Project finance for AI data center construction

In short

In 2025, U.S. AI data center financings were projected to reach $60 billion. That was nearly double the $30 billion from 2024. Equipment Leasing & Finance Foundation report, Law firm analysis Lending commitments hit $121 billion. The number of loans fell, but the average deal size grew sharply. Data Center Dynamics The money comes from many sources. Big Tech hyperscalers, commercial banks, private credit funds, and a fast growing asset backed securities market all participate. AI data center developers typically rely on mini perm construction loans that cover construction plus three or four years and finance a project’s construction and its first years of operation. Much of the AI data center buildout has also been financed through off balance sheet structures involving bankruptcy remote special purpose vehicles, where lenders rely on project level cash flows for repayment. Law firm analysis, Law firm analysis Federal tax law adds powerful incentives. 100 percent bonus depreciation on servers and cooling equipment can sharply reduce first year taxable income. Tax alert The real deals show the range of options. CoreWeave’s $7.5 billion GPU backed loan, Meta’s $27 billion off balance sheet joint venture with Blue Owl, and Oracle’s Stargate syndicated facility each illustrate a different financing approach.

How big is the AI data center construction market?

The 2025 numbers are large. They show a market now operating at a new scale and urgency.

U.S. data center financings were $30 billion in 2024 and could double to reach $60 billion in 2025. Foundation study Lenders committed $121 billion in credit to US data center properties, up $30 billion from the year before, while the total number of loans fell by 32. That means each loan was larger. Data Center Dynamics

AI data center debt issuance soared to approximately $182 billion in 2025, up from roughly $92 billion in 2024. Market data Global AI data center dealmaking reached a record $61 billion across more than 100 transactions. CNBC

Construction spending hit a single month record of $14.0 billion in July 2025, nearly twice the previous record of $7.4 billion from August 2024. Full year 2025 AI data center construction starts are projected to top $46 billion, following a monthly record of $14.0 billion in July 2025. ConstructConnect

Even with that build out, demand far outpaces supply. The vacancy rate in primary AI data center markets fell to a record low 1.4 percent at the end of 2025, even as supply jumped 36 percent year over year to 9,432 megawatts of total capacity. CBRE

Morgan Stanley projects total global AI data center spending of $2.9 trillion through 2028. The bank estimates a $1.5 trillion financing gap, meaning roughly half the needed capital must come from external capital, with credit markets a major source. Morgan Stanley Research

Where does all this capital come from?

The largest share comes from the Big Tech hyperscalers themselves. Their combined capital expenditure for AI infrastructure in 2025 was approximately $405 billion. Financial analysis

• Amazon, $100 to 125 billion
• Microsoft, $80 billion
• Alphabet, $75 to 85 billion
• Meta, $60 to 65 billion

But the hyperscalers do not build everything on their own balance sheets. A deep and diverse lending market now finances AI data center construction. More than 50 commercial banks are active, along with private credit funds, traditional project finance lenders, and regional banks. Project Finance Law

A newer source of capital is GPU backed lending. Lenders advance money secured by high value NVIDIA chips, which are both essential to AI computing and rapidly depreciating. Over $11 billion in such loans have been issued to AI cloud startups like CoreWeave, Lambda, and Crusoe. Lenders typically advance 50 to 70 percent of the chips’ value at interest rates of 12 to 15 percent to account for that depreciation risk. Industry report

The asset backed securities market, or ABS, is also growing fast. In 2025, SASB CMBS issuance hit a record $11.2 billion, and AI data center deals made up about 11 percent of that total. CBRE These bonds repackage the steady lease payments from an AI data center into tradable securities that institutional investors can buy.

Not every developer accesses capital the same way. Large companies that own big portfolios of existing AI data centers can often borrow unsecured debt at the parent company level. That happens about 80 percent of the time. Smaller or single purpose entities rely on secured, project level financing. Law firm analysis

A $25 billion ABS market is already in place, but the amount of debt that will need to be refinanced in the coming years could approach $300 billion, creating both risk and opportunity for lenders. DoubleLine research, Law firm analysis, DOE report

What are the main financing structures for an AI data center?

The market uses several distinct structures, each suited to different sponsors and risk profiles. The table below outlines the most common ones.

Mini perm construction loans are the standard way to get a project built. A mini perm loan typically runs the construction period plus three or four years. It covers the construction period and the first few years of operation until the project reaches a steady stream of lease income. At that point, the developer refinances into commercial mortgage backed securities, asset backed securities, or private placements. Data Center Financing Structures

How does a mini perm construction loan work?

