Fifty-Seven Jobs: The Fiscal Bargain at the End of the Data-Center Series

Subtitle: State auditors are cutting their own job counts by seventy percent, four of twenty audited Texas facilities missed the headcount they were paid for, and the largest transfer isn’t the tax break at all — it’s the utility bill

The same thousand acres, promised twice

In 2017, Mount Pleasant, Wisconsin was promised 13,000 jobs and $10 billion. By 2019 the deal had been renegotiated down to a target of 1,454 jobs and about $672 million. The last publicly verified count, through 2022, was 1,029 jobs and $571.2 million — under the reduced target, and under eight percent of the original number. A manufacturing plant, not a data center, and still the most thoroughly documented arc a jobs promise has ever been made to travel in American economic development: promised, renegotiated, and then missed again.

Microsoft’s Fairwater campus now sits on part of that same land — 1,363 acres, acquired for $225.8 million.

Readers of the last edition will recognize the address. Mount Pleasant is where roughly 1,000 residents filed a noise-and-light class action against Microsoft in July 2026. The same village has now hosted two consecutive generations of the same transaction: public money and public patience, exchanged for a headcount. The first time, the headcount was renegotiated once and missed anyway. This time, the promise is smaller and the check is larger.

This is the fourth and final edition of the data-center series. The first followed the power — megawatts living in paperwork the county recorder has never seen. The second followed the water — permits that are licenses, not property. The third followed the neighbors, and what a decibel table does to a buildable envelope. This one follows the money in both directions: what a community actually pays to land one of these campuses, what it actually receives, and what happens when the two don’t match. For the attorneys, CPAs, and lenders this series is written for, the uncomfortable finding is not that the deals are bad. It is that almost nobody — including the states writing the checks — can tell you whether they are, and the mechanism designed to answer that question, the clawback, has almost never been tested.

Key takeaways for attorneys, CPAs, and lenders

• The audits have turned. Georgia’s Department of Audits published data-center job figures in December 2025 and cut them by roughly 70% five weeks later. Virginia’s JLARC found the data-center exemption consumed 53% of a decade of state incentive spending. Texas has audited 20 of 138 certified facilities and found 6 out of compliance — four of them for failing to create the jobs they were paid to create.
• Clawbacks are drafted almost everywhere and enforced almost nowhere. The statutes typically delegate the recapture formula to a negotiated MOU rather than fixing it in law. Public reporting establishes no data-center clawback that has been triggered and recovered for a disclosed dollar amount. Where real teeth have appeared, they are utility collateral, not tax recapture — Wisconsin regulators required security reported at over $7 billion for a single campus.
• The tax break is not the biggest transfer. PJM’s own market monitor attributed $6.3 billion of $16.4 billion in the capacity auction reported July 2026 to data centers. Virginia’s answer was not to repeal its sales-tax exemption but to invent a new one-cent-per-kilowatt-hour tax on the load itself — including power the facility generates for itself.

Beyond the Core Four examines specialized properties outside the traditional industrial, retail, multifamily, and office categories — quarries, marinas, agricultural facilities, processing plants, and other assets where the real estate is only part of the story. This is the closing edition of our data-center series: power, water, noise and light, and now the ledger. If you are joining here, the earlier three stand on their own.

What the jobs evidence actually shows

Start with the honest numbers, because both sides of every county hearing are quoting the ones that flatter them.

The federal baseline is not in dispute. The Bureau of Labor Statistics counts roughly 460,000 employees nationally in NAICS 518 — data processing, hosting, and related services — in mid-2026, at average hourly earnings of $59.66. The Census Bureau, working from a different instrument, put the sector at 501,000 workers in 2023, up from 306,000 in 2016. That is real growth in a real, well-paid industry.

Then look at what a single campus employs. Not the sector — the building.

Abilene granted its project an 85% property-tax discount for twenty years of operation. Fifty-seven people work there. When the President described the Stargate program as producing “over 100,000 American jobs,” the reporting could identify no methodology behind the figure and no connection to site employment; it appears to describe indirect, industry-wide activity. Set it beside the fifty-seven and you have the entire rhetorical problem of this asset class in one comparison.

The academic evidence has now caught up, and it is more careful than either camp. Dany Bahar and Greg Wright have produced two related studies. The version published through Brookings in May 2026 examines roughly 1,500 facilities against 52 announced-but-cancelled projects and finds that counties receiving their first large data center see data-processing-sector employment rise 56% over the first decade, with telecommunications up 43%. Their separate academic working paper, dated August 2026, narrows to 341 satellite-verified hyperscale facilities and finds a 26% rise in data-processing employment — a figure that, as the authors note, includes the facility’s own jobs.

That second paper is the one that tests what everybody actually wants to know, and its answers are flat: advertised salaries near facilities rose 1.3%, statistically indistinguishable from zero; new firm registrations fell half a percent; supplier job postings moved within a confidence interval so wide it excludes nothing. The authors’ own summary: they “find little evidence” of local spillovers through labor pooling, supplier linkages, or firm clustering. Their sentence is better than any I could write: “A hyperscale campus can cost more than a billion dollars and draw hundreds of megawatts of continuous power, yet it may employ only a few dozen people on site.”

Against that, the industry’s own number. PwC, commissioned by the Data Center Coalition, reports 1,005,080 direct jobs in 2024 and a multiplier of “more than 4.5 jobs elsewhere” per direct job — revised down, notably, from “more than six” in the prior edition. That direct figure is roughly twice the BLS count for the same period, which almost certainly reflects a far broader definition of the industry: construction trades, facilities services, real estate, and IT services attributed to data-center activity rather than establishment employment in the sector itself. That is a defensible way to count economic contribution. It is not the same thing BLS is counting, and any hearing where the two numbers are quoted at each other is a hearing where nobody is going to learn anything.

The practical translation for professionals: permanent operations employment is a rounding error in these deals, and everyone in the room knows it. The construction phase is where the labor income is — Virginia’s JLARC found roughly 80% of the jobs and 70% of the GDP contribution occur during construction, at a peak of about 1,500 workers per project, over 12 to 18 months. Construction jobs are genuine, well-paid, and temporary. Whether a twenty-year tax abatement is the right price for eighteen months of them is the actual policy question, and it is rarely the one being argued.

The audits have turned

For a decade these programs ran without much scrutiny because the numbers were small. They are no longer small, and the auditors have arrived.

Georgia audited itself and had to publish a correction. The Department of Audits and Accounts released its evaluation of the high-tech data center exemption in late December 2025, crediting the program with 28,350 construction jobs and 5,471 operations jobs. On January 15, 2026, it revised those figures to 8,505 and 1,641 — cuts of roughly 70% — along with parallel reductions to the economic-impact estimates, attributing the error to a calculation problem in drafting the summary rather than to the underlying university analysis. The corrected report puts forgone state revenue at $474.2 million in FY2025 against $41.5 million generated back: a fiscal return of about four cents on the dollar. Cost per construction job: $42,646. And the finding that should end most arguments about additionality — the evaluators estimated that roughly 70% of the data-center construction would have occurred without the tax break at all.

(My own arithmetic, offered as illustration and not as a reported figure: $474.2 million of forgone revenue against 1,641 operations jobs is roughly $289,000 per permanent job — in a single year. That is a snapshot, not a lifetime cost-per-job, and it deliberately sets aside construction employment, which the audit prices separately. It is the number a county commissioner is going to compute on the back of an envelope, so it is worth having the honest version at hand.)

Virginia’s legislative auditor did the arithmetic differently and arrived somewhere equally striking. JLARC’s 2025 review of economic development incentives found the data-center exemption accounted for roughly 53% of all state incentive spending from FY2015 through FY2024 — $2.7 billion of $5.2 billion — and $1.02 billion in FY2024 alone, about 79% of that year’s total incentive spending. JLARC’s earlier standalone study, briefed in December 2024, reported companies saving $928.6 million in FY23 across state, local, and regional sales taxes, and called the exemption “by far the state’s largest economic development incentive.” Note what that means structurally: Virginia’s economic development policy is, to a first approximation, its data-center policy.

Texas found out what happens when you actually check. Under Tax Code § 151.359, a qualifying facility over 100,000 square feet must create at least 20 jobs at 120% of the area median wage and invest $200 million over five years; over 250,000 square feet, the thresholds rise to 40 jobs, $500 million, and 20 reserved megawatts. As of July 2026, 138 facilities were certified. Twenty had been audited. Six of those twenty were found non-compliant — four for failing to create the required jobs, one for square footage, one for a power purchase agreement that fell through. Those companies must repay previously waived sales tax. None have been named. Meanwhile the Comptroller’s chief revenue estimator testified that the program’s original 2014–15 biennium cost estimate was $14.6 million; the projection for the 2028–29 biennium is $3.3 billion, and he described even that as stale, with roughly ten new certifications arriving each month.

And nationally, the disclosure problem is the finding. Good Jobs First counts 37 states offering data-center-specific sales and use tax exemptions as of June 2026, up from 32 a year earlier, and reports that taxpayer costs “routinely exceed $1 million per permanent job.” Its more damning observation is procedural: only 11 of 36 states disclose which companies receive the breaks, and not one state reports both jobs promised and jobs actually created. Mississippi — home to the largest noise class action in the country, covered in the last edition — discloses nothing at all. There is no credible nationwide total for revenue foregone, and anyone who quotes you one has extrapolated it.

The clawback: drafted everywhere, tested almost nowhere

Every professional who reads an incentive agreement finds a clawback in it. Very few find one that has been exercised.

The structural weakness is usually in the same place. Virginia’s statute, for instance, requires the company’s memorandum of understanding with the state to include “repayment obligations should those goals not be achieved, and any conditions under which repayment … may be required” — but the statute leaves the formula to the negotiated MOU. The enforceable arithmetic lives in a document that is not the statute, is often not published, and was drafted by the party who benefits from it being soft. Georgia takes a firmer approach on paper — a seven-year compliance window with recapture of all exempted tax plus interest — which makes the Georgia audit’s own finding, that most of the construction would have happened anyway, the more interesting problem.

What a rigorous clawback looks like, for comparison: the model language circulating among reform advocates prorates recapture to the share of jobs not created; recaptures at twice the rate of job loss if the parent’s statewide employment falls below 90% of its base; and rescinds the subsidy entirely if job creation misses target by 25% or more for three consecutive years, with the obligation surviving for the life of the subsidy or five years, whichever is longer. The data-center-specific version adds a twenty-year post-completion operating requirement with 5% annual recapture for each year permanent headcount is not maintained. Whatever one thinks of the politics, that is what a provision designed to be enforced reads like — and it is not what most executed agreements read like.

Two honest caveats, because this is where advocacy usually outruns the record.

First, there is no verified example of a data-center clawback being triggered and recovered for a disclosed amount. Texas’s audit findings are the strongest evidence the mechanism is live, and the recovery figures are not public. The useful analogs come from outside the sector: North Carolina recovered $28 million from Dell after a plant closure; Ohio ordered General Motors to repay roughly $28 million over Lordstown; Illinois initiated recapture against eleven non-compliant projects in a single year. Clawbacks work when someone chooses to enforce them. That choice is discretionary, political, and rare.

Second — and this is the practical point for lenders and deal counsel — the real security in these transactions is migrating from tax recapture to utility collateral. When Wisconsin regulators revised their very-large-customer tariff in 2026, Oracle’s obligation for the Port Washington campus was reported as a letter of credit likely exceeding $7 billion, carrying an annual cost that could exceed $100 million, imposed expressly to address the risk of transmission costs shifting onto existing customers. No county’s clawback provision is going to reach that number. The regulator got there first, and got a bank behind it. If you are papering one of these deals, the collateral question is now a utility-commission question before it is a tax question.

The bill that arrives anyway

Which brings us to the transfer nobody voted on.

Edition one walked through the new large-load tariffs from the customer’s side — the minimum-billing percentages, the multi-year terms, the exit fees that can exceed the land’s value. Read them from the community’s side and they are something else: an admission by regulators that, absent intervention, the cost of serving these loads was going to land on residential ratepayers.

Ohio’s tariff is on appeal to that state’s Supreme Court, brought by the Ohio Manufacturers’ Association Energy Group — which tells you something useful: industrial customers, not just residential advocates, believe the allocation is still wrong.

The quantification has arrived alongside the rules. PJM’s independent market monitor reported that in the capacity auction of July 2026, data centers drove $6.3 billion of $16.4 billion in total capacity charges — roughly 38% — and that across the last four base auctions the figure is $29.4 billion of $63.6 billion, about 46%. The monitor’s characterization was blunt: the market operator “is continuing to act like it’s business as usual,” when what has occurred is “really a paradigm shift.” Harvard’s Electricity Law Initiative, surveying nearly fifty regulatory proceedings, framed the structural point a year earlier: the rate structures that once spread the cost of reliable service across everyone “are now forcing the public to pay for infrastructure designed to supply a handful of wealthy corporations.” Virginia’s own legislative auditor projected that a typical Dominion residential customer could see generation and transmission costs rise $14 to $37 a month in constant dollars by 2040.

And then Virginia did something genuinely novel. After a budget standoff in which the Senate sought outright repeal of the sales-tax exemption and the industry offered a lump sum, the biennial budget signed June 30, 2026 kept the exemption intact and created a Data Center Electricity Consumption Tax: $0.011 per kilowatt-hour on all electricity consumed at a Virginia data center — including electricity the facility generates for itself — billed monthly, with collections above $600 million a year refunded pro rata beginning in FY2028.

Read that carefully, because it is the most important structural development of the year. Virginia declined to tax the equipment and chose instead to tax the load. That moves the fiscal question out of the assessor’s office and onto the meter — a base that cannot be depreciated, cannot be abated by an assessment appeal, and cannot be avoided by building your own generation. Edition one argued that the megawatts were the property. Virginia has now agreed with that proposition and taxed it.

The backlash ledger

2026 is the year the political consensus cracked. The record needs precision, though, because the press coverage has been sloppy about what actually became law.

• Enacted or ordered. Washington narrowed its exemption by statute in June 2026, removing replacement server equipment and ending refurbishment certificates in the largest counties. Arizona’s FY2027 budget imposed a three-year pause on new incentive applications — and developers filed nearly thirteen years’ worth of applications in the two weeks before it took effect, which is its own commentary on how these programs are used. Ohio’s governor directed the Tax Credit Authority to stop accepting new exemption requests in May 2026, pending a legislative study, while stressing it was “not a data center ban.” Illinois paused processing of new Data Center Investment Program agreements as of July 1, 2026, honoring those already executed. New York’s governor issued an executive order in July 2026 imposing a one-year permitting moratorium on AI data centers over 20 megawatts. Oklahoma enacted a Ratepayer Protection Act in May 2026 requiring large loads to cover their own grid costs. And in Monterey Park, California, voters approved a permanent citywide ban on data centers on June 2, 2026 — reported as the first of its kind decided at the ballot box.
• Passed but not law. New York’s legislature passed a Responsible Data Center Development Act on June 4, 2026 by wide margins. The governor has not signed it. The moratorium in force is the executive order, which is narrower and legally distinct. Anyone describing New York as having “enacted a moratorium law” is describing something that has not happened.
• Introduced and died. Georgia is the instructive case. Six separate bills — repeal, sunset, suspension, and a construction moratorium — were filed across the 2025 and 2026 sessions. Every one of them failed. The Georgia exemption remains fully intact through December 31, 2031, notwithstanding its own state auditor’s four-cents-on-the-dollar finding. A ratepayer-protection bill from a Republican senator died when the chamber adjourned before a floor vote. The lesson for anyone modeling political risk is that audits do not repeal statutes; legislatures do, and mostly they haven’t.

Locally, it is a different story. Good Jobs First counted 54 local moratoriums already adopted and 9 pending as of March 2026, in places as different as Denver and Ypsilanti. State-level incentive policy has held. Local land-use policy has not.

What replaces the handshake

If the tax exemption is politically fixed and the clawback is rarely enforced, the negotiating action moves to the instruments that actually bind. Three are worth knowing by name.

The community benefits agreement. The best-documented executed example in this asset class is the Lancaster AI Hub in Pennsylvania, negotiated in late 2025: $20 million in community contributions paid upfront and split between a county community foundation and a municipal clean-energy fund, secured by a $20 million letter of credit or a corporate guarantee from an entity with at least $100 million in net worth; a hard cap of 20,000 gallons per day of municipal water per campus; a clean-energy requirement backed by tiered penalties of up to $10 million per building; noise limits benchmarked to pre-construction ambient levels; a local hiring plan and first-responder training. It is also instructive for what it lacks — no specific hire targets, no independent oversight body, no air monitoring. And note the sequencing, because it is the reason the rest of it has teeth: the agreement was executed in November 2025, while the city was still preparing the zoning amendments the project needed. The money was secured while the entitlement was still outstanding. For counsel on either side, that ordering is the whole ballgame. Whatever is not locked down before the approval is negotiated from a weaker position afterward, if at all.

The decommissioning bond. Susquehanna County, Pennsylvania adopted the most specific formula I have found, in an ordinance effective at the end of 2025: decommissioning required within twelve months of end of useful life, defined as twelve consecutive months of non-use; security equal to net decommissioning cost plus 10% — gross cost less equipment salvage — estimated by a licensed professional engineer after the first year of operation and re-estimated every five years; posted as a bond or county-approved assurance before construction begins; plus general liability coverage of at least $3 million per occurrence. That is a serious instrument, and it raises a question every appraiser and lender should be asking about the salvage assumption: the formula nets out equipment value, which means the bond is sized on a residual estimate for hardware whose useful life is being compressed by the same technological churn that built the facility. Get the salvage assumption wrong in year one and the security is thin by year ten.

The abatement structure itself. Property tax is where the real estate professionals earn their keep here, because the mechanisms vary enormously. South Carolina’s fee-in-lieu regime can negotiate the assessment ratio down from 6% on real property and 10.5% on personal to as low as 4%, with millage frozen or reset on a five-year cycle over a twenty- to thirty-year term. Ohio counties are offering choices between a fifteen-year, 50% abatement with tax-increment financing and a thirty-year, 100% abatement through a community reinvestment area; one nearby comparable pairs a fifteen-year, 75% abatement with a $500,000 annual payment in lieu of taxes. The Tax Foundation’s survey of a dozen major jurisdictions found tangible personal property tax accounts for more than 20% of the total tax burden on these facilities, with effective combined rates ranging from about 21% in Cheyenne to roughly 80% in Santa Clara County. The same building, the same equipment, the same operator: a fourfold spread in effective tax rate depending on the county line. That is not a footnote in a pro forma. It is often the deal.

What this does to value

Four consequences, in the order they will reach your clients.

1. The incentive is a value component, and its assignability is an open question. If a campus trades, does the abatement, the fee-in-lieu, the certification, and the community benefits obligation follow the buyer? Sometimes explicitly; often only with agency or county consent; occasionally not at all. This is the same structural problem edition one identified with the power contract, and it has the same answer: read the instrument, condition the closing on consent, and escrow the value of anything that might not transfer. An appraisal that reports value without stating whether the incentive package was assumed to survive the transaction has not answered the question that was asked.
2. Assessors are no longer conceding. Cook County’s assessor defended data-center valuations at appeal hearings through mid-2026 with no reduction granted to four contested properties — including assessments of $106 million, $102 million, and $70 million — against owner requests for cuts ranging from 24% to 90%, while a separate operator obtained reductions between roughly 29% and 68% by settlement. The methodological fight is over land rates. Expect more of these, expect them to be well-briefed on both sides, and expect the appraisal question — what portion of value is land, improvements, equipment, and intangible operating advantage — to be genuinely contested rather than stipulated.
3. Fiscal concentration is a risk factor, and one county is saying so out loud. Loudoun County, Virginia derives roughly 38% of its general fund from data centers occupying about 4% of its commercial parcels, with more than $100 million a year in new personal-property tax revenue. That revenue funded successive residential rate cuts and the elimination of a vehicle license fee. And the county’s own published materials warn that this revenue growth “will plateau in the next five to 10 years.” The reason is depreciation: Loudoun assesses business personal property on a steep declining schedule — 60% of original cost in the first year, then 45%, 30%, 15%, 10%, and 5% for anything older. Nearly two-thirds of the taxable value of a server is gone by year four. Front-loaded revenue funds a permanently lower residential rate — a structure that works beautifully while the buildout continues and becomes a budget problem the moment it doesn’t. Any client underwriting residential or commercial real estate in a data-center-concentrated jurisdiction should understand that the local tax rate they are modeling is being subsidized by an accelerating depreciation schedule.
4. The decommissioning liability is real, unbonded in most jurisdictions, and nobody has priced it. The equipment in these buildings has a short and shortening life. The buildings themselves are single-purpose in a way that a warehouse is not. Susquehanna County’s bond formula is the exception, not the rule, and where no bond exists, the restoration obligation is an unsecured contingency sitting with an entity that may not outlive the asset. That is a valuation issue, a credit issue, and — for the CPAs — a disclosure question that belongs in the conversation before it belongs in a footnote.

