Data Center Feasibility Study: Colocation, Enterprise, Edge and Hyperscale Shells

A data center feasibility study tells a lender whether one specific data center can get the power it needs, on the terms the utility will offer, and lease or operate well enough to pay its debt at the coverage the capital source requires. The building matters less than it does for most property types. The power contract, the tariff, the tenant and the timing carry the file.

Four facts frame almost every data center loan request in late 2026:

• Construction has never moved this fast. Private data center construction ran at a seasonally adjusted annual rate of $96,462 million in September 2026 (preliminary), up from $50,700 million a year earlier and $1,922 million in September 2014, the first September Census reports [1].
• Electricity use has roughly tripled in a decade. Data centers used 58 TWh in 2014 and 176 TWh in 2023, about 4.4% of U.S. electricity, and a federal laboratory estimate puts 2028 between 325 and 580 TWh [8][9].
• Utilities now write rules for large loads. Georgia, Kansas, Missouri and Pennsylvania regulators have approved large-load terms with longer contracts, minimum bills, collateral or study deadlines since January 2025 [18][19][21][22].
• Federal grid rules are in motion. The Federal Energy Regulatory Commission opened a co-location investigation of the largest regional grid operator in February 2025 and has a large-load interconnection docket open since October 2025 [12][15].

This page sets out what the study must prove for each kind of data center, how each capital source reads the file, and where the public data a lender will recognize can be found. It covers colocation (retail and wholesale), enterprise facilities, edge sites and powered or hyperscale shells. It also includes a 50-state panel built from Census and Energy Information Administration data (Exhibit 21) and a calculator that turns critical IT load, power usage effectiveness and the power price into an energy cost and a coverage ratio.

Every study ends in one of three conclusions: feasible as proposed; feasible with stated conditions, such as a signed utility agreement, pre-leasing, a smaller first phase, more collateral or a larger reserve; or not feasible as proposed, with the changes that would make it work.

What is a data center feasibility study?

It is an independent analysis, written for the lender's file, of whether a proposed or existing data center can support its financing. It covers power supply and its cost, demand for space and capacity, competing supply, the site and its permits, the building and its cost, the tenant or operator, the financing structure and the cash flow that pays the debt. For the difference between this kind of report and other documents in a loan file, see what makes a feasibility study bankable, feasibility study vs. appraisal and feasibility study vs. business plan.

Lenders see five kinds of data center projects, and each shifts the weight of the study:

1. Retail colocation. Many customers lease cabinets, cages or small suites and buy power, cooling, security and network access as a service.
2. Wholesale colocation. Larger suites or whole halls leased to a few customers, often with power billed at or near cost.
3. Enterprise data centers. Facilities a company builds or buys to run its own systems; the operating company's credit decides the loan.
4. Edge sites. Small facilities close to users or network nodes, resting on latency, routes and a narrow local market.
5. Powered and hyperscale shells. Buildings, sometimes with substations and cooling plant, delivered to one large user that installs its own equipment; the lease, the tenant's credit and the power contract carry the loan.

The Lawrence Berkeley National Laboratory, in its 2024 report to Congress, sorts the same market into space types that run from small edge rooms to large-scale colocation and hyperscale modules. It modeled average module sizes of 11,000 square feet for large colocation spaces and 30,000 square feet for hyperscale facilities, measured per module rather than per campus [9].

Exhibit 1. What the lender asks and what the study shows

Source: 13 CFR Part 120 and 121.201 (revised as of January 1, 2026); SBA SOP 50 10 8.1; 7 CFR Part 5001 (revised as of January 1, 2026), as cited in each row.

Many data center sites start as industrial land; the industrial feasibility study page covers the warehouse and plant tests that a powered shell shares.

What does a lender need the study to prove?

A lender needs proof that rent or operating cash flow repays the loan under the capital source's rules, and that the power supply, the site and the tenant do not break the projections. SBA's regulation states the standard plainly: "Loans must be so sound as to reasonably assure repayment" [41]. The same regulation allows SBA to require "a survey, or a feasibility study" [41].