A mini perm loan connects a construction start to permanent long term financing. Here is a simplified timeline.

1. A sponsor forms a bankruptcy remote SPV to own the project.
2. The SPV signs a lease, often with a hyperscaler, or a power purchase agreement that shows future revenue.
3. A bank or group of lenders provides a loan that covers most of the construction cost, secured by the project’s assets and lease.
4. During construction, the lender monitors progress and releases funds in tranches.
5. Once the AI data center is built and the lease starts producing income, the borrower refinances the mini perm into a longer term loan or sells bonds to investors.

This structure shields the sponsor’s other assets and focuses the lender on the project’s own economics. Law firm analysis

What federal tax incentives power these deals?

Federal tax law gives AI data center investors several powerful tools to reduce taxable income in the early years of a project, when cash flow is tightest. The most important ones are described below.

100 percent bonus depreciation is back

The One Big Beautiful Bill Act permanently restored 100 percent bonus depreciation for qualified property acquired and placed in service after January 19, 2025. Tax alert Qualifying property includes tangible personal property with a recovery period of 20 years or less. That covers servers, graphics processing units (GPUs), specialized cooling systems, uninterruptible power supplies, dedicated electrical distribution, raised flooring, and structured cabling. Both new and used property can qualify if the taxpayer did not previously own it.

There is no annual dollar cap. A developer can deduct the full cost of all qualifying equipment in the year it is put into use. The deduction can even create or increase a net operating loss that carries forward.

Because an AI data center is a mix of real estate and equipment, a cost segregation study is needed to identify which parts qualify for bonus depreciation. Without that study, the default rule would treat most of the building as 39 year property, which is depreciated slowly and does not qualify for the 100 percent write off.

Section 179 creates a smaller, parallel option

Internal Revenue Code Section 179 lets a business deduct up to $2.5 million of qualifying equipment purchases in 2025, subject to a phase out that begins when total asset purchases exceed $4 million. Tax alert Unlike bonus depreciation, the Section 179 deduction cannot create or increase a net operating loss. It is limited to the business’s taxable income. For a large AI data center project, $2.5 million is a small number, but for a small colocation operator buying a few racks of servers, it can be useful.

The component election for self constructed assets

A taxpayer building its own AI data center can make a component election under the bonus depreciation rules. This allows the taxpayer to claim 100 percent bonus on specific components, such as a cooling system or a backup generator, even if the overall building does not fully qualify. The component must be acquired or constructed after January 19, 2025, be qualified property, and be placed in service on time. Construction is deemed to have begun when more than 10 percent of total qualifying costs have been incurred, which is a safe harbor. Tax alert

Cost segregation turns buildings into faster deductions

A professional cost segregation study can reclassify 35 to 60 percent of the total cost of an AI data center project into shorter lived categories. For example, server racks go into five year property, and dedicated electrical distribution goes into five year property. The land and the building shell remain 39 year property. The result is a much larger first year deduction. For a $10 million AI data center project acquired after January 19, 2025, a cost seg can push first year deductions into the $3.5 million to $6 million range. Cost segregation analysis

Interest deduction rules now favor high leverage

Before 2022, a company’s business interest deduction was limited to 30 percent of its adjusted taxable income (ATI), and ATI was calculated after subtracting depreciation and amortization. The One Big Beautiful Bill Act permanently removed depreciation and amortization from the ATI calculation for tax years beginning after 2024. Tax alert That change means a highly leveraged AI data center project can deduct more of its interest expense, which lowers the effective cost of debt.

Watch for state decoupling. Several states, including California, Massachusetts, and New York, do not automatically follow federal bonus depreciation, and California also does not follow the federal interest deduction limitation. A project in one of those states may face higher state tax bills even while enjoying federal benefits, because those states generally do not follow federal bonus depreciation and some, like California, also do not follow the business interest deduction limitation. Mass.gov, NCSL report, Mass.gov, N.Y. Tax Dept.

How do lenders protect themselves?

An AI data center loan is only as good as the repayment stream behind it. Lenders use several tools to manage risk.

Special purpose vehicles (SPVs). The developer typically creates a separate legal entity, an SPV, to own the project. The SPV is bankruptcy remote, meaning that if the developer or its other businesses fail, the project’s assets remain insulated. Law firm analysis

Reserves. Lenders often require the borrower to keep six to twelve months of lease payments in a reserve account, plus additional reserves for tenant improvements, property taxes, and insurance. Law firm analysis

Cross collateralization. When a developer owns multiple data centers, a lender may take a lien on more than one site. Cash flows from an existing operational site can then support debt on a new construction project. DC BLOX secured a $265 million green Senior Secured Credit Facilities loan in October 2024 to develop preleased hyperscale-driven edge sites across the southeastern U.S. DC BLOX press release

Power hedging. Electricity is the single largest operating expense for an AI data center over its life. Developers are increasingly signing virtual power purchase agreements, or VPPAs. A VPPA is a financial contract with a renewable energy generator. The AI data center agrees to pay a fixed price for power and receives renewable energy credits. If wholesale electricity prices rise above that fixed price, the generator pays the difference, and vice versa. This stabilizes the project’s long term power cost and makes its cash flows more predictable to lenders. Verse

ESG pricing. Facilities with strong environmental metrics can sometimes borrow at slightly better rates. For example, Vantage Data Centers secured a $3 billion green loan in 2024, supporting nearly 1.4 gigawatts of new capacity. The pricing advantage for ESG linked financing is in the range of 15 to 25 basis points. Data center ESG financing analysis

What do the biggest real deals look like?

A few landmark transactions show how the different pieces of project finance come together in the real world.

CoreWeave raises billions with GPU backed debt

In May 2024, CoreWeave closed a $7.5 billion debt facility led by Blackstone and Magnetar. It was one of the largest private credit financings in history at the time. The structure was a senior secured term loan collateralized by CoreWeave’s inventory of NVIDIA GPUs and its customer contracts. The collateral sat inside a bankruptcy remote SPV. This deal was the largest in a GPU-backed lending market that reached over $11 billion by late 2024, as Wall Street lenders financed AI cloud companies using Nvidia chips as collateral. ABF Journal, CoreWeave press release, Law firm analysis, Business Insider, CoreWeave press release, Financial Times, Law firm analysis

The deal’s price tag reflects the rapid depreciation of GPUs. CoreWeave later went public in March 2025 at $40 per share, and followed with a $2.6 billion dedicated debt facility in July 2025, with the interest rate set at SOFR plus 4 percent, specifically to build out infrastructure for OpenAI. MLQ.ai

Meta pushes its largest campus off its balance sheet

In October 2025, Meta and Blue Owl Capital formed a $27 billion joint venture to build the Hyperion AI data center campus in Richland Parish, Louisiana. Blue Owl funds own 80 percent, and Meta retains 20 percent. Meta contributed land and partially built facilities. Blue Owl contributed about $7 billion in cash, and Meta received a $3 billion distribution. The project will span roughly 2,250 acres, with about 4 million square feet across at least nine buildings. Initial power capacity is 2 gigawatts, with room to expand beyond 5 gigawatts. Construction is expected to run through 2030.

The financing came through a private bond offering led by PIMCO, the largest project finance bond on record. The debt is off Meta’s balance sheet. Meta leases the facilities under four year terms, extendable, and guarantees the residual value for 16 years. That means Meta owns only a 20% stake in the real estate, but it leases the capacity it needs, while Blue Owl and the bondholders take the real estate risk, though Meta provides a residual value guarantee for the first 16 years of operations. CNBC, Bisnow, Data Center Dynamics

Other structures at scale

• Vantage Data Centers closed a $3 billion green loan in 2024, funding 1.4 gigawatts of IT capacity across North American leased and greenfield sites. Press release
• DC BLOX used a hybrid approach in September 2024. It cross collateralized cash flows from operational data centers with new construction across nine southeastern U.S. locations, later expanding the loan to $850 million. Data Center Financing Structures, DC BLOX, ABF Journal
• Apollo Funds acquired Stream Data Centers in 2025, securing a development pipeline of more than 4 gigawatts. CBRE
• BlackRock, MGX, and AIP agreed to acquire Aligned Data Centers for $40 billion in 2025, one of the largest enterprise value AI data center deals ever. CBRE

These deals show the variety. Pure play AI cloud debt, hyperscaler co investment, ESG linked loans, and infrastructure fund acquisitions each use a different structure.

What are the main constraints on new development?

Capital is abundant. Power is not. The U.S. Department of Energy predicts AI data centers will consume between 6.7 and 12.0 percent of the nation’s electricity by 2028. DOE Getting a new electrical grid connection can take up to four years. That makes power the true bottleneck for new projects.

Some developers are turning to bring your own power (BYOP) solutions, building dedicated generation alongside the AI data center. The Meta Hyperion campus, for example, will draw power partly from a new Entergy natural gas plant in Louisiana.

Construction costs are also rising. The average AI data center now costs about $220 million to build, measured over the trailing twelve months through July 2025. ConstructConnect GPUs and custom AI accelerators account for roughly one third of total AI data center capex. Dell’Oro Group analysis

Refinancing risk is real. The market for AI data center asset backed securities is about $25 billion today, but the amount of debt that will need to be refinanced as construction loans mature could approach $300 billion. Law firm analysis If interest rates stay high or investor appetite wanes, some projects could struggle to refinance on favorable terms.

Finally, geography matters. Arizona led the nation with over $41 billion in data center loans in 2025, followed by Illinois, Texas, and New Mexico. Virginia, the world’s largest AI data center market, ranked ninth with only $3 billion in loans. Data Center Dynamics Several states offer aggressive tax breaks to attract AI data centers, while others, like California and Colorado, do not. Stream Data Centers, RSM, Bloomberg Tax, NCSL report

Key takeaways

• The U.S. AI data center construction market is attracting record capital. Lending commitments hit $121 billion in 2025, and the market may double again.
• The main sources of construction financing are hyperscaler capex, banks, private credit funds, and a growing ABS market.
• Mini perm construction loans of construction plus three or four years are the standard. They bridge the gap to permanent financing once the AI data center shows stable lease income.
• Federal tax law now provides 100 percent bonus depreciation on servers, cooling, and electrical equipment, with no annual cap. A cost segregation study is essential to capture these deductions.
• The interest deduction limit has been relaxed by permanently removing depreciation and amortization from the adjusted taxable income calculation, cutting the after tax cost of debt.
• Lenders protect themselves with bankruptcy remote SPVs, cash reserves, cross collateralization, and power price hedges.
• GPU backed lending has become a substantial new asset class, with over $11 billion deployed at interest rates of 12 to 15 percent.
• Power, not capital, is the principal constraint. Grid connections can take years, pushing developers toward on site generation.
• States matter. Arizona, Texas, and Illinois attract the most debt capital, partly due to favorable tax policies.

Frequently asked questions

Q:What is bonus depreciation, and why does it matter for AI data centers?

A:Bonus depreciation lets a business deduct the full cost of qualifying equipment in the year it is placed in service, instead of spreading the deduction over many years. For an AI data center, this can cover servers, cooling systems, and electrical distribution. The new law makes 100 percent bonus permanent and applies to property placed in service after January 19, 2025. Tax alert

Q:What is a mini perm construction loan?

A:It is a short term loan, usually covering construction plus three or four years, that finances a project from the start of construction through its first years of operation. Once the AI data center is leased and generating steady income, the borrower refinances into longer term bonds or a commercial mortgage. Mini-Perm Loan

Q:Can a startup get project financing for an AI data center?

A:Yes, if it has contracted revenue or valuable collateral. GPU backed lenders will advance money against NVIDIA chips, even for a young company like CoreWeave or Lambda, as long as the chips and customer contracts are pledged. But interest rates are high, typically 12 to 15 percent.

Q:How does a borrower keep the lender’s risk separate from its other assets?

A:By using a special purpose vehicle (SPV). The SPV is a separate legal entity that owns the project and the debt. The sponsor’s other business assets are shielded from the project’s lenders, and the project’s assets are shielded from the sponsor’s other creditors. Financial Edge

Q:Why did Meta put its Louisiana AI data center into a joint venture instead of building it directly?

A:To keep the $27 billion in construction debt off its balance sheet. Meta leases the facility, so it gets the computing capacity without recording the real estate debt as a liability. Blue Owl and the bond investors hold the real estate risk.

Q:What is a virtual power purchase agreement (VPPA)?

A:A VPPA is a financial hedge on electricity prices. The AI data center agrees to buy power at a fixed price from a renewable energy project. If the market price goes above that fixed price, the project pays the data center the difference. If the market price falls below, the data center pays the project. It stabilizes long term power costs. Quick Guide to Virtual Power Purchase Agreements

Q:Why is power availability the biggest constraint on AI data center building?

A:A new electrical grid connection can take up to four years in congested markets. Meanwhile, AI data center electricity demand is projected to reach up to 12 percent of total U.S. consumption by 2028. Finding enough power is harder than finding enough capital. Law firm analysis

Q:Do all states follow the same tax rules for AI data centers?

A:No. Several states, including California, Massachusetts, and New York, have decoupled from the federal bonus depreciation and interest deduction rules. Others offer generous incentives. Alabama offers up to 30 years of tax breaks, and Arizona exempts data center equipment from transaction privilege and use taxes for up to 20 years. Stream Data Centers

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.