The Fifty-Four-Million-Dollar Job: USDA Financing, Rural Data Centers, and the Feasibility Question

The AI build-out is moving to the countryside. The federal program most associated with rural development was built for something it cannot finance. Here is what that means for lenders, developers, and the studies that decide which deals are real.

Drive an hour past the edge of almost any American metro and you will start to see them. Concrete boxes the size of warehouses, set back behind new substations, ringed by transformers and chain-link, going up on land that grew soybeans two seasons ago. The data center has become the most aggressive new entrant in rural and exurban real estate, and the numbers behind it are genuinely large. Roughly two-thirds of planned U.S. data center capacity is now headed to rural and frontier markets (10), pulled there by cheap land, available power, and tax incentives that small counties are eager to write.

For anyone who works in rural development finance, that migration raises an obvious question. The United States Department of Agriculture runs the country's marquee rural business lending programs. Its Business and Industry guarantee, its Rural Energy for America Program, its electric and broadband portfolios exist precisely to put capital into the kinds of places where data centers are now landing. So the pitch writes itself: federal money, rural mission, digital infrastructure, a match made in Washington.

It is also, on close inspection, mostly wrong. Not because the programs are badly run, and not because data centers are unwelcome in rural America. It is wrong because of a single number, and because of what that number does to the logic of every program USDA administers.

A program built for payroll

Start with what the Business and Industry program is actually for. B&I guarantees loans that commercial banks make to rural businesses, and its founding purpose, written into statute and repeated across every piece of agency guidance, is to create and preserve jobs in communities of fifty thousand people or fewer (5). The program does not buy buildings or write checks directly. It stands behind a bank, absorbing 70 to 85 percent of the loss if a rural borrower defaults, so that capital flows to places lenders would otherwise treat as too thin or too risky.

The Rural Energy for America Program follows the same instinct in a different sector, financing renewable energy systems and efficiency upgrades for agricultural producers and small rural businesses (6). The electric program lends to the cooperatives that keep the lights on across most of rural America. The broadband programs, ReConnect chief among them, have pushed more than five billion dollars of fiber toward unserved communities (described in the rural development research). Every one of these instruments was designed around the same theory of the case: rural prosperity comes from rural employment and rural connectivity, and federal credit should tip projects that deliver them.

A data center delivers almost none of the first thing. That is not a slur on the asset class. It is what the asset class is. A modern hyperscale campus is among the least labor-intensive large structures in the economy, and the gap between its capital and its payroll is not a rounding error. It is the entire story.

The math that breaks the pitch

Here is the number. An analysis by Food and Water Watch found that across data centers built in Virginia over the prior five years, each permanent job represented roughly fifty-four million dollars of investment, about a hundred and sixty-eight times the capital behind the average job in the state (1). Measured across all Virginia facilities since 1990, the figure was still around thirteen million dollars per job, roughly a hundred times the cost of a comparable position in other industries (1). Virginia's own legislative review commission put concrete staffing behind the ratio: a typical large data center runs on about fifty full-time workers, and roughly half of those are contractors (2).

The independent research lands in the same place, with more caution. A Brookings study of roughly seven hundred and seventy facilities found that counties receiving their first large data center did see private employment rise four to five percent over five to six years, but that the naive comparisons used in many industry-sponsored reports overstate the jobs effect by a factor of three (3). The same work drew a distinction that matters enormously for anyone underwriting one of these deals: the employment benefit concentrates in dense hyperscale clusters, not in single facilities or colocation sites, and incentives represent about two percent of hyperscale construction cost but more than sixty percent in colocation counties (3). In plain terms, the projects that generate the smallest job benefit are exactly the ones that lean hardest on public subsidy.

Now set that against the way USDA writes credit. B&I's standard loan ceiling is twenty-five million dollars (5). The Secretary can approve more in narrow circumstances, but twenty-five million is the working number. A single modern AI campus costs between several hundred million and several billion dollars to build, at development costs that now run nine to fifteen million dollars per megawatt of critical load (11). The guarantee, in other words, would cover a fraction of one building's electrical room.