SBA's procedures name the situations that may prompt it to ask for a study. Three fit data centers closely: a "highly specialized Project property", a "project size disproportionate to size of community it will serve" and "market saturation by industry type and location" [43]. The procedures also say that independent reports "may be beneficial in mitigating any weaknesses identified in the credit analysis" [43]. USDA goes further. For Business and Industry guarantees above $1,000,000 to a new entity or new activity, a study by an independent qualified consultant acceptable to the Agency is required, and the Agency defines the study as an evaluation of "economic, market, technical, financial, and management feasibility" [44]. A summary of how bank regulators, SBA and USDA each treat these reports is in lender feasibility study requirements.

For a data center, the proof falls into five questions:

1. Power. Will the utility deliver the capacity on the date the projections need it, under which tariff, with what minimum bill, collateral and exit terms?
2. Demand. Who will lease or use the capacity, at what rent per kilowatt, and how fast?
3. Supply. What competing capacity is operating, under construction or holding a utility commitment in the same market?
4. Cost and timing. What will the building, the electrical and cooling plant and the network connections cost, and when can each be energized?
5. Coverage. After power, operating costs and reserves, does cash flow cover debt service in the base case and in the downside cases the lender cares about?

The weight shifts with the project type: absorption and the utility schedule for speculative colocation; the lease, tenant credit and re-leasing case for a powered shell; the operating company's need and cost savings for an enterprise facility.

How fast is data center construction growing?

Faster than any other private nonresidential category Census tracks. Census reports data center construction as its own line inside the office category, and the series starts in 2014. In September 2014 the seasonally adjusted annual rate was $1,922 million. By September 2026 it was $96,462 million, about 50 times larger, and up 90.3% from September 2025 [1].

Exhibit 2. Private data center construction, September of each year, 2014 to 2026

Source: U.S. Census Bureau, Value of Private Construction Put in Place, seasonally adjusted annual rate, release of November 2, 2026 [1]. Shares calculated by FSC.

Two things stand out. First, data centers are now two-thirds of private office construction by value. A lender who reads "office construction is rising" in a national summary is mostly reading data center spending. Second, the other industrial categories moved the other way. Warehouse construction fell 35.5% from September 2023 to September 2026, and manufacturing construction fell 32.9% from its September 2024 level [1]. Capital, contractors and electrical equipment have shifted toward data centers.

The pace accelerated through 2026. Spending rose 73.9% from December 2025 to September 2026, with monthly gains of 10% or more in June, July and August [1].

Exhibit 3. Private data center construction by month, December 2025 to September 2026

Source: U.S. Census Bureau, Value of Private Construction Put in Place, seasonally adjusted annual rate, release of November 2, 2026 [1]. Changes calculated by FSC; recent months will be revised.

Construction spending measures money going into the ground, not demand being met. For a specific market the study asks whether new supply is already leased or owner-built, whether equipment and labor will be available on schedule, and whether earlier projects will claim the utility's capacity first.

How big is the data processing and hosting industry, and where is it?

County Business Patterns counted 18,544 establishments and 629,527 mid-March employees in NAICS 518210 in 2023, with annual payroll of $92.96 billion [2]. That industry code, retitled in the 2022 NAICS as computing infrastructure providers, data processing, web hosting and related services, is the one SBA uses for the size standard in Exhibit 1 [42]. It counts business establishments classified by their main activity, not buildings. A colocation operator, a hosting firm and a cloud computing operation can all fall in it, and the count can miss facilities run by companies whose main business is something else. The panel is still the best public measure of where the industry employs people.

Most establishments are small. Two in three have fewer than 10 employees, and only 526 have 250 or more [2].

Exhibit 4. NAICS 518210 establishments by employment size class, United States, 2023

Source: U.S. Census Bureau, County Business Patterns 2023, cbp23us, NAICS 518210 [2]. Shares calculated by FSC.

Employment is concentrated. California alone holds 19.7% of the industry's employees, and five states (California, Texas, New York, Virginia and Florida) hold 45.1% [3]. Payroll per employee varies widely among the leading states, from $111,869 in North Carolina and $112,836 in Pennsylvania to $247,172 in New York, which reflects the mix of engineering, software and facility jobs counted in each state.

Exhibit 5. Top 12 states by NAICS 518210 employment, 2023

Source: U.S. Census Bureau, County Business Patterns 2023, cbp23st, NAICS 518210, all legal forms [3]. Payroll per employee calculated by FSC.

A study uses these counts in three ways. They show whether a market has the trained workforce a facility needs to staff around the clock. They help size the local base of hosting and network businesses that lease retail colocation. And county-level counts, from the same dataset, give a first read on whether an edge site's local market is real or thin. They do not measure megawatts, and a study should never treat employment as a proxy for capacity.

How much electricity do data centers use?

About 176 TWh in 2023, or 4.4% of all U.S. electricity, according to the Lawrence Berkeley National Laboratory report released by the Department of Energy in December 2024 [8][9]. Use was flat at about 60 TWh from 2014 to 2016, reached 76 TWh in 2018 and then more than doubled between 2017 and 2023 as servers with graphics processors for artificial intelligence spread through the installed base [9].

Exhibit 6. U.S. data center electricity use, 2014 to 2028

Source: U.S. Department of Energy release of December 20, 2024 [8]; Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report [9].

The 2028 range is wide on purpose. The laboratory built it from scenarios for chip shipments, how hard the installed servers run and how they are cooled [9]. A study should cite the range, not a point inside it, and should not use a national range to justify demand in one market. What the range does tell a lender is that the national load is large enough to strain transmission and generation in the places where data centers cluster. The Department of Energy said as much in its September 2025 request for information, describing data centers and other large loads that "are driving multi-gigawatt increases in electric demand" [17].

Power usage effectiveness (PUE) is the ratio of all the power a facility draws to the power its IT equipment uses. A PUE of 1.40 means the building draws 1.4 kilowatts for every kilowatt delivered to servers, storage and network gear; the extra 0.4 goes to cooling, power conversion, lighting and losses. The laboratory estimates that the national average PUE fell from 1.6 in 2014 to 1.4 in 2023, largely because more servers moved into larger hyperscale and colocation buildings [9].

Exhibit 7. Efficiency and water estimates for U.S. data centers

Source: Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report [9].

Two cautions belong in any study that leans on a design PUE. The laboratory's figures come from simulations that "assume systems are commissioned and operate as designed", and the report adds: "This is rarely the case" [9]. Its simulations also assumed uninterruptible power supply efficiencies as low as 77% to 85% for small spaces [9]. A study should test the cash flow at a PUE worse than design, at least through the first years of operation when halls are partly filled and cooling plant runs below its best point.

Is power available, and on what terms?

Power availability is a finding, not an assumption. The study must show, from the utility's own documents, how much capacity the site can draw, from what date, under which tariff and on what conditions. A letter that says a utility is "studying" a request is not a capacity commitment, and the study should say so in plain words.

The power section of a data center study answers these questions:

1. Capacity and date. What load will the utility serve, in what steps and from what dates, and which upgrades must finish first?
2. Tariff and contract. Which rate schedule applies, with what minimum bill, term, ramp, exit fee or collateral (Exhibit 10)?
3. Cost responsibility. Who pays for the substation, line extension and network upgrades, and when?
4. Queue position. Where does the project sit among other large loads seeking the same capacity?
5. Backup supply. What generators, batteries or on-site generation are planned, under which permits and limits?
6. Price path. What will the delivered price be, including demand charges and riders?

Public data can frame the utility's answers. EIA's state tables show where retail electricity sales and generating capacity are growing fastest. Commercial-sector retail sales rose 3.3% nationally in the first eight months of 2026 compared with the same period of 2025, but 14.4% in Ohio, 14.2% in Arizona, 13.8% in Indiana and 18.4% in Nebraska [5]. Data centers are not the only cause, and EIA does not break them out, but a lender should expect the utility in a fast-growing state to be writing tighter large-load terms.

Exhibit 8. Largest increases in commercial retail electricity sales, January to August 2026 vs 2025

Source: U.S. Energy Information Administration, Electric Power Monthly, Table 5.4.B, data through August 2026, released October 23, 2026 [5][7]. Changes calculated by FSC.

A note on reading this table: EIA reports sales by the sector each utility assigns a customer to. A data center may be billed as a commercial or an industrial customer depending on the utility and the tariff, so the study should confirm the rate class with the utility rather than infer it from the table. Virginia's commercial sales rose only 2.6% from a much larger base; the absolute increase still ranks seventh.

Generating capacity is moving too. Net summer capacity from all sources rose 57,152.5 MW nationally between August 2025 and August 2026, and Texas alone added 16,238.0 MW, 28.4% of the national gain [6].

Exhibit 9. Largest increases in net summer generating capacity, August 2025 to August 2026

Source: U.S. Energy Information Administration, Electric Power Monthly, Table 6.2.A, utility-scale units, all sources, August 2026 and 2025 [6]. Changes calculated by FSC.

New capacity in a state does not mean deliverable capacity at a site. Transmission between new generation and a substation near the project can be the binding limit, and that is a question only the utility and the grid operator can answer. The table tells a lender where to expect more supply over time; the utility letter tells the lender what the project can actually draw.

What do large-load tariffs require from a data center?

They shift the cost and risk of serving a very large new customer onto that customer. Since January 2025, regulators in at least four states have approved terms that apply above a load threshold, lengthen contracts, set minimum bills and require collateral. A study must read the project's actual tariff, but the pattern is consistent enough that a lender can expect these items.

Exhibit 10. Large-load terms approved or adopted by state regulators

Source: Georgia Public Service Commission [18]; Kansas Corporation Commission [19]; Missouri Public Service Commission [20][21]; Pennsylvania Public Utility Commission [22]; Proclamation 11014, 91 FR 11439 [11]. The pledge is a voluntary commitment, not a tariff.

The pledge deserves a short explanation because lenders ask about it. Proclamation 11014 records that on March 4, 2026, seven technology companies accepted the terms of a "Ratepayer Protection Pledge" [11]. Under it, those companies say they will build, bring or buy new generation, "pay for all new power delivery infrastructure upgrades to service their data centers" and pay negotiated rates and infrastructure costs "whether they use the electricity or not" [11]. The pledge binds the companies that signed it. It does not change any utility's tariff, and it does not bind a developer, a colocation operator or a smaller user. A study for a project leased to a signatory should still read the actual utility agreement.

What these terms mean for the cash flow:

• Minimum bills become fixed costs. An 80% minimum on contracted demand in Kansas, or a demand charge set at 80% of the agreed rate in Missouri, is owed whether or not the halls fill on schedule [19][21]. In a colocation model, that cost sits with the operator until tenants arrive, unless the leases pass it through.
• Collateral ties up equity. Two years of minimum bills in Kansas and Missouri is a cash or credit commitment that belongs in sources and uses [19][21].
• Contract terms outlast many leases. A 17-year utility commitment can run longer than the first lease term of a wholesale tenant. The study should show who carries the minimum bill if that tenant leaves.
• Exit costs are large. In Kansas an early exit pays every remaining minimum bill; in Missouri the customer must give 24 months' notice and pay an exit fee [19][21].
• Thresholds matter by phase. A small first phase may fall below a threshold that later phases cross.

How are FERC co-location and large-load proceedings changing interconnection?

They are moving the rules for connecting large loads, including loads that sit next to a power plant, from case-by-case negotiation toward written tariff terms. Most retail service terms remain with states, but where a data center connects to the interstate transmission system or shares a site with a generator, federal proceedings now matter to the schedule and the cost.

Exhibit 11. Federal proceedings on large loads and co-location, 2024 to 2026

Source: Federal Register, as cited in each row [10][12][13][14][15][16][17].

The regulatory agenda describes the rulemaking's aim as reforms that would "standardize interconnection procedures and agreements" for large loads, "including those seeking to share a point of interconnection with new or existing generation facilities" [15]. Until a final rule or tariff change takes effect in the project's region, the study should treat interconnection terms as the ones in the utility's and grid operator's current documents, and it should name the open dockets that could change them. The study should also say whether the project relies on co-location with a generator, because that structure has been under active review in the largest regional market since February 2025 [12].

Executive Order 14318 is a permitting order, not a power order. It defines a "Data Center Project" as a facility that "requires greater than 100 megawatts (MW) of new load dedicated to AI inference, training, simulation, or synthetic data generation" [10]. It sets up an initiative to support qualifying projects, steers agencies toward categorical exclusions under the National Environmental Policy Act, directs EPA to identify brownfield and Superfund sites for these projects, and revokes Executive Order 14141 of January 14, 2025 [10].

Exhibit 12. Executive Order 14318 definitions that a study may need to apply

Source: Executive Order 14318 of July 23, 2025, 90 FR 35385 [10].

Most projects financed with SBA or USDA support are far below these thresholds. The order matters to them mainly through the faster environmental reviews and the brownfield guidance it calls for. For larger projects, the study should say whether the project meets a qualifying test and what federal permits it actually needs.

How do power price and PUE change the cash flow?

Directly and in proportion. Annual energy cost equals the IT load, times PUE, times 8,760 hours, times the price per kilowatt-hour. Every one of those terms is a study input, and two of them, PUE and price, are where optimistic projections hide.

State average prices give a starting range, not a site price. In the first eight months of 2026 the average industrial price ran from 5.19 cents per kilowatt-hour in New Mexico to 37.83 cents in Hawaii, against a national average of 9.09 cents, up 6.9% from 8.50 cents a year earlier [4]. The national commercial average was 14.04 cents, up 5.6% [4].

Exhibit 13. The 12 states with the lowest average industrial electricity price, January to August 2026

Source: U.S. Energy Information Administration, Electric Power Monthly, Table 5.6.B, average price to ultimate customers in cents per kilowatt-hour, year to date through August 2026 and 2025 [4].

State averages blend many customers, from small plants to large contract loads. A data center's delivered price depends on its tariff, its load factor, demand charges, riders and any special contract. The study should build the price from the tariff sheets and the utility's terms, then show what happens if it rises. In the calculator's default case, each one-cent increase in the price adds $637,728 a year in energy cost for 8 MW leased at 70% average draw and a PUE of 1.30.

Who pays the energy bill matters as much as its size. When leases pass power through at cost, the tenant carries price and PUE risk; when contracts bundle power into a fixed charge, the operator does. The study models the lease structure the operator will actually sign.

The calculator below tests this arithmetic. It turns leased megawatts, average IT draw, PUE and power price into annual energy cost, then adds rent per kilowatt per month and the share of energy cost recovered from tenants to reach net operating income and a debt service coverage ratio. It also solves for the rent needed to hit a target coverage. Its inputs are illustrative defaults chosen to make the arithmetic easy to follow, not market figures, and its result is not a feasibility conclusion. For how SBA and USDA define and test coverage, see debt service coverage ratio for SBA and USDA loans.

How is demand measured for colocation, enterprise, edge and hyperscale projects?

By the capacity tenants will contract for, in kilowatts or megawatts of critical IT load, and by the rent they will pay per kilowatt per month. Square feet matter for the building budget; kilowatts matter for revenue. A demand section that counts square feet and applies an office-style rent per square foot has measured the wrong thing.

The evidence differs by project type:

Exhibit 14. What demand evidence looks like for each kind of data center

Source: FSC method; space types follow Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report [9].

For speculative colocation, the market survey is the heart of the demand section. It identifies every operating and planned facility within the competitive area, its capacity in critical megawatts, its occupancy where it can be learned, and its power status. Planned supply is sorted into four groups: under construction with power secured, under construction without a power date, announced with a utility commitment, and announced without one. Only the first two belong in the near-term supply forecast; the others are risks to rent and lease-up later.

Absorption is then estimated in kilowatts per year, from the leasing history of comparable buildings and from tenants in the market that are known to be seeking space. The study converts absorption into a lease-up schedule for the subject, net of the supply that will deliver at the same time. If the schedule needs the subject to capture a larger share of absorption than its quality, location and power date support, the study says so.

How does the study test a single-tenant hyperscale lease?

As a credit exposure first and a real estate question second. When one tenant leases the whole building, the loan is repaid by that tenant's rent until the lease ends or the tenant leaves. The study reads the lease and the tenant's financial condition with the same care a lender would give a corporate borrower.

The lease review covers:

• Term and options. A lease shorter than the loan's amortization leaves a re-leasing gap the study must price.
• Rent and escalations. Rent per kilowatt, escalators and which costs the tenant pays directly.
• Commencement. A rent start tied to energization moves the utility's schedule into the lender's cash flow.
• Termination and contraction. Early exit, give-back of capacity and remedies for late delivery.
• Assignment and guarantee. Whether a parent guarantees the lease and whether it can pass to a weaker affiliate.

