Methodology & sources

Revenue and expenditure figures come from the US Census Bureau's 2022 Census of Governments Individual Unit File (FY2022, the latest full-coverage vintage, revision of July 2026) — the audited, unit-level record of every local government in the country. We aggregate all ~89,900 units (counties, cities, townships, school districts, special districts) to county areas. "Total local revenue" sums taxes, charges, and federal/state aid; transfers between local governments inside the same county are excluded so nothing double-counts. Function budgets are current operations plus construction for K-12 education, police, fire, and parks. FY2022 dollars are uplifted ×1.10 to current dollars (CPI-U annual average 2022→2025, BLS series CUUR0000SