Anchor tenant and colocation leases for AI data centers

In short

Large AI data centers depend on anchor tenant and colocation leases rather than owner occupied facilities. The anchor tenant, often a hyperscaler like AWS or Meta, signs a long term lease that gives the developer a predictable revenue stream to finance construction. Leases are priced per kilowatt of power capacity per month, not per square foot. Three lease structures dominate. Triple net (NNN) is used for anchor tenants, modified gross plus electric (MG+E) for wholesale colocation, and full service gross for retail tenants. Anchor tenant leases commonly run 10 to 20 years at $100 to $150 per kilowatt per month, with the tenant paying base rent plus all operating costs. CREFC Data Center E-Primer, p.4. In 2024 and 2025, such leases backed massive debt financings. TeraWulf issued $3.2 billion in bonds backed by a Google credit enhancement at its Lake Mariner site. CoreWeave obtained a $7.5 billion GPU collateralized loan. Law firm analysis. Key lease provisions include PUE caps that shift cooling inefficiency risk to the landlord, strict service level agreements for uptime, and expansion rights. CREFC E-Primer, pp.4-5, Law firm leasing analysis, Jan. 2026. While these structures open capital markets, they also introduce risks such as the mismatch between short GPU depreciation and long lease terms, and the growing reliance on opaque off balance sheet vehicles. Policy analysis on data center finance, Nov. 2025, Law firm financing risks analysis, 2025.

What scale of investment relies on AI data center leases?

The top five U.S. hyperscale cloud providers spent about $443 billion on capital expenditures in 2025, a 73 percent jump from the prior year. That number is expected to reach $602 billion in 2026. CREFC E-Primer, p.3. Much of that spending flows through long term leases rather than direct ownership. Since 2021, U.S. AI data center securitization, including asset backed securities and commercial mortgage backed securities, reached about $57 billion. Morgan Stanley projects another $130 billion in net issuance from 2026 through 2028. CREFC E-Primer, p.3. In the Americas at the end of 2025, 25.3 gigawatts of AI data center capacity was under construction, with nearly 89 percent already pre-committed and a vacancy rate of only 4.2 percent. Cushman & Wakefield, H2 2025. Anchor tenant leases make this capital flow possible because they give lenders a predictable cash flow to lend against.

What are the three main lease structures for AI data centers?

AI data center leases fall into three categories based on who pays for operating costs and power. The structure chosen depends on the tenant’s size and the deal size. The table below summarizes them.

CREFC E-Primer, pp.5, 24.

A hyperscale anchor tenant lease is the backbone of an AI data center project. It provides stable rent for a decade or longer, making the project financeable. Wholesale colocation is common for multi tenant buildings where several companies lease space and power. Retail colocation serves small users with short term needs.

How are AI data center leases priced?

Unlike office or industrial leases priced per square foot, AI data center leases are priced per kilowatt of power capacity per month. The reason is simple. Power, not floor space, is the scarce resource that runs the servers. Monthly base rent for a hyperscale NNN lease runs from $100 to $150 per kilowatt. For wholesale colocation, it is $150 to $250, and retail colocation can reach $200 to $400. CREFC E-Primer, p.4.

Power reimbursement can represent 30 to 50 percent of total revenue. CREFC E-Primer, p.16. In an MG+E lease, the tenant pays for metered electricity on top of base rent. In an NNN lease, the tenant pays the utility directly or reimburses the landlord. Investment grade hyperscale tenants can negotiate low power rates. Rating agencies model lifecycle reserves as low as $5 per kilowatt per month for long term investment grade leases, with KBRA estimating $2 to $9 for hyperscale and $7 to $18 for colocation. CREFC E-Primer, p.5.

Modern leases commonly include a power usage effectiveness cap, known as a PUE cap. PUE measures how much extra energy an AI data center uses for cooling and other overhead beyond the computing itself. A PUE of 1.4 means that for every watt used by servers, another 0.4 watts go to cooling. A PUE cap in the lease sets a limit. If the actual PUE exceeds the cap, the landlord must pay for the extra cooling inefficiency, not the tenant. This shifts operational risk to the landlord. CREFC E-Primer, p.5.

In the Chicagoland market, modified gross base rent typically escalates 1 to 3 percent each year. Data center TCO primer. Operating expense pass throughs in MG+E leases often use a base year index, with the tenant paying increases above that year’s level.

Power ramp schedules are another important feature. A tenant that commits to 100 megawatts rarely uses all that power on day one. The lease often specifies a schedule for drawing more power over time, and the rent may increase accordingly. This helps the developer match revenue to the actual cost of delivering power gradually. Data Center Development Model.

What key provisions appear in AI data center leases?

Beyond base rent and term, AI data center leases contain many covenants that allocate risk and give tenants flexibility. The most important ones are below.

Service level agreements (SLAs)

SLAs set measurable performance standards that the landlord must meet. Common metrics include power availability of 99.999 percent, often called five nines, temperature and humidity ranges, incident response times, and physical security. If the landlord fails to meet an SLA, the tenant receives a service credit against rent. Repeated SLA breaches can give the tenant the right to terminate the lease. Law firm leasing analysis, Jan. 2026, Ch. 2.

Expansion rights

Most anchor tenants negotiate the ability to add capacity later without renegotiating the entire lease. This may take the form of an option to lease additional space, a right of first refusal on adjacent space, or a right of first offer. Growth can also be structured across multiple leases on a campus. Modular building design makes phased expansion easier. Law firm leasing analysis, Jan. 2026, Ch. 1.

Assignment and subleasing

Hyperscale tenants often restrict the landlord’s ability to assign the lease or sell the property without the tenant’s consent. On the other side, tenants typically negotiate the right to assign the lease or sublease space for financing purposes, including the right to pledge the leasehold as collateral. Assignment and encumbrance rights, notice and cure periods, and lender protection provisions ensure lenders can step in if the tenant defaults. Law firm analysis on long term leases, Part I.

Landlord waiver of improvement rights

The landlord usually waives any statutory or common law right to own tenant installed improvements, such as server racks, cooling improvements, or generators. Lenders require a first lien on the data center and all associated improvements, so the landowner’s waiver of improvement rights is essential for the tenant and its lenders to retain ownership and control of those assets. Law firm analysis on long term leases, Part I.

Credit support

Because the rent stream secures the developer’s construction loan, landlords require strong credit support. This can be a security deposit, a letter of credit, or a guarantee from the tenant’s parent company. In large deals, a third party may also backstop the tenant’s obligations. The Google credit enhancement for Fluidstack’s leases at TeraWulf’s sites is a high profile example. Google agreed to backstop certain of Fluidstack’s lease obligations, with the option to pay a termination fee or assume the lease and pay rent if Fluidstack defaults, under a backstop totaling $1.3 billion for the Abernathy, Texas project. TeraWulf Form 8-K, Oct. 28, 2025.

Colocation sublicense rights

Many anchor tenants, and even wholesale tenants, reserve the right to sublicense portions of their space to other users under colocation agreements. They often need only give notice to the landlord, not obtain consent. The Applied Digital lease with CoreWeave explicitly defines a colocation agreement and permits colocation license agreements without landlord consent, while expressly stating such agreements are not the grant of a leasehold interest. Applied Digital/CoreWeave lease, SEC Exhibit 10.1, May 28, 2025.

Cooling system specifications

Modern AI workloads use high density liquid cooling. Leases now specify direct to chip cooling, closed loop systems, and coolant distribution units. The lease may require acceptance testing and commissioning, often to a Level 5 standard, before rent begins. The lease requires commissioning of the premises’ infrastructure, which includes cooling systems, to pass Level 5 criteria, and rent does not begin until the Commencement Date, when all delivery conditions are satisfied. Applied Digital/CoreWeave lease, Schedule 1 (Building Systems, Commissioning, Commencement Date) and Section 2.1.

Data sovereignty and residency

Some leases include clauses giving the tenant the right to audit the location of its data. The landlord may warrant that operational data stays within the agreed jurisdiction. There may be protocols for handling law enforcement requests. While these clauses are more common in cross border deals, they now routinely appear in AI data center lease and colocation agreements as well. Law firm analysis on next generation data centre agreements.

Technology proofing

Because AI hardware evolves quickly, tenants want the right to install next generation cooling or draw higher power density later. Technology proofing clauses permit such upgrades, often with a cost sharing framework for any base building upgrades needed. Law firm analysis on next generation data centre agreements.

