Data Center Underwriting Method: Power, Tenant Commitments and a Thin SBA Record

A building sold by the kilowatt

A data center is a building whose product is electricity delivered into a controlled room. Its customers pay for the right to draw a stated amount of power for their servers, storage and network equipment, together with the cooling that carries the heat away and the backup plant that keeps the load running when the grid does not. Floor area matters, but floor area is not what is sold. A feasibility study of a data center therefore opens with two questions that few other classes in this series ask in the same form: how much power the serving utility has actually committed to the site, and how much of that power customers have signed contracts to take. Nearly every other line of the cash flow follows from those two numbers.

The class has grown fast at the national level. The Department of Energy, summarising a 2024 report from Lawrence Berkeley National Laboratory, put data centers at about 4.4% of U.S. electricity use in 2023 and projected 6.7 to 12% by 2028 ^[1]. The same summary has data center electricity use rising from 58 TWh in 2014 to 176 TWh in 2023 ^[2]. Those are national quantities. They show a class in expansion; they say nothing about whether one building in one county will fill, and a study that cites them to support a project's absorption has answered a local question with a national trend.

This article sets out the method a study for the class should follow. It covers the demand model, built from utility capacity and signed contracts instead of population; the measures a credit officer will look for, with the documents behind each; the SBA, USDA and bank rules that bear on a data center borrower; the class's record in SBA's loan files, which is thin and, for reasons the next section gives, mostly not a record of buildings; the outcome floors that record shows; and a short worked illustration. The general methods are written elsewhere in the series: trade areas and capture rates, counting supply and choosing comparables, ramp-up and absorption and the coverage test against program floors. The site's archived scorecard of metropolitan markets on power, land and pipeline compares regions; the method here is the one a single project's loan file needs.

What code 518210 counts, and what it cannot separate

The loan dataset assigns the class to a single industry code, 518210, and the code's own definition explains much of what follows. NAICS describes the industry as establishments primarily engaged in providing computing infrastructure, data processing services, Web hosting and related services ^[3]. Its illustrative examples run from application hosting, cloud storage and infrastructure as a service to optical scanning services ^[4]. Nothing in the definition requires the establishment to own or operate a building. A firm that rents virtual servers out of another company's hall, a document scanning bureau and the operator of a purpose-built colocation facility can all carry the same code.

Colocation is placed in the code on purpose. The manual's cross-references send establishments primarily engaged in renting server and networking space in data centers to 518210, away from the real estate lessors ^[5]. Two neighbouring activities are classified elsewhere. Operating another company's data processing facility on that company's premises belongs to 541513, computer facilities management ^[6], and an owner acting as lessor of a nonresidential building belongs to 531120 ^[7]. On the manual's definitions, a landlord that leases a powered shell to an operator, with no colocation service of its own, reads as a lessor, and a loan coded as the manual directs would not appear in the data center class at all.

The consequence for the loan record is direct, and the series article on codes and classes treats the general problem: the data center class holds hosting and processing businesses beside colocation operators, and the dataset's asset map records that colocation buildings cannot be separated from hosting services inside the code. Every loan figure later in this article carries that limitation, and the text repeats it where the figures are printed.

Construction statistics draw a different line. In the Census Bureau's construction spending definitions, the office category covers the office types listed beneath it ^[8], and data center is one of them, covering buildings that contain the hardware needed for storing, processing and transmitting digital information ^[9]. The same definitions keep special purpose equipment out of construction and give racks or servers in data centers as an example ^[10]. The series therefore measures the shell and its building systems, not the computing equipment inside. The monthly release prints office as one line: private office construction ran at an annual rate of $145,633 million in September 2026, against $104,376 million a year before ^[11]. The data center part of that line cannot be read from the release, because general and financial office buildings sit in it too, and a study that cites the line as data center construction has claimed more than it shows.

Power is the unit of capacity

A data center's capacity is stated in kilowatts, and a study has to keep two units apart. The Department of Energy's metering guide defines a kilowatt-hour as the energy from one kilowatt of power used for one hour, and power as the rate at which energy is generated or used ^[12]. The method used here states both capacity and price per kilowatt of critical IT load, meaning the power the customers' equipment is entitled to draw. The energy bill, in kilowatt-hours, then depends on how much of that entitlement the customers actually use, how many hours they use it, and how much the building spends to deliver each kilowatt.

That last quantity is the overhead between the meter and the racks. Power usage effectiveness, or PUE, is the ratio of a facility's total energy use to the energy its IT equipment uses ^[13]. Total facility energy takes in the power delivery chain, from uninterruptible power supplies, switchgear and transformers to generators, distribution units and batteries ^[14], and the cooling plant, from chillers and cooling towers to the air handlers in the computer rooms ^[15]. At basic metering, PUE for a stand-alone facility is the utility meter reading divided by the output of the uninterruptible power supply ^[16]. The guide also records The Green Grid's caution that PUE is neither a productivity metric nor a complete efficiency measure ^[17]. It tells a lender how much power the building burns to deliver a kilowatt to the racks; it says nothing about whether the racks earn any revenue.

PUE is what converts utility capacity into sellable capacity. An agreement for 1,000 kW at the meter supports only that capacity divided by the design PUE at the racks, and the building can sell no more than the smaller of that figure and the load its electrical and cooling plant was designed to carry. The two limits come from different documents, the utility's agreement and the engineer's design, and a study should print both, say which one binds, and show what happens to the binding one if the design PUE proves optimistic. A building with spare floor area and no spare power has nothing more to sell, and a building with spare power and a full plant needs capital before it can sell any of it.

In the model below electricity is the one operating cost that moves with load, and its price differs by tariff class and by place. EIA's table of average retail prices by sector and state ^[18] puts the U.S. average at 14.47 cents per kilowatt-hour for commercial customers and 9.64 cents for industrial customers in August 2026 ^[19]. In Virginia the commercial average that month was 11.14 cents ^[20]. A sector average blends many tariffs and many load shapes. The illustration later in this article uses the national commercial figure only because it has no site. A study for a real facility should price energy from the serving utility's own schedule for the class the facility will be billed under, including demand charges, riders and any minimum bill, and should state the date of the schedule it used.

Reliability terms belong in the same chapter. Texas S.B. 6, as passed by the 89th Legislature in 2025 (the enrolled text is the version on file here), defines on-site backup generating facilities as generation that cannot export to the ERCOT grid and that can serve at least 50 percent of on-site demand ^[21], and the enrolled text provides that large load customers with such facilities may be directed, before or during an energy emergency, either to run that generation or to curtail load ^[22]. It also calls for a protocol, with any equipment installed before interconnection, that allows certain transmission-voltage loads to be curtailed during firm load shed ^[23]. A study of a project in a market with rules of this kind should state how many hours of generator operation the fuel storage, permits and customer contracts allow, and what the contracts promise customers if the grid operator orders the load curtailed.

The utility's commitment: capacity, ramps and minimum bills

Power availability is not a matter of a utility's goodwill. For a large load it can be a filed contract, with terms a lender has to read as carefully as any lease, and two public documents show what those terms look like.

The first is a tariff. AEP Ohio's Schedule DCT, on file with the Public Utilities Commission of Ohio, applies to a data center whose aggregate monthly maximum demand exceeds 25,000 kW ^[24]. Under it a new customer's contract capacity ramps in over a period that may not exceed four years ^[25], with the ramp capacity set no lower than 50% of the final contract capacity in the first year, 65% in the second, 80% in the third and 90% in the fourth ^[26]. During the ramp, billing demand may not fall below 85% of the ramp capacity ^[27]. A new contract runs at least the ramp period plus eight years, twelve years where the ramp lasts four ^[28], and for that initial term the customer owes the minimum charges "regardless of the customer choosing to curtail, reduce, suspend, or terminate service" ^[29]. After the fifth contract year following the ramp, a customer may leave by paying an exit fee of 36 months of minimum charges ^[30].

The same schedule asks for security and for land. A new-load customer without the specified credit ratings and liquidity ^[31] must post a guarantee or collateral equal to 50% of the total minimum charges for the full term ^[32], and the amount falls by one year's minimum charges for each year the customer is energized and pays on time ^[33]. To sign, the customer must designate a site and own, or hold the exclusive right to use, the land ^[34], and service does not start until the utility has the capacity to meet the contracted load ^[35]. The Office of the Ohio Consumers' Counsel, a party to the settlement in the tariff's original case, describes the result as an obligation on new large data centers to pay for at least 85% of contracted capacity for up to 12 years even when they use less ^[36].

The second is a bill. In its enrolled text, as passed by the Texas Legislature in 2025, S.B. 6 directs the state's utility commission to set interconnection standards for large loads, with a demand threshold of 75 megawatts at a single site unless the commission finds a lower one necessary ^[37]. The standards must require a large load customer to disclose whether it is pursuing a substantially similar request for service elsewhere in the state ^[38], to pay a flat charge of at least $100,000 for the initial transmission screening work ^[39], to show site control through ownership, a lease or another accepted legal interest ^[40], and to meet uniform financial commitment requirements for the transmission built to serve it ^[41]. The bill's disclosure requirement speaks to a problem that the site's archived post on the scale of Texas data center requests describes: a queue in which a large share of requests will never become load.

Neither document would reach a small facility, and the 1,000 kW building in the illustration would fall under neither. What they show is the anatomy of a power commitment. A study of a project of any size should collect the same facts from whatever agreement the serving utility uses: the capacity committed and the date it will be energized; any ramp schedule; the minimum bill and how long it runs; the exit terms; the security posted; and the utility's own written statement that the capacity exists. Bank examiners look for the same evidence. The OCC's examination questions ask whether a bank's construction loan files include a list of documents ^[42], and evidence of the availability of utilities to the site is on that list ^[43]. The OCC handbook also names the availability of public utilities among the matters a site analysis covers ^[44].

A minimum bill changes the shape of the risk. A utility obligation that runs for twelve years and charges for most of the contracted capacity whether or not it is used behaves as a fixed cost, and it may stay fixed for longer than the customer contracts that are meant to pay it. A study should set the utility term and the minimum charge beside the schedule of customer contract expiries, and should test coverage in a year when contracted customer load has fallen but the minimum bill has not.

Tenant commitments: the demand evidence the study rests on

For most classes in this series, demand is estimated from a trade area: residents, households, visitors or traffic. In the method used here the trade area is weak evidence for a data center. What a customer buys is power, network access and a price, none of which a population count measures, and a single customer can take a large share of a small building. The evidence that carries weight is contractual: which customers have signed, for how much load, at what price, for how long, and with what right to leave early.

Bank supervisors put the same point in writing. Among the standards the interagency guidelines expect a lending policy to set is "Pre-leasing and pre-sale requirements for income-producing property" ^[45], listed with requirements for feasibility studies and sensitivity analyses on development projects and completed commercial properties ^[46]. The OCC handbook explains the purpose: prudent loan policies set minimum levels of preleasing or sales as a condition of commitment or funding ^[47], and banks typically analyze preleasing to determine whether it represents bona fide commitments ^[48]. Among the underwriting weaknesses the handbook says may signal credit deterioration is a loan for a commercial development without significant preleasing and without adequate mitigants ^[49].

A study should sort the contractual evidence into tiers and weight each tier in the open.

• Signed contracts. Each is recorded with the customer, the contracted critical IT load, the monthly price per kilowatt and what it includes, the start date and any ramp, the term and renewal options, the termination and downsizing rights, and any deposit, guarantee or letter of credit behind it. Only these count in full.
• Letters of intent. Load a customer has said it intends to take, without a binding contract. They count at a stated weight, printed among the inputs, with the reason for that weight. The OCC's list of loan file documents places letters of intent to lease next to assignments of signed leases ^[50], which is a reason to collect them, not a reason to count them as leases.
• Proposals and conversations. Named prospects with nothing in writing. The study lists and describes them and gives them no weight.

Credit quality runs through every tier. The OCC handbook says a lower coverage ratio may in some cases suit properties with stable and certain cash flows, such as long-term net leases to highly creditworthy tenants ^[51], and that properties with volatile cash flows may warrant a higher one ^[52]. A data center study should not claim the first treatment by default. It should show which customers are rated or guaranteed, how long each commitment runs against the term of the loan, and what share of contracted load the largest customer holds. Short commitments have their own exposure: the handbook warns that properties with shorter lease terms are vulnerable to falling values as rents decline and leases renew at lower rates ^[53], and that in a weakening economy tenants could cut the space they use or stop paying altogether ^[54].