The coverage test is then run twice: with the tenant paying under the lease, and with the tenant gone at the earliest date it could leave. In the second case the study estimates downtime, re-leasing cost and the rent a new tenant would pay for a building designed around another company's equipment. If the building is so specialized that re-leasing is not realistic, the study says the loan depends entirely on the tenant's credit, and the lender can decide on that basis.

The utility contract must line up with the lease. If the landlord signs a 17-year utility commitment with an 80% minimum bill, as in the Kansas tariff, and the tenant's lease runs 10 years, someone carries that minimum for seven years after the lease ends [19]. The study names that party.

What site, fiber, water and zoning issues decide feasibility?

Five site issues decide most data center files: power delivery, fiber routes, water and cooling, zoning and community approval, and environmental condition. Power is covered above. The other four are below.

Fiber and latency. Tenants need diverse routes, more than one carrier and, for retail colocation and edge sites, low latency. The study maps the routes, the carriers, the cost and time to build laterals and whether routes are physically separate; signed carrier commitments outweigh discussions.

Water and cooling. Cooling design sets both the PUE and the water draw. The laboratory found that direct water use by U.S. data centers rose from 21.2 billion liters in 2014 to 66 billion liters in 2023, with hyperscale and colocation facilities using 84% of the 2023 total [9]. It found that water-cooled chiller systems without economizers show the highest water use, and that airside economizers allow chilled water systems to shut down in good weather [9]. The study checks the water supply, the utility's capacity and charges, any local limits on consumption and the trade-off between water and power in the chosen design.

Zoning and community process. The study confirms whether the use is permitted by right, conditional or in need of rezoning, and lists the approvals still required: site plan, building, stormwater, generator air permits and utility easements. Local concerns usually center on noise, building mass, water, construction traffic, tax revenue and grid capacity. The study reports hearings ahead and conditions accepted; it does not predict how an open public process will end.

Environmental condition and reuse sites. Former industrial and utility sites often come with power and transmission nearby, which is why they attract data centers. Executive Order 14318 directs EPA to identify brownfield and Superfund sites for qualifying projects and to issue guidance within 180 days to speed their reuse [10]. A study for a reuse site reads the environmental reports, the remediation status and any land-use limits recorded against the property.

Which environmental permits apply to backup generators?

Two federal rules govern most diesel backup engines at data centers: the new source performance standards for stationary compression ignition engines (40 CFR Part 60, Subpart IIII) and the national emission standards for stationary reciprocating engines (40 CFR Part 63, Subpart ZZZZ). Both treat an engine as an emergency engine only if it stays within strict operating limits [35][39].

Exhibit 15. Federal rules that bind data center backup engines

Source: 40 CFR Parts 60 and 63 (revised as of July 1, 2025), as cited in each row; definition of emergency stationary RICE at 40 CFR 63.6675 [40].

These limits drive three parts of a study. First, the operating plan: a design that counts on running generators to shave peaks, sell power to the grid or sit out utility curtailments for long periods may lose emergency status and fall under the much stricter non-emergency standards [35]. Second, the permit: the number and size of engines can push a campus toward major-source status, which brings permitting cost and time. Third, the utility contract: if a tariff or service agreement asks the facility to curtail during grid emergencies, the study must check that the curtailment plan fits the engine limits. State and local air permits add their own conditions, which the study lists for the site.

What state tax incentives apply to data center equipment?

Many states exempt data center equipment, and sometimes electricity, from sales and use tax, but only for projects that meet investment, job and wage tests. The exemptions are large enough to change the equipment budget and the operating budget, so the study confirms which tests the project meets and when.

Exhibit 16. Data center sales and use tax incentives in eight states

Source: Code of Virginia 58.1-609.3 [23]; NRS 360.754 [24]; Minnesota Statutes 297A.68, subd. 42 [25]; RSMo 144.810 [26]; N.C. G.S. 105-164.13(55a) and 105-164.3(201) [27][28]; Iowa Code 423.3(95) [29]; Georgia Department of Revenue [30][31]; A.R.S. 41-1519 [32]. Summaries; read the statute for every condition.