Take or pay provisions

Some colocation agreements require the tenant to pay for a minimum amount of power or space whether or not they actually use it. This guarantees the landlord a baseline revenue even if the tenant’s demand fluctuates. Data center take-or-pay provisions.

How do anchor tenant leases support debt financing?

An anchor tenant’s promise to pay rent for 10 to 20 years is the collateral that makes an AI data center project bankable. Developers and sponsors have developed several financing structures that ride on these lease payments.

Off balance sheet special purpose vehicles (SPVs). Large tech companies often do not want data center lease obligations on their own balance sheets. They form a captive SPV that leases the facility and issues debt. The parent company may guarantee a portion of the payments, but the liability often appears only in a footnote to the annual report. In under two years through late 2025, more than $120 billion of AI data center spending moved off hyperscaler balance sheets through SPVs and similar structures. Law firm financing risks analysis, 2025, p.1. Meta’s 5 gigawatt Hyperion project in Louisiana used a captive SPV that issued $30 billion in debt, with Meta committing to cover any revenue shortfall. That residual value guarantee was described only in the footnotes to Meta’s annual report and no liability was recorded. Law firm financing risks analysis, 2025, p.5.

GPU collateralized loans. Some operators borrow against the value of the graphics processing units they install, with the lease payments servicing the debt. CoreWeave, a GPU cloud provider, secured a $7.5 billion debt facility in May 2024 led by Blackstone and Magnetar. The loan is secured by CoreWeave’s GPU inventory and the customer contracts that generate revenue. CoreWeave S-1, SEC.gov.

Bond offerings with lease backstops. TeraWulf, the former bitcoin miner turned HPC operator, issued $3.2 billion of 7.75 percent Senior Secured Notes due 2030 to finance its Lake Mariner campus in New York. The bonds were rated Ba2/BB by the major agencies and were heavily oversubscribed with about $10 billion in orders. The key credit strength was the backstop from Google, which agreed to cover Fluidstack’s lease obligations if Fluidstack defaulted. TeraWulf Form 8-K, Nov. 10, 2025. This structure effectively lifted the bonds to investment grade credit quality despite TeraWulf’s own rating.

Commercial mortgage backed securities (CMBS) and asset backed securities (ABS). Once an AI data center is built and leased, the lease itself can be securitized. In 2025, Blackstone and QTS issued a $3.5 billion CMBS single asset single borrower deal backed by ten QTS data centers. The most junior tranche was reportedly 23 times oversubscribed. A separate QTS Phoenix ABS in September 2025 raised ~$600 million, backed by a single hyperscale tenant under an NNN lease with a ~58.6 percent loan to value ratio. CREFC E-Primer, pp. 14 15. These instruments allow developers to recycle capital into new projects.

What do the biggest real world anchor tenant deals look like?

The following deals show how anchor tenant leases work in practice and the scale of current commitments.

• TeraWulf / Fluidstack / Google. TeraWulf, once a bitcoin miner, converted its Lake Mariner site in New York to host high performance computing. It leased 450 megawatts of capacity to Fluidstack under three ten year leases, producing about $6.7 billion in contracted lease payments. Google provided a credit backstop that covers Fluidstack’s obligations. A separate joint venture in Abernathy, Texas, added a 25 year, 168 megawatt lease with about $9.5 billion in contracted revenue, also backstopped by Google. TeraWulf controls up to 51 percent of that joint venture. In total, the two deals carry $16.2 billion in contracted lease payments and $4.5 billion in Google credit enhancements. TeraWulf Form 8-K, Aug. 18, 2025, TeraWulf Form 8-K, Oct. 28, 2025, TeraWulf Q3 2025 Investor Presentation, Nov. 10, 2025.

• Stargate / Oracle / OpenAI. The Stargate project, announced in January 2025, is a joint venture among OpenAI, Oracle, and SoftBank with a $500 billion, 10 gigawatt target. The Abilene, Texas campus developed by Crusoe Energy is the first large implementation. Oracle holds a roughly 15 year lease and subleases to OpenAI as the anchor tenant. The campus plans 1.2 gigawatts across eight halls with more than 400,000 GPUs. Oracle committed $40 billion for NVIDIA chips there. In July 2025, OpenAI agreed to lease 4.5 gigawatts of compute from Oracle in a deal worth about $30 billion per year, nearly triple Oracle’s annual cloud infrastructure revenue. In September 2025, five more U.S. sites were announced, including locations in Texas, New Mexico, the Midwest, and Ohio, bringing planned capacity to nearly 7 gigawatts. SoftBank press release, Sept. 24, 2025, OpenAI press release, Sept. 23, 2025, IntuitionLabs Stargate guide, Vantage Data Centers press release, Oct. 22, 2025.