USDA's guarantee rules put a value on a contract. Under 7 CFR 5001, a start-up holding sales contracts that cover its debt service, that run at least as long as the guaranteed loan and whose counterparty's credit the agency accepts ^[55] can meet the capital test at the lowest level the rule sets for new businesses, for instance 10 percent balance sheet equity, or debt no more than nine times that equity, at closing ^[56]. The rule's summary table sets higher bars for the rest: 25 percent for a construction project whose guarantee the lender will seek before completion ^[57], and 20 percent balance sheet equity for every other new business ^[58]. It reaches only a project that is otherwise eligible for the program, a question the eligibility section takes up, but it is a published rule that treats signed demand from a sound counterparty as lowering the risk of a start-up.

Two physical facts can make a contract worth less than its face. The first is adaptability. Writing of manufacturing facilities, the OCC handbook says they are often built to a specific user's needs and that a building's adaptability to other users is an important underwriting consideration ^[59]. A hall designed around one customer's power density and cooling arrangement raises the same question, and a study should say what a different customer would need changed before it could move in. The second is technology. The handbook lists changes in technology among the factors that can leave an existing appraisal or evaluation out of date ^[60]. A study of a data center should date its description of the plant and say what the design would need in order to serve rack densities above the ones it was built for.

The demand model, equation by equation

The model below replaces population with power and contracts. In plain notation:

1. IT load the utility capacity supports: P = U ÷ PUE, where U is the capacity in the signed electric service agreement, at the meter.
2. Sellable critical IT load: S = the smaller of P and D, where D is the IT load the power and cooling plant is built for.
3. Committed load: C = signed load + w × load under letters of intent, where w is the study's stated weight.
4. Contracted load at stabilisation: L = S × s, where s is the stabilised share of sellable load under contract.
5. Load still to be contracted after opening: G = L less C. Months to stabilise: T = G ÷ a, where a is the load signed per month in dated lease-ups of comparable facilities.
6. Facility energy at stabilisation: E = L × average draw × PUE × hours in a year.
7. Contracted load the loan needs: B = (binding coverage minimum × debt service + fixed costs) ÷ margin per kilowatt-year, where the margin is the annual price of a contracted kilowatt less the cost of the energy that kilowatt consumes at the stated draw, PUE and price.

Every input is a labelled assumption except the electricity price, the prime rate and the program floor, which are sourced. Table 1 lists them, and Table 2 carries them through the equations in a base case and under three shocks.

Rounded as shown; lines may not foot by one unit in the last place.

In the base case the utility capacity supports 690 kW at the racks and the plant 650 kW, so the plant binds and 650 kW can be sold. Signed contracts cover 50.8% of it. With letters of intent at half weight, committed load is 390.0 kW against 552.5 kW at stabilisation, which leaves 162.5 kW to sign after opening, 13.5 months at the assumed pace. The coverage test in the mini-example needs 453.6 kW under contract, 69.8% of what can be sold.

The shocks show why power, not floor area, is the first variable a lender should move. If the design PUE proves two tenths higher, the utility capacity supports only 606 kW, power becomes the binding limit, and part of the plant can never be sold. Less load is left to absorb, so stabilisation arrives sooner, at 10.4 months, but the contracted load the loan needs rises to 80.4% of sellable load, because each contracted kilowatt now costs more energy to serve. A fifth less utility capacity than planned, which is what a downsized agreement or a delayed substation upgrade would mean, leaves 552 kW to sell and pushes the needed share to 82.2%. Half the absorption pace leaves capacity and the coverage threshold where they were and doubles the time to stabilise, to 27.1 months.

The energy line, 4,211 MWh a year at stabilisation in the base case, is the quantity the utility's tariff will price, and it is the number a lender can check against the utility's own load letter for the site.

The measures a credit officer reads, and where each is evidenced

The first four measures are particular to the class and decide the others: capacity and PUE fix the ceiling, and the signed share says how much of it is demand rather than expectation. A study should print each with its source document and date, so that a reader can trace the sellable load in the cash flow back to the utility agreement and the engineer's design.