Ten questions before your client signs

1. What does the incentive agreement actually promise in jobs, wages, and capital — and where is the recapture formula written, in the statute or in an unpublished MOU?
2. Who verifies compliance, on what schedule, and has this jurisdiction ever audited a data center? (Texas has audited fewer than one in six.)
3. If the recapture is triggered, what secures it — a guaranty, a letter of credit, a lien, or nothing but the operator’s willingness to pay?
4. Do the abatement, the fee-in-lieu, the certification, and any community benefits obligation survive a sale, a change of operator, or a foreclosure — and whose consent is required?
5. What is the county’s revenue concentration, and what depreciation schedule is that revenue running on? What happens to the residential rate when the buildout pauses?
6. Which utility rate class applies, what minimum billing and term does it impose, and how much collateral does the commission require — this is now where the real security in the deal sits?
7. Is there a decommissioning obligation, is it bonded, how is the bond sized, and what salvage assumption is it netting out?
8. What did the developer commit to in the conditions of approval versus in a side agreement negotiated afterward — and which of the two is actually enforceable?
9. If the state enacts a consumption tax on the load, as Virginia did in June 2026, does the pro forma survive it, and who bears it under the lease or service agreement?
10. For the client on the other side of the fence: what did this jurisdiction give up, what does the public record show it received, and what does that establish for an assessment appeal?

Where the series lands

Four editions, four constraints, one recurring lesson: the value is not where the camera points.

Power lives in tariffs. Water lives in revocable permits. The buildable envelope lives in an octave-band table. And the fiscal bargain — the part everyone argues about at the county hearing — lives in a recapture formula that is usually written somewhere other than the statute, verified rarely, and enforced almost never.

Across all four, the pattern holds: the attributes that decide what these properties are worth, and what they cost the people who live near them, are contractual and regulatory rather than physical. None of them appear on a title report. Most of them do not appear in the appraisal, either, unless somebody asks.

That is the work our team does on complex real estate and personal property: inventorying what a client actually owns and actually owes across land, improvements, equipment, contracts, and regulatory positions, and turning that inventory into numbers counsel, CPAs, and lenders can defend. Data centers were a useful place to demonstrate the method because they compress twenty years of that problem into thirty-six months. The method is not about data centers. The next asset class that comes apart this way will look nothing like these buildings, and the questions will be the same ones.

A question for your practice: When your client’s incentive agreement says the jobs must be created, does anyone in the room know who checks — and what happens to the deal if the answer is nobody?

If a client’s holdings raise these questions — an incentive package of uncertain assignability, a jurisdiction whose tax base has concentrated in one asset class, a facility where the real estate, the equipment, the contracts, and the public obligations intertwine — that intersection of complex real estate and personal property is where our team works. We regularly speak to law firms and legal teams (including CLE-style briefings), CPA firms and societies (including CPE-style sessions), lenders, family offices, owner-users, and industry groups on complex property and specialized asset issues. If your firm would like a private session on these topics, feel free to reach out.

Daniel Boring, CRE®, MAI, ARA, ASA | Senior Vice President – Valuation Advisory Services | Kidder Mathews

Sources and further reading

Employment research: Dany Bahar & Greg Wright, “New evidence on data center employment effects,” Brookings (May 4, 2026, updated Aug. 10, 2026); Bahar & Wright, “The Local Economic Impact of Data Centers,” working paper (Aug. 7, 2026); Daniel Goetzel, Mark Muro & Shriya Methkupally, “Turning the data center boom into long-term, local prosperity,” Brookings (Feb. 5, 2026); Anthony F. Pipa & Adam Aley, “The local implications of data centers for rural communities in the US,” Brookings (Mar. 2, 2026); U.S. Bureau of Labor Statistics, NAICS 518 industry data (extracted Aug. 2026); U.S. Census Bureau, “Employment in Data Centers Increased by More Than 60% From 2016 to 2023” (Jan. 6, 2025); Michael J. Hicks, “Data Centers and Local Job Creation,” The Country Economist (Nov. 10, 2025) — self-published; PwC for the Data Center Coalition, 2026 Impact Study — industry-funded.

State audits and fiscal analysis: Georgia Department of Audits and Accounts / Carl Vinson Institute of Government, “Tax Incentive Evaluation: Georgia Data Center Sales & Use Tax Exemption” (Dec. 2025, rev. Jan. 2026), with correction coverage by Georgia Public Broadcasting and The Current GA (Jan. 15–16, 2026); Virginia JLARC, “Data Centers in Virginia,” Report 598 (Dec. 2024) and “Economic Development Incentives 2025,” Report 611 (Nov. 10, 2025); Virginia Department of Taxation / VEDP biennial exemption report RD40 (Jan. 2, 2026); Georgia Office of Planning and Budget, FY2027 Tax Expenditure Report; Texas Tribune, “Audits lag for Texas data centers getting sales tax break” (July 27, 2026) and Texas Public Radio coverage of Senate Finance testimony (July 28, 2026); Policy Matters Ohio, “Indefensible tax breaks for data centers will cost Ohio” (Jan. 7, 2025) — advocacy analysis; Tax Foundation, “State Taxation of Data Centers” (Dec. 2025, upd. Feb. 2026).

Subsidy and clawback research: Good Jobs First, “Even Cloudier with a Greater Loss of Spending Control” (June 2026), “Cloudy Data, Costly Deals” (Nov. 2025), “Cloudy with a Loss of Spending Control” (Apr. 2025), “Data Center Moratorium Bills Are Spreading in 2026” (Mar. 9, 2026), and model clawback legislation; ProPublica, Lordstown clawback coverage.

Utility regulation and ratepayer impact: PUCO Case No. 24-508-EL-ATA (AEP Ohio Schedule DCT, approved July 9, 2025) and Ohio Supreme Court Case No. 2025-1458; Georgia PSC Docket 55378 (eff. Feb. 1, 2025); Virginia SCC, Dominion Energy Virginia 2025 Biennial Review (order of Nov. 25, 2025); Indiana URC Cause No. 46097; Texas SB 6 (2025); Wisconsin PSC very-large-customer tariff and The Register reporting on Oracle’s collateral obligation (July 21, 2026); Monitoring Analytics / PJM capacity auction reporting via Utility Dive (July 20, 2026); Eliza Martin & Ari Peskoe, “Extracting Profits from the Public,” Harvard Electricity Law Initiative (Mar. 5, 2025); Williams Mullen and Cardinal News coverage of the Virginia Data Center Electricity Consumption Tax (June 2026).

Ordinances, agreements, and assessments: Susquehanna County, PA Ordinance No. 2026-03 (adopted Dec. 29, 2025); Lancaster AI Hub community benefits agreement, via Federation of American Scientists and the Columbia Climate Law Blog CBA database; NAACP model data-center CBA template (2026); Food & Water Watch, “Tax Incentives and Data Centers: Warnings From Adams County, Ohio” (May 2026); Cook County Assessor’s Office, statement on defending data-center assessments at appeal (2026); Loudoun County, VA, “Data Centers – Tax Revenues and the County Budget”; Michigan HB 6135–6142 (introduced June 25, 2026).

Case details, figures, and identifiers discussed from public sources are cited to those sources; allegations in pending litigation and matters under appeal are exactly that. The per-permanent-job calculation in the Georgia section is the author’s arithmetic applied to that state’s published audit figures and is presented as illustration, not as a reported statistic. The discussion is grounded in real-world appraisal experience and is presented for educational purposes only. It is not valuation, legal, tax, utility-regulatory, or investment advice for any specific property, transaction, or dispute. Incentive programs, utility tariffs, ordinances, and pending legislation vary materially by jurisdiction and change frequently — several matters described here were unresolved as of publication; readers should confirm current requirements with the relevant agency, commission, and counsel.