• Applied Digital / CoreWeave. In North Dakota, Applied Digital built two AI data center buildings totaling 250 megawatts of critical IT load and leased them to CoreWeave under two roughly 15 year leases. The deal is expected to generate about $7 billion in total lease revenue. CoreWeave holds an option for a third 150 megawatt building. The campus can eventually scale to over 1 gigawatt. Applied Digital press release, June 2, 2025.

• AWS / OpenAI. In November 2025, AWS signed a seven year, $38 billion partnership giving OpenAI access to hundreds of thousands of NVIDIA GPUs. All capacity under the agreement was targeted for deployment by the end of 2026. Aragon Research report on AWS-OpenAI.

• Meta / Hyperion (Louisiana). Meta planned a 5 gigawatt AI data center using a $30 billion SPV structure to keep the obligation off its balance sheet. Meta reportedly provided a residual value guarantee of up to $28 billion but disclosed the commitment only in footnotes. Law firm financing risks analysis, 2025, p.5.

• CoreWeave / Nvidia capacity guarantee. In September 2025, Nvidia agreed to purchase up to $6.3 billion of unused cloud capacity from CoreWeave through April 2032. This effectively guarantees demand for CoreWeave and supports its financing model. Datacenters.com news.

What risks should counsel monitor in AI data center leasing?

The numbers are enormous, but several structural risks deserve attention when reviewing an AI data center lease or its financing.

Revenue shortfall relative to capital spending. In 2025, AI companies generated about $60 billion in revenue while spending around $400 billion on capital projects. Law firm financing risks analysis, 2025, p.1. The gap raises questions about whether future cash flows will cover the debt taken on today.

Asset liability mismatch. The hardware that fills an AI data center, primarily GPUs, depreciates in 3 to 6 years. Yet the AI data center mortgage loans can extend 20 to 30 years, and the anchor tenant lease may last only 10 to 20 years. When the lease ends and the GPUs are obsolete, the developer must refinance or find a new tenant. If demand weakens, the loan may not be repaid. The asset liability mismatch between developers’ term loans and their tenants’ lease terms creates refinancing risk at the developer level. Policy analysis on data center finance, Nov. 2025.

Growing reliance on opaque debt. Four U.S. Senators sent a letter in January 2026 asking regulators to scrutinize the AI sector’s growing use of complex debt, warning that hidden losses could trigger a broader financial crisis. Law firm financing risks analysis, 2026, p.3. The $120 billion moved off balance sheets through SPVs is an example. A parent company’s support may be limited to a footnote promise, not a full guarantee.

Construction and utility delays. Over 30 percent of U.S. AI data center project delays in 2024 and 2025 were tied to utility interconnection issues and permitting. SiteBid market report, Aug. 2025 citing Data Center Dynamics. A delayed start can push a lease commencement date later, reducing the window to earn return on the developer’s capital.

High replacement cost of building systems. Mechanical, electrical, and plumbing systems account for 60 to 70 percent of total development costs. Power distribution alone, at 25 to 30 percent of capital expense, has a 25 to 30 year replacement cycle. Cooling systems need replacement every 15 to 20 years. CREFC E-Primer, p.6. A landlord under a long NNN lease must budget for these major capital replacements even though the tenant pays operating expenses, because the lease may not cover full replacement of core systems. Counsel should check whether the lease reserves or amortization schedules match these lifecycles.

Key takeaways

• Anchor tenant leases are the foundation of AI data center finance. A 10 to 20 year lease with a creditworthy tenant unlocks debt and securitization markets.
• Know the three lease structures. NNN for hyperscale, MG+E for wholesale colocation, and full service gross for retail. Pricing and risk allocation differ sharply.
• Pricing is per kilowatt per month, not per square foot. Power pass throughs can make up 30 to 50 percent of revenue. PUE caps protect the tenant from cooling inefficiency.
• Critical provisions include SLAs with five nines availability, expansion rights, improvement waivers, credit support, and colocation sublicense rights. Document these early.
• Financing comes through SPVs, GPU backed loans, bonds with lease backstops, and CMBS. The credit quality of the anchor tenant often drives the rating, not the developer’s own balance sheet.
• Watch for asset liability mismatch, revenue shortfall relative to capex, opaque off balance sheet structures, and utility interconnection delays. These risks can unwind a deal that looks solid on paper.
• Real world deals like TeraWulf/Google ($16.2 billion contracted payments), Stargate ($500 billion ambition), and AWS/OpenAI ($38 billion) show the size and complexity of the market. Each deal structure carries lessons for counsel.

Frequently asked questions

Q:What is an anchor tenant in an AI data center?

A:
An anchor tenant is the primary, and often sole, occupant of a large AI data center building or campus. The anchor tenant, usually a hyperscale cloud provider, commits to a long term lease and takes most of the power capacity. The lease with the anchor tenant is what makes the project financeable.

Q:What is the difference between a wholesale colocation lease and a retail colocation lease?

A:
Wholesale colocation involves a few tenants each leasing significant power capacity, typically 5 to 40 megawatts, with 5 to 10 year terms. Retail colocation serves many small tenants each taking under 10 megawatts with 1 to 3 year terms. The lease structure moves from modified gross to full service. CREFC E-Primer, p.4.

Q:What does triple net (NNN) mean in an AI data center lease?

A:
In a triple net lease, the tenant pays base rent plus all property related expenses. Those include property taxes, building insurance, and common area maintenance. The tenant also pays all utility costs independently. The landlord is responsible for the building shell and core infrastructure. CREFC E-Primer, p.5.

Q:Why are AI data center leases priced per kilowatt instead of per square foot?

A:
Power is the binding constraint on computing, not floor space. A kilowatt per month pricing model directly ties rent to the tenant’s ability to deploy servers and reflects the developer’s cost structure, which is dominated by power delivery and cooling.

Q:What is a PUE cap and why does it matter?

A:
PUE stands for Power Usage Effectiveness. It measures total facility energy divided by IT equipment energy. A cap in the lease, such as 1.4, limits the amount of wasted energy the tenant must pay for. If the AI data center’s cooling system is inefficient and PUE goes above the cap, the landlord absorbs the extra electricity cost. CREFC E-Primer, p.5.

Q:Can an anchor tenant sublease its space to other users?

A:
Yes, many anchor tenant leases allow subleasing, often without the landlord’s consent beyond simple notice. This lets the anchor tenant run a colocation business inside its leased space. The Applied Digital/CoreWeave lease explicitly permits such colocation activity. Applied Digital/CoreWeave lease, SEC Exhibit 10.1.

Q:How does a credit backstop work in an AI data center lease?

A:
A credit backstop is a promise by a third party, often a parent company or a large customer, to cover lease payments if the tenant defaults. For example, Google backstopped Fluidstack’s leases at TeraWulf’s sites. If Fluidstack defaults on payment or becomes insolvent, Google has the option to pay all rent due and assume the lease at a discounted rate or terminate and pay a termination fee, backstopping the landlord in support of project debt. TeraWulf Form 8-K, Oct. 28, 2025.

Q:What is the biggest risk in financing an anchor tenant lease?

A:
The largest structural risk is the asset liability mismatch. The hardware inside the AI data center becomes obsolete in a few years, but the development loan and the lease run much longer. When the lease ends, the developer may face a large refinancing need without a guaranteed renewal, especially if technology or demand shifts. Policy analysis on data center finance, Nov. 2025.

Q:Are off balance sheet SPV structures risky?

A:
They can be. A captive SPV keeps lease obligations off the parent company’s balance sheet, but the parent’s support may be limited to a footnote agreement to cover revenue shortfalls up to a cap. If the project underperforms, lenders may find the parent’s commitment less solid than a full guarantee. Counsel should examine the specific support documents.

Q:What cooling specifications are common in modern AI data center leases?

A:
Most AI focused leases require liquid cooling, specifically direct to chip cooling with closed loop systems and coolant distribution units. The lease may also set commissioning standards, such as Level 5 testing, before the tenant begins paying rent. Applied Digital/CoreWeave lease, Schedule E-1.

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.