Coverage and debt yield are defined as for any income property. The OCC handbook computes coverage by dividing net operating income by annual debt service ^[66], and debt yield by dividing it by the loan amount ^[67]. Debt yield earns its place in this class because it does not move with the interest rate or the amortisation, which matters where a lender shortens the amortisation to match the life of the electrical and cooling plant.

The value basis needs care in a building that is sold before it is finished. For supervisory loan to value, the OCC handbook says the as-completed value should generally be used for an owner-occupied building or a preleased property to be built, and an as-stabilised value for a property to be built that is not preleased to stabilised levels ^[70]. The supervisory limit for construction of commercial property is 80 percent, and for improved property 85 percent ^{[68] [69]}. A feasibility study does not supply the value; that is the appraiser's work, as the article on studies and appraisals explains. What the study supplies is the evidence of preleasing and stabilisation that decides which value basis applies, and the date on which each contract was verified.

The cost side has public anchors of its own, covered in the article on operating expense evidence and the article on construction cost evidence. For this class the anchor that matters most is the energy price, and the price a study uses should be the one the serving utility will actually bill; the EIA averages ^[19] are only a starting point.

Which data center borrowers the SBA and USDA rules reach

Eligibility comes before the model for any program loan, and in this class it turns on what the borrower does with the building. SBA may require a feasibility study, as it may require appraisals or a survey ^[71], so the study can itself become part of the eligibility file.

The first test is the passive business rule. The rule at 13 CFR 120.110(c) names, among ineligible businesses, "Passive businesses owned by developers and landlords that do not actively use or occupy the assets" a loan would buy or improve, while sparing Eligible Passive Companies ^[72]. A colocation operator that runs the power, cooling, security and network services its customers pay for is classified as a service establishment in 518210 rather than as a lessor ^[5], but an industry code does not decide SBA eligibility. A study should describe what the borrower's staff and contracts actually provide, so that the lender can judge whether the borrower actively uses the assets it asks to finance. A landlord that leases a powered building to an unrelated operator stands differently: it is the kind of owner the passive rule describes, and NAICS places it among the nonresidential lessors ^[7].

The Eligible Passive Company is the route for an owner that holds the real estate in one entity and runs the operation in another. Under 13 CFR 120.111 the holding company may spend proceeds only on property that its Operating Company, or several of them, will rent and use in their business, or else on transferring ownership between its current owners ^[73]. The Operating Company has to be an eligible small business whose proposed use would qualify if it borrowed in its own name ^[74]; each company has to be small under part 121 ^[75]; and the lease between them has to be written down and placed below SBA's lien on the property ^[76]. The conditions bind the two companies together. The Operating Company must guarantee the loan or join it as co-borrower ^[77], and the guarantor option falls away when the 7(a) proceeds also cover the Operating Company's working capital or other assets, since it must then borrow jointly ^[78]; the rent, for its part, may not exceed the loan payment plus the holding company's direct costs of keeping the property ^[79]. The structure therefore suits an owner and its own operating business, not a developer seeking a market rent from an unrelated operator.

Occupancy rules apply where part of a financed building is leased as real property to other businesses. The test works in shares of rentable property. In a building that already stands, the borrower keeps 51 percent or more for itself and may let no more than 49 percent to tenants ^[80]. A building newly built needs 60 percent in the borrower's own use, with no more than 20 percent let permanently, among other conditions the rule sets ^[81]. A study should map the floor area between the operator's own halls, plant rooms and offices and any space let to others on a lease, and name the rule it applied. Whether a customer's colocation contract is a lease of rentable property or a contract for services is a question for the lender and its counsel, and the study should describe each contract precisely enough for them to answer it.

Size is the second test. The size standard for 518210 is $40.0 million in average annual receipts ^[82], against $34.0 million for lessors of nonresidential buildings ^[83]. The applicant is measured together with its affiliates, against the higher of its own industry's standard and that of the group ^[84], so an operating company owned by a large hosting or cloud group fails the test if the group's receipts exceed the standard, however modest the single site. A 7(a) or 504 applicant that fails the industry standard can try a balance sheet and earnings test instead, again counting its affiliates: tangible net worth capped at $20 million, and net income after federal tax, averaged across its last two full fiscal years, capped at $6.5 million ^[85].