Four points keep these incentives from distorting a study:

1. They are earned, not assumed. Most require a certificate or agreement and can be clawed back if targets are missed; the study shows the cost if the project falls short.
2. They follow the equipment buyer. Virginia applies the exemption to the operator and its tenants if they meet the tests together [23]. Arizona extends relief to qualified colocation tenants [32]. A colocation study must show which party buys which equipment, because a tenant's exemption does not reduce the operator's budget.
3. They change over time. Iowa narrowed the electricity and fuel exemptions for construction completed on or after June 6, 2025 [29]. Minnesota added a large-scale tier for investment periods starting after June 30, 2025 [25]. The study cites the version in force on its date.
4. Local taxes still apply. Property taxes and local abatement agreements vary by county and city.

How do depreciation rules affect a data center budget?

They change after-tax cash flow, not the debt service coverage most lenders test, but they matter to equity and to an owner-operator's projections. Federal law treats a data center as two very different asset classes: the building and its structural components, recovered over 39 years, and servers and other computer equipment, which are 5-year property as qualified technological equipment [45]. Public Law 119-21 made the 100 percent special depreciation allowance permanent for property acquired after January 19, 2025 [46].

Exhibit 17. Federal recovery rules that apply to data center assets

Source: 26 U.S.C. 168 (2024 edition) [45]; Public Law 119-21 [46]. Cost segregation and the classification of electrical and mechanical plant are matters for the owner's tax adviser.

For a study, the practical point is the replacement cycle. Servers and network gear owned by an operator wear out and become obsolete far faster than the building. A projection that shows strong cash flow but no reserve for replacing owned equipment overstates what is available for debt service over a long loan.

Which financing paths fit a data center?

It depends on the borrower and the size. Most large data centers are financed with conventional construction loans, private credit or the balance sheets of their owners and tenants, and those lenders set their own terms. Government-guaranteed programs fit a narrower set: operating businesses that own their facilities, smaller colocation and hosting firms, edge sites and enterprise facilities in rural areas. The study is written for whichever path the borrower takes, and it reads the program rules that come with SBA or USDA money.

Exhibit 18. Financing paths and the rules that shape a data center study

Source: 13 CFR Part 120 (revised as of January 1, 2026) [41]; SBA SOP 50 10 8.1 [43]; 7 CFR Part 5001 (revised as of January 1, 2026) [44]. Conventional and private terms vary by lender.

SBA 7(a)

SBA's 7(a) program fits an operating business that runs the data center itself. The first question is whether the borrower is an operating business at all. SBA excludes "passive businesses owned by developers and landlords that do not actively use or occupy the assets acquired or improved with the loan proceeds" [41]. A colocation operator that runs the facility, sells power, cooling, security and network services and employs the staff is a different business from a landlord that leases an empty powered shell. The study describes what the borrower actually does every day, who controls operations, and whether any management agreement gives a third party sole discretion, which SBA treats as making the applicant an ineligible passive business [43].

The occupancy rules then apply. For new construction, the borrower must occupy at least 60% of the rentable property and may lease up to 20% long term; for an existing building, the borrower must occupy at least 51% and may lease up to 49% [41]. How those rules apply to a colocation operator, whose customers occupy cabinets inside space the operator controls and runs, is a judgment the lender makes on the facts; the study supplies the facts, including how the space is measured, who controls it and what the customer contracts say. More detail on the current procedures is on our page about the SOP 50 10 8.1 update and the SBA loan programs.

Coverage must be at least 1.15 on a historical or projected basis and 1:1 globally [43]. For start-ups and construction projects, the projections must show 1.15 within 2 years of funding or of the end of construction [43]. Rental income from the project property may count in the global analysis but not in the repayment analysis [43].

SBA 504

The 504 debenture limit rises from $5,000,000 to $5,500,000 per project when the operating company's energy use falls by at least 10% [41]. A data center that replaces an older, less efficient facility may document such a reduction, and the study can supply the energy model.

SBA asks whether the project property is limited or special purpose, defined as "a limited-market property with a unique physical design, special construction materials, or a layout that restricts its utility to the use for which it was built" [43]. SBA's list of examples does not name data centers, but it states that the list "is not intended to be all-inclusive" [43]. If the property is limited or special purpose, the borrower must contribute at least 15%, or 20% if the project also involves a new business [43]. The study gives the lender the evidence on design and alternative use so the lender can decide.