People sometimes assume the job-creation rule is the binding obstacle, that a low-headcount data center would be disqualified for failing some federal cost-per-job test. It would not, and the reason is worth understanding precisely. Under the consolidated rule that governs B&I today, job creation is a scoring factor, not a gate. The priority point system awards five points out of a hundred where a project creates or saves at least five jobs above 150 percent of the federal minimum wage (5). Eligibility itself turns on rural location, financial feasibility, collateral, equity, and ordinary credit underwriting, not on headcount. A capital-heavy, low-labor project is not categorically barred.

So the obstacle is not the jobs rule. The obstacle is arithmetic and incentive. Twenty-five million dollars cannot finance a building that costs a thousand million, and a company with the balance sheet of a major cloud provider has no reason to ask the federal government to guarantee a loan it can fund three other ways. Search the public record for a data center financed through B&I, REAP, or the rural electric program and you will not find one. None appears in USDA's own materials, none in the cooperative trade press, none in the inspector general or accountability reports through the first half of 2026 (research synthesis). The absence is itself the finding. Given the program design, an actual USDA-financed hyperscale data center would be an anomaly, and anomalies that large tend to leave a paper trail.

It is worth saying clearly, because loan brokers have begun marketing the opposite. A handful of lenders now advertise USDA "data center financing," and the claims describe eligibility in principle rather than any closed transaction. Read them as sales copy, not evidence. The B&I statute permits financing commercial and industrial property in rural areas; that is true and unremarkable. It has not, on the record, financed a data center.

Where the money actually comes from

If not USDA, then what? The named projects answer the question without ambiguity.

Meta's Hyperion campus in Richland Parish, Louisiana sits on roughly 2,250 acres of former farmland and began as a ten-billion-dollar investment, scaling toward twenty-seven billion through an October 2025 joint venture with Blue Owl Capital (19). Meta says the site will employ more than five thousand skilled-trade workers during peak construction and support more than five hundred operational roles (19). A regional impact analysis projected that direct local employment would plateau closer to three hundred and twenty-six jobs (19). Louisiana's incentive package, by one published analysis, runs to roughly 3.3 billion dollars, anchored by a twenty-year exemption on equipment that includes the chips themselves (19). The capital came from Meta's balance sheet and private credit. None came from Washington's rural programs.

The first Stargate site in Abilene, Texas tells the same story in sharper relief. The developer committed to spending up to several billion dollars and secured an eighty-five percent property-tax abatement, and in exchange contractually promised three hundred and fifty-seven permanent jobs at minimum salaries near fifty-eight thousand dollars (20). The financing came through a multibillion-dollar joint venture with Blue Owl and roughly 2.3 billion dollars of construction debt from JP Morgan (20). Fifteen hundred people were on site building it; three hundred and fifty-seven will stay (20).

Amazon's investment in Madison County, Mississippi was announced as the largest in state history, ten billion dollars for about a thousand jobs, supported by a forty-four-million-dollar state package, and has since grown toward twenty-five billion dollars across three counties (21). In central Ohio, the village of New Albany built an entire economic model around the asset's peculiar shape. Because the community relies on income rather than property taxes, it structured payment arrangements that require minimum annual payments in lieu of taxes, so that a facility employing two hundred people delivers the fiscal weight of an office tower employing two thousand (22). One representative small project there, a twenty-five-million-dollar data center, received a sixty-five percent property-tax abatement in exchange for eight jobs (22).

The pattern is consistent across every case. Data centers are financed by hyperscaler equity, by private credit funds, and by state and local tax incentives negotiated facility by facility. The public money that flows is overwhelmingly state and local, and it flows as abatement, not as a federal loan guarantee. Good Jobs First, the most rigorous skeptic of these subsidies, found the average data center megadeal costs about 1.95 million dollars per job and that at least ten states already forgo more than a hundred million dollars a year in revenue to the sector, with Texas alone giving up more than a billion (4). The organization's recommendation, a cap of fifty thousand dollars in combined subsidy per permanent job, would foreclose nearly every deal on the board (4).

The narrow door

So is there any honest intersection between USDA financing and a rural data center? Yes, but it is narrow, and a feasibility advisor owes a client precision about exactly where the door sits.

Three pathways survive contact with reality. The first is the owner-operated facility at genuine edge or enterprise scale. A regional hospital system, a manufacturer, a bank, or an agricultural cooperative that builds its own compute, rather than leasing to a cloud provider, in a community under fifty thousand people, can plausibly seek a B&I guarantee on the building and equipment, capped near twenty-five million dollars. This is the most credible route, and it works only because the borrower is a real rural enterprise rather than a hyperscaler's special-purpose vehicle.

The second is energy rather than compute. REAP can finance an on-site solar array or an efficiency retrofit at a facility owned by an agricultural producer or a qualifying small rural business, up to a twenty-five-million-dollar loan guarantee (6). It cannot finance the data center itself, and the recipient has to clear the program's size standards, which a hyperscaler never will. One important caveat sits on this pathway as of early 2026: USDA has paused new REAP grant awards pending a rule rewrite, leaving the guaranteed loan as the live instrument (6).

The third is the grid, through the cooperative that serves the load. The rural electric program lends to cooperatives for generation, transmission, and distribution, and a co-op serving a data center could in theory finance supporting infrastructure that way. In practice, cooperatives are managing this surge through large-load tariffs that require the customer to pay for new generation, precisely to keep those costs off the ratepayer-supported, federally backed balance sheet. The cooperative trade association has told federal regulators that some of its generation-and-transmission members, which historically serve peaks of five hundred to two thousand megawatts, are now fielding single interconnection requests of five hundred to a thousand megawatts each (rural development research). No public record ties a specific federal electric loan to a named data center customer. The pressure is real; the documented deal is not.

These are the genuine edges. They share a ceiling, a borrower profile, and a scale that confine them to small enterprise and edge facilities. For anything resembling the projects in the headlines, USDA is not the answer, and a study that pretends otherwise is doing a client a disservice.

What a feasibility study has to prove instead

Here the analysis turns from financing to physics, because the more useful question is not which federal program a rural data center can tap. It is whether the project is real at all. And the discipline that answers that question has almost nothing to do with jobs and everything to do with four constraints: power, water, fiber, and land. A feasibility study earns its keep by proving those out before a dollar of debt is committed, and the order is not arbitrary. Power comes first, because power is the constraint that kills deals.

Power, or nothing

The demand shock to the grid is the single most important fact in the sector. The International Energy Agency projects global data center electricity demand more than doubling toward 945 terawatt-hours by 2030, with data centers accounting for close to half of all U.S. electricity demand growth this decade (13). Lawrence Berkeley National Laboratory found U.S. data centers already consumed about 4.4 percent of national electricity in 2023, on a path toward as much as 12 percent by 2028 (12). Grid Strategies tallied utility load forecasts that rose roughly sixfold in four years, to an additional 166 gigawatts of expected peak growth by 2030 (14).

The grid cannot deliver that on the schedule developers want, and the bottleneck has a name: interconnection. In the largest U.S. grid region, the wait from interconnection request to commercial operation has stretched past eight years, up from under two in 2008 (15). New generation is no faster. The three companies that build large gas turbines are sold out into 2029 and beyond, with backlogs measured in tens of gigawatts (16). Capacity prices in that same region rose roughly tenfold across three auctions, and the market monitor attributed about forty percent of one auction's cost directly to data centers (17). The nuclear deals that make headlines, such as the Microsoft-backed restart of a reactor at Three Mile Island targeted for 2028, are real but slow, and small modular reactors remain a 2030-and-later proposition (18).

For a feasibility study, this converts into a single underwriting test that separates a financeable project from a press release. There is a hierarchy of power certainty. At the bottom sits announced or pipeline capacity and a utility letter of intent, which is not bankable and which no serious lender will accept as proof. In the middle sits an executed interconnection agreement, a power purchase agreement, or a binding will-serve letter with a committed energization date, which is the stage at which a deal becomes real. At the top sits delivered, energized power feeding an operating building. A study that cannot document where a project sits on that ladder has not assessed the project. The market has a word for capacity claimed without dated power behind it. They call it bragawatts, and applying historical conversion rates to the announced North American pipeline suggests only a fraction of it ever energizes (research synthesis).

This is also why the rural setting cuts both ways. Cheap land near a substation looks like an advantage until the study examines whether the rural grid, often the least modernized infrastructure in the country, can actually carry the load, and whether the serving cooperative has the capacity and the tariff structure to absorb a request that may dwarf its existing peak.

Water is the second power

The constraint that has gone from afterthought to deal-killer fastest is water. Cooling a large data center can consume three to five million gallons a day, comparable to a town of thirty to fifty thousand people, and Lawrence Berkeley estimates the sector's direct water use at roughly seventeen billion gallons in 2023, with the indirect figure tied to electricity generation many times larger (12). The trouble is that the build-out is concentrating in exactly the wrong places. By one analysis of water-stress data, about two-thirds of new U.S. data centers since 2022 have gone into areas already under high water stress (water and permitting research).

The flashpoints are instructive for anyone siting a project. Tucson's city council voted seven to nothing in August 2025 to reject a large data center over water and energy concerns (water and permitting research). Chandler, Arizona, the first U.S. city to regulate data center water, has effectively closed itself to water-cooled development (water and permitting research). The community resistance is now quantifiable: industry trackers counted on the order of a hundred billion dollars in projects blocked or delayed in a single quarter of 2025, with more than a hundred local moratoriums and hundreds of state bills in play (downside research). A feasibility study has to treat water rights and cooling design as risk line items with the same weight as power: documented supply, secured rights or a reclaimed-water agreement, a cooling architecture matched to the basin, and a clear-eyed read of the permitting timeline, which can run two to three years where a federal nexus triggers environmental review. None of that is environmental site assessment work, which MMCG does not perform. It is risk identification, and it belongs in the study because it determines whether the project gets built.

The case the asset can actually make

If a rural data center cannot honestly sell itself on jobs, what can it sell? The fiscal case, and it is a strong one when it is told straight. In Loudoun County, Virginia, the densest data center market on earth, the facilities occupy roughly four percent of commercial parcels yet generate about thirty-eight percent of the county's general fund revenue and nearly half its property tax (8). The assessed value per data center parcel runs many times that of the next-highest commercial use, and the county has been able to cut its residential tax rate every year as a result (8). A regional study calculated that without the data centers, the typical Loudoun homeowner would pay roughly 5,800 dollars more in property tax annually (8).

That is the honest pitch for a rural data center: not employment, but tax base, capital investment, and infrastructure that a small county could never fund itself. A feasibility study models that case with the same input-output multipliers used across commercial real estate, but it presents the result honestly, as gross effects, and it pairs the jobs line with the tax line so a county knows which one it is actually being offered. The fundamentals support the demand side of that case. National colocation vacancy hit a record low of about 1.4 percent at the end of 2025 even as inventory grew thirty-six percent, with record absorption and asking rents above 195 dollars per kilowatt-month, and more than ninety percent of capacity under construction already pre-committed (9)(10). Demand, for the moment, is not the question. Deliverable power and water are.

Pricing the downside

A study that only models the upside is marketing. The discipline is in the stress test, and 2025 and 2026 have handed analysts a rich set of downside scenarios to run.