Loan limits and terms follow. The 7(a) ceiling on any one loan is $5,000,000 ^[61]. The rule on maturities starts from "The shortest appropriate term, depending upon the Borrower's ability to repay" ^[86]. Beyond that, ten years is the ceiling for a loan whose assets last no longer than that; real estate, and equipment expected to serve beyond ten years, can carry a longer term ^[64], to a limit of 25 years with extensions ^[87]. A data center loan mixes assets with very different lives: the shell, the electrical and cooling plant and, where the borrower rather than its customers owns them, servers and network equipment. A study should list the financed assets by class with an estimated useful life for each and show the blended term the cash flow supports, since the term is where the asset mix reaches the debt service.

The 504 program adds a job test that a power-dense building may struggle to meet. SBA's 504 rule ties the funding to employment, matching each Job Opportunity a project creates or keeps to a funding amount that SBA publishes by notice ^[88]. The notice of September 30, 2025 fixes that amount at $95,000 of SBA guarantee for each job, rising to $150,000 where the borrower is a small manufacturer or the project qualifies under an energy public policy goal ^[89]. The Ohio Consumers' Counsel observes that data centers create relatively few permanent jobs once they are built ^[90]. Falling short is not always fatal: a project that serves one of the listed community development or public policy goals can still be eligible, so long as the CDC's whole 504 portfolio, this loan counted, reaches the job average the CDC is required to keep ^[91]. An energy goal may fit: the statute sets a 504 limit of $5,500,000 for a project that reduces the borrower's energy consumption by at least 10 percent ^[92], and a study that claims the reduction should show the baseline and the result in the same metered units the PUE discussion uses. The 504 program article sets out the job test in full.

For USDA's Business and Industry guarantee the rural test matters: the rule's definition of a rural area leaves out cities and towns above 50,000 people and the urbanized fringe joined to them ^[93], except where the Under Secretary of Rural Development rules that a particular piece of that fringe is rural in character, which keeps it inside the program's reach ^[94], a ruling made one project at a time ^[95]. One borrower's guaranteed total stops at $25 million ^[96]. The Secretary may go higher, to $40 million, only for rural cooperatives that process value-added agricultural commodities, which a data center operator is not ^[97]. Where a data center project is otherwise eligible, the equity rule for new businesses with qualifying contracts, described in the section on tenant commitments, is the program's own recognition that signed demand reduces risk; the article on 7 CFR 5001 covers the rest of the program.

For assisted living the gate is a state care licence. For this class, in the public documents this article draws on, the gate is power: a filed tariff in Ohio ^[24], a bill the Texas Legislature passed in 2025 ^[37], and, wherever a facility is built, the serving utility's own rules for connecting a large load. For a data center, the utility documents do the work that a licensing file does for other classes, and a study should present them as eligibility evidence, not as background.

The class in SBA's loan files: few approvals, mostly not buildings

Over fiscal years 2017 to 2026 the data center class holds 911 ^[98] 7(a) approvals for $476,913,400 ^[98] and 50 ^[99] 504 approvals for $41,581,000 ^[99]. Every one of those figures carries the code's limitation: 518210 also holds hosting, cloud and data processing businesses, and the file cannot separate a colocation building from a hosting service. The code on an approval describes the borrower's business, not what the loan bought, as the article on the file's limits explains, so nothing in the cells marks which approvals financed a building.

The loan sizes sit far below the program maximum. The median 7(a) approval in the class is $200,000 ^[98], with a quarter of approvals at or below $84,400 ^[98] and a quarter above $440,500 ^[98], all a long way inside the $5,000,000 cap ^[61]. The 504 median is $591,000 ^[99]. The loan assumed for the 1,000 kW facility in the illustration, a building small beside the loads the Ohio tariff and the Texas bill address, sits above the upper quartile of the class's 7(a) record. The record should be read as the SBA history of businesses in the code, of which data center real estate is at most one part, and not as evidence about data center buildings.

Franchising is almost absent from the class: 4 ^[98] of the 7(a) approvals, a share of 0.44 percent ^[98], carry a franchise code. The 7(a) approvals are spread across 52 ^[98] of the file's state codes, with California first at 103 ^[98] approvals and Texas second at 97 ^[98].