USDA Business and Industry

USDA's Business and Industry guarantee fits rural projects that create or save jobs. The site must be outside any city or town of more than 50,000 and its contiguous urbanized area [44]. The borrower limit is $25 million [44]. A study by an independent qualified consultant is required for guarantees above $1,000,000 to a new entity or new activity [44]. For B&I projects above $1 million that increase direct employment by more than 50, USDA applies a market capacity test against existing competitors [44]. Most data centers employ fewer people than their cost suggests, so the jobs test is often met easily, but the study should still show the job count and the market analysis. See USDA loan programs.

Conventional and private capital

Banks, private credit funds and large tenants finance most large projects, on terms negotiated rather than published. A construction lender wants the power schedule, pre-leasing and cost to complete; a term lender wants stabilized coverage and the re-leasing case. See conventional loan programs.

What downside cases should the study run?

At least one for each assumption that can fail. A data center's cash flow is unusually sensitive to dates, because rent often starts only when the utility energizes capacity, and to fixed power commitments, because minimum bills run whether or not tenants arrive. The study runs each case through the same model as the base case and reports coverage and any equity or reserve needed to hold it.

Exhibit 19. Downside cases for a data center study

Source: FSC method; minimum bill and term examples from the Kansas Corporation Commission and the Missouri Public Service Commission [19][21]. Case sizes are model inputs, not market figures.

The calculator's worked cases show how quickly coverage moves. In its default case, coverage is 1.22. With power fully passed through to tenants, coverage rises to 2.07; with a PUE of 1.60 and a higher power price the operator absorbs, it falls to 0.50; and with only 5 MW leased instead of 8, it falls to 0.64. Those results come from illustrative inputs, but the pattern holds for real projects: who pays for power and how fast capacity leases matter more than almost anything else in the model.

What must a data center feasibility study contain?

It must contain enough evidence for a credit officer who has never seen the site to follow every number to its source and to see the cash flow under stress. A study for a data center usually includes the parts in Exhibit 20.

Exhibit 20. Contents of a data center feasibility study

Source: FSC method; SBA SOP 50 10 8.1 and 7 CFR 5001 for the program content noted in each part [43][44].

USDA's rule describes the study as an evaluation of "economic, market, technical, financial, and management feasibility" [44]. For a data center, "technical" means the power and cooling plan as much as the building, and "management" means the team's record of running a facility around the clock.

How does the engagement run, and what comes after closing?

The engagement follows the order of risk: power first, then demand, then the rest. Scope depends on the project type, the capital source and how much of the utility and leasing work is already documented. This page describes scope and deliverables only; the proposal sets out the rest.

1. Scoping. Confirm the capital source, the lender's requirements and the documents available: utility agreements, leases or letters of intent, design, budget and permits.
2. Power review. Read the utility documents and tariff; list what is committed, pending or open to change.
3. Market and site work. Survey competing supply, network routes, the site and remaining approvals.
4. Model. Build the budget, the operating projection and the coverage tests, including downside cases.
5. Draft, review and final. Correct factual errors raised in review without changing independent conclusions, then deliver the signed study.

After closing, some lenders ask for periodic monitoring: construction progress against the power schedule, leasing against the projection, and actual PUE and power cost against the model. The same evidence base supports that work. The study is signed by Sarrah Allen, MAI. More about the firm is on the about page, and other property types are on the asset classes page.

How do the states compare on data center fundamentals?

Exhibit 21 brings the main public series together for every state. It shows the industry's establishments and employees from County Business Patterns, average commercial and industrial electricity prices, the change in commercial retail sales from the first eight months of 2025 to the same months of 2026, and total net summer generating capacity. Read it as a first screen. A site's power terms come from its utility, not from a state average.

Exhibit 21. Data center fundamentals by state

Source: U.S. Census Bureau, County Business Patterns 2023, NAICS 518210 [2][3]; U.S. Energy Information Administration, Electric Power Monthly, Tables 5.4.B, 5.6.B and 6.2.A, data through August 2026 [4][5][6][7]. Changes calculated by FSC. CBP counts for small states carry Census noise flags; employment is mid-March 2023.

The lowest-price states are not all the states with the most data processing employment; price is one input among network, workforce and utility capacity [3][4]. Commercial sales grew fastest in Nebraska, Ohio, Arizona and Indiana, each above 13% [5], and the largest capacity additions were in Texas, Arizona and New Mexico [6].

Download the evidence behind this page's exhibits (CSV, 821 data points with sources)