Begin with the question hanging over the entire sector: whether this is a bubble. The bear case is serious and quantified. Bain and Company estimated the industry needs roughly two trillion dollars in new annual revenue by 2030 to justify the build, and saw a shortfall approaching eight hundred billion even after counting AI-driven savings (25). A widely cited MIT study reported that the large majority of enterprise AI pilots had produced no measurable return, though that figure is methodologically contested. The International Monetary Fund warned in its October 2025 stability report that risk-asset valuations sat well above fundamentals and that a shift in expectations about AI could trigger a sharp repricing (24). The bull case is also real: the largest cloud providers still fund their capital spending out of enormous operating cash flow, demand and pre-leasing are at records, and physical power and fiber are long-lived even when first movers lose money. The point of a feasibility study is not to settle the debate. It is to ensure the project survives if the bears are right.

Three downside vectors deserve explicit modeling. The first is obsolescence, and it is structural. The chips inside these buildings turn over on a roughly two-to-three-year cycle, and an accounting argument has broken into the open over whether operators are understating depreciation by assuming five- and six-year useful lives (downside research). The feasibility-relevant distinction is between the equipment and the shell. The building can be physically sound and still be functionally obsolete if it was designed for one generation of hardware and the next generation demands far higher rack densities and liquid cooling. A study should model the risk that the facility needs a major mechanical and electrical retrofit, given that those systems can be sixty to seventy percent of development cost.

The second vector is the tenant. A purpose-built single-tenant data center is not a diversified building; it is a bet on one counterparty paying rent for the full term. The credit gap between an investment-grade cloud provider and a cash-burning AI startup is enormous, and the weaker counterparties are precisely the ones expanding fastest. A prudent study underwrites the lease, not the headline rent, and treats a non-investment-grade tenant the way a lender would, with a guarantor analysis and a renewal probability that the rating agencies put as low as one in four under stress (downside research).

The third vector is recovery, and it is where rural single-purpose assets are most exposed. The one large U.S. data center bankruptcy that offers a comparable, Cyxtera, ended in a going-concern sale in a strong market, and even then secured creditors took a loss and unsecured creditors recovered pennies (downside research). The agencies that rate this debt model the worst case bluntly: Moody's assigns terminal value based on land alone, assigning effectively nothing to the building, because the specialized infrastructure inside it does not transfer to another use (23). The same firm flagged that the major cloud tenants carried hundreds of billions in lease commitments not yet on their balance sheets (23).

Translate all of this into the pro forma and the downside case for a rural, single-tenant, speculative build writes itself: a pre-leasing haircut, a lease-up delay of twelve to twenty-four months, an anchor-tenant default with a long re-tenanting gap, rent compression, an obsolescence-driven retrofit, and a refinancing test that values the asset near its land-only floor rather than its replacement cost. If the deal only works at replacement-cost terminal value, or assumes a liquid exit to another operator from a remote location with a thin buyer pool, it has not passed the test. A monitoring framework watches the leading indicators: slipping pre-leasing, the lease cancellations that one analyst flagged at a major cloud provider in early 2025, falling chip rental rates, and funding stress at the AI labs themselves (downside research).

What it comes down to

The rural data center boom is real, and the capital behind it is staggering. But the instinct to route that capital through USDA mistakes the shape of both the asset and the programs. USDA was built to finance rural jobs and rural connectivity at the scale of a single community. A data center is a capital-and-power machine that creates very few permanent jobs and costs more than any USDA instrument can carry. The two intersect only at the margins, in small owner-operated and edge facilities, in on-site energy, and in the cooperatives that supply the power. For the projects that make headlines, the honest answer to a client asking about USDA options is that the programs are the wrong tool, and the right capital stack is hyperscaler equity, private credit, and state incentives.

That clarity is the point of doing the work. A feasibility study for one of these projects is not an exercise in finding a financing program. It is a disciplined test of whether the power can be delivered on a bankable timeline, whether the water exists and the permits will clear, whether the fiscal case is told honestly, and whether the deal survives a serious downside. Get those answers right and the financing follows. Get them wrong, or skip them in favor of a federal program that was never going to apply, and the project stalls in an interconnection queue or strands on a site no one else can use.

