tied to critical path events, not just general completion. For example, a contractor might earn a bonus for early energization or for completing integrated systems testing ahead of schedule. If they are late, liquidated damages apply.
Incentive structures paired with shared cost savings are common, so both parties benefit when the project comes in under budget. Price adjustment clauses tied to material pricing indices, tariffs, or equipment cost increases are also common. Construction delivery analysis
Key takeaways
- EPC shifts cost and schedule risk onto the contractor and is the preferred choice for many large, repeatable AI data center builds. Smaller projects often use a guaranteed maximum price under CMAR.
- Powered shell delivers speed. The tenant gets a shell and power in 9 to 18 months and owns its own critical infrastructure inside.
- A joint venture between a developer operator and an institutional investor aligns expertise with capital. The operator typically gets a promote after a 10 to 12 percent IRR hurdle.
- Project finance ties debt to the project’s leases and assets, not to the sponsor’s balance sheet. LTV can reach 90 percent when a hyperscaler lease is in place.
- Modular and prefabricated construction can cut build times by half, and the hybrid approach is now the industry standard.
- Contracts should name specific force majeure events, set performance guarantees tied to uptime and power density, and use detailed commissioning protocols with color coded testing levels.
Frequently asked questions
Q:What is the difference between EPC and design-build?
A:Under EPC, the owner is nearly hands off. The contractor takes full turnkey risk. Design-build keeps more owner involvement in the design. Both overlap design and construction, but EPC puts more cost and schedule risk on the contractor. Construction Dive
Q:When should an owner choose a powered shell over a build-to-suit?
A:Choose a powered shell when speed matters most and the tenant wants to control its own internal infrastructure. A powered shell delivers in 9 to 18 months versus 18 to 36 months for a custom build.
Q:How much equity is needed to start an AI data center project?
A:A greenfield project with entitlement and construction risk usually requires 50 to 70 percent equity. Once the facility is operational and leased, the debt portion can rise to 65 to 75 percent. Global Data Center Hub
Q:What is a typical preferred return in an AI data center joint venture?
A:The financial investor usually gets a preferred return of 10 to 12 percent internal rate of return before the developer operator earns a promote. Alantra
Q:Can an AI data center project get built without a long term lease in place?
A:It can, but financing is much harder. Lenders rely on the creditworthiness of a long term lease from a hyperscale tenant. Without that lease, debt terms are stricter and equity requirements are higher.
Q:How do sale leasebacks work for AI data centers?
A:The owner operator sells the completed facility to an investor and then leases it back on a long term triple net lease. This shifts the asset off the balance sheet and turns it into an operating expense, freeing capital. Yields now can be as high as 8.5 percent, up from a previous range of 4.5 to 6.5 percent. Data Center Dynamics
Q:What is the cost difference between a powered shell and a turnkey facility?
A:Powered shell leases are priced per square foot, typically $10 to $25 per square foot annually. Turnkey facilities are priced per kilowatt of critical load because the developer delivers all the interior infrastructure. LandGate
Q:How long does it take to build an AI data center from breaking ground to handover?
A:A powered shell can be delivered in 9 to 18 months. A full custom build usually takes 18 to 36 months. Modular liquid cooled builds can be done in 8 to 10 months. Introl
Q:What is commissioning and why is it important?
A:Commissioning is the systematic testing of every system in the AI data center before the owner accepts it. It moves through five or six levels, from factory testing to integrated failure scenario testing. It confirms that the facility meets its performance guarantees. constructandcommission.com
Q:Does the federal government offer any special incentives for AI data center construction?
A:Not directly through tax credits or grants tied to the construction structure itself. The July 2025 executive order aims to speed up permitting for AI data center projects over 100 MW, which can shorten the pre construction timeline, but it does not override local zoning or state permitting. White House
Subscribe to The Compute Law Brief
The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.
Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.