The yearly series moves without any explanation the file can supply. The class's 7(a) approvals numbered 98 ^[98] in FY2017, fell to 55 ^[98] in FY2020, rose to 151 ^[98] in FY2024 and 145 ^[98] in FY2025, and stood at 87 ^[98] in FY2026. The 504 series peaks at 11 ^[99] approvals, in FY2019. No cause is offered here, because the dataset's report gives none.

The method in this article is written so that it serves a conventional lender as well as an SBA one. The SBA record holds few approvals in the class and cannot show which of them financed a building; the bank rules set out in the conventional loan article and the utility terms described above are the standards a data center loan of any size will be read against.

What the class's floors show against the programs

The loan file supports two outcome measures for the class, and both are floors. A cancellation rate counts approvals cancelled before any disbursement; a charge-off rate counts disbursed loans charged off to date. The series piece on the two outcomes explains the difference, and the cohort floor article shows why neither can do anything but understate.

In 7(a), 11.19 percent ^[98] of the class's approvals in the FY2017 to FY2024 cohort ended in cancellation, 76 ^[98] of 679 ^[98], a floor that sits beside a program-wide 7(a) floor of 11.24 percent ^[100] over the same years. If the 10 ^[98] class approvals still undisbursed were all cancelled, the class's rate would reach its ceiling of 12.67 percent ^[98].

For the FY2017 to FY2021 cohort the class's 7(a) charge-off floor is 6.95 percent ^[98], 23 ^[98] of 331 ^[98] disbursed loans, a figure the program as a whole puts at 6.33 percent ^[101]. Of the loans in the class's denominator, 85 ^[98] remain open; SBA keeps a disbursed loan's status out of the public release until it ends in payoff, charge-off or cancellation ^[102], which is one reason the rate can only rise. On the fixed five-year horizon, which lets one cohort be set against another, the class shows 3.32 percent ^[98], the program 3.72 percent ^[103]. Counted by dollar amount instead of by loan, the class reaches 3.02 percent ^[98] and the program 1.95 percent ^[104].

In 504 the class's cancellation floor is 10.87 percent ^[99], 5 ^[99] of 46 ^[99], below the program's 13.60 percent ^[105] for the cohort. No 504 charge-off rate is printed for the class, since too few loans, under 30, sit in its FY2017 to FY2021 cohort; the thirty-loan rule explains why no rate is printed on so few. For the program as a whole, the 504 charge-off floor for that cohort is 0.35 percent ^[106].

What the floors can say is narrow. The class's 7(a) charge-off floor to date sits above the program's, its five-year measure below it and its dollar floor above it; its cancellation floors sit near or below the programs'. Read together, they describe the SBA experience of the businesses in code 518210 over those cohorts, of which buildings may be only a part. They cannot say why loans in the class failed, and nothing in the file allows a cause to be assigned. Nor can they be carried over to a colocation building financed on signed contracts, whose risks are the ones the earlier sections set out. A study that prints them should print them as context, with their keys and the code's limitation, and leave the underwriting to the power and contract evidence. The block below lists every figure this section uses.

Loan Data Behind This Article

Every figure in this table comes from the Loan Analytics loan dataset, analytics-loan-data freeze-3 (commit 360c856), and carries its key. SBA rates are approval-cohort floors and are withheld when fewer than 30 loans stand behind them.

A worked mini-example: a small colocation facility on its signed contracts

The series has no worked example for the class yet, and this section is not a full one. It carries the inputs of Table 1 through a coverage test for a small colocation facility with 650 kW of sellable critical IT load, financed with a $3,000,000 loan inside the 7(a) maximum ^[61]. Customers pay an all-in monthly price per contracted kilowatt and the operator pays for the energy. The electricity price, the prime rate and the program floor are sourced; every other figure is an assumption, labelled as such, and no real facility stands behind any of them.

Conventions. This article follows the conventions of the Loan Analytics worked examples. Coverage is tested against the program floor the site's program page states, where the program has one, and against a stated 1.25x lender assumption, and the higher of the two binds. Equity is measured at the loan the cash flow supports, with fees recomputed on that loan. An acquisition is tested on the seller's historical earnings, not on projections.

The site's SBA program page puts the program floor this way: "Under SOP 50 10 8.1, standard 7(a) loans need at least 1.15x business coverage, and startups must reach it within two years of funding." ^[65] Against that floor and against the series' lender assumption of 1.25x, the illustration tests the cash flow, and whichever is higher governs.

On signed contracts alone, 330 kW, the facility earns $205,080 of net operating income against $381,371 of debt service, coverage of 0.54x. That income supports a loan of $1,290,588 at the binding minimum, well short of the $3,000,000 requested. At stabilisation, with 552.5 kW under contract, coverage is 1.82x, the debt yield 23.1%, and the income would support $4,367,153. The loan works only on contracts not yet signed, and the distance between those two columns is the absorption risk the lender is being asked to carry.

The shocks are the class's own. A design PUE two tenths higher leaves stabilised coverage at 1.40x, because the building has less load to sell and spends more energy serving each contracted kilowatt. Electricity 4 cents a kilowatt-hour dearer leaves 1.38x. A rise in the average draw from 60% of contracted load to three quarters of it leaves 1.42x: under an all-in price the operator pays for the energy its customers choose to use, which is a reason for a study to read each contract's power terms instead of assuming a draw. Losing the largest customer before stabilisation turns net operating income negative. That is the concentration test, and a lender should see it printed for any building in which one customer holds a large share of signed load.

Time matters as much as level. In the base case Table 2 puts stabilisation 13.5 months after opening; at half the absorption pace it is 27.1 months, past the two-year limit the site's program page sets for a startup ^[65]. A study should carry coverage quarter by quarter through that period, with the reserve that funds each shortfall, as the article on ramp-up sets out, and should measure the equity at the loan the cash flow supports, as the article on sources and uses does. A lender reading this illustration would see a loan that coverage supports only once the building is close to stabilised, and the question the study has to answer is how close the signed contracts and the utility commitment bring it on the day the loan closes.

Questions to put to a data center study before credit committee

The list below restates the method as questions. Each should be answered on the study's own pages, with the document behind the answer named and dated.

• Power. What capacity has the utility committed in an executed agreement, from what date, with what ramp, minimum bill, term, exit terms and security? Does the file hold the utility's evidence that the capacity exists ^[43]?
• The binding limit. Is the sellable load set by the plant's design or by the power, and what happens to it if PUE comes in above design?
• Contracts. How much load is signed, at what price, for what term and with what termination rights? How are letters of intent weighted, and are they shown apart from signed load?
• Bona fide commitments. Do deposits, guarantees or letters of credit stand behind the contracts, the test the OCC handbook describes banks applying to preleasing ^[48]?
• Terms in line. Does customer income outlast the debt service and the utility's minimum bill, and what does coverage look like in the year the largest contract expires?
• Concentration. What share of signed load does the largest customer hold, and is coverage printed without it?
• Energy cost. Is the price taken from the serving utility's tariff for the right customer class, with demand charges and riders, and who bears the energy cost under each contract?
• Eligibility. Does the borrower actively use the assets, is it small together with its affiliates, and does the asset mix support the term requested? For 504, how is the job opportunity test met?
• The loan record. Is any SBA figure for the class presented as evidence about data center buildings? It should be printed only with the code's limitation beside it.

With those answers on file, the reviewer's next step is to test the two numbers the whole study stands on against their own sources: the capacity and energization date, in writing from the serving utility, and each signed contract's kilowatts, price and term, matched line by line to the schedule in the study. If either number moves, Table 1 is where the change goes, and Table 5 then shows what the loan can carry.

Related Pages on Loan Analytics

• The Loan Analytics methodology series (all articles)
• How a feasibility study is done: the Loan Analytics methodology
• Bankable feasibility studies for conventional loans
• SBA feasibility studies (7(a) and 504)
• Data center: loan data and feasibility studies
• Loan data by state
• Loan data by asset class
• Worked examples (case studies)
• Data Center Site Readiness Scorecard: Ranking U.S. Metros on Power, Land & Pipeline

Commission a Feasibility Study

Loan Analytics prepares lender-grade feasibility studies for SBA 7(a), SBA 504, USDA Business and Industry, conventional and EB-5 financing, built on the methods this series sets out. A study is not an appraisal: it tests whether a project's market, costs and cash flow can carry the debt a lender is asked to approve. To discuss a study for your project, contact Loan Analytics.

Daniel Smith, MAI, Loan Analytics