MMCG Invest builds those studies for lenders and developers across the asset classes where the answers are hard and the capital is large. If you are evaluating a rural data center, a power-constrained site, or any project where the feasibility question is the whole question, the next step is a conversation.

June 23, 2026, by Michal Mohelsky, J.D. Principal of MMCG Invest, LLC, feasibility study company serving feasibility studies for assisted living facilities.

Reach out to discuss how our methodology supports your lending decision.

Michal Mohelsky, J.D. | Principal | mmcginvest.com 

Contact: michal@mmcginvest.com

Phone: (628) 225-1125

Disclaimer: This report is provided for informational purposes only and does not constitute investment advice. Data presented herein is derived from proprietary MMCG databases and third-party sources believed to be reliable; however, MMCG Invest makes no representation as to the accuracy or completeness of such information. Figures from third-party industry databases have been independently verified and, where appropriate, adjusted to reflect MMCG's proprietary analytical methodology. Past performance is not indicative of future results.

Sources

1. Food & Water Watch, Artificial Jobs: The Illusion of Big Tech's Data Center Promises (January 2026).

2. Virginia Joint Legislative Audit and Review Commission (JLARC), Data Centers in Virginia (December 2024).

3. Brookings Institution, New Evidence on Data Center Employment Effects (Bahar and Wright, 2026).

4. Good Jobs First, Money Lost to the Cloud (2016) and Cloudy with a Loss of Spending Control (2025).

5. U.S. Department of Agriculture, Rural Development, Business and Industry Guaranteed Loan Program materials and 7 CFR Part 5001, including the §5001.318 project priority point system.

6. U.S. Department of Agriculture, Rural Development, Rural Energy for America Program (REAP) program materials and March 2026 program notice.

7. National Rural Lenders Association / Summit LLC, USDA B&I Guaranteed Loan Program: Economic Assessment 2025, written testimony to the U.S. House Agriculture Subcommittee (September 18, 2025).

8. Loudoun County, Virginia, FY2026 budget materials and Northern Virginia Technology Council data center fiscal analysis (2025–2026).

9. CBRE, North America Data Center Trends, H2 2025.

10. JLL, North America Data Center Report, Year-End 2025.

11. Cushman & Wakefield, Data Center Development Cost Guide (2025) and Global Data Center Market Comparison(2026).

12. Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (Shehabi et al.) and Queued Up: 2025 Edition.

13. International Energy Agency, Energy and AI (April 2025).

14. Grid Strategies, Strategic Industries Surging: National Electricity Demand Forecasts (2025).

15. RMI / PJM Interconnection, interconnection queue timeline analysis (2025).

16. GE Vernova and Siemens Energy public order-backlog disclosures (2025–2026).

17. PJM Interconnection Base Residual Auction results and Monitoring Analytics market monitor reports (2024–2026).

18. Utility Dive, reporting on the Constellation Three Mile Island / Crane Clean Energy Center restart and Microsoft power purchase agreement (2024–2025).

19. Meta press materials, Louisiana Economic Development, and Sherwood News analysis of the Richland Parish "Hyperion" campus (2024–2025).

20. Texas Observer and Data Center Dynamics reporting on the Stargate / Crusoe / Oracle Abilene campus (2025).

21. WJTV and Mississippi Development Authority reporting on Amazon Web Services' Madison County investment (2024–2026).

22. News 5 Cleveland and Data Center Dynamics reporting on New Albany, Ohio and central Ohio data center tax structures (2024–2026).

23. Moody's Ratings, data center asset-backed securities methodology (February 2025) and hyperscaler lease-commitment analysis (2025–2026).

24. International Monetary Fund, Global Financial Stability Report: Shifting Ground beneath the Calm (October 2025).

25. Bain & Company, Global Technology Report 2025 (September 2025), with reference to Sequoia Capital's "$600B question" and TD Cowen channel-check reporting on hyperscaler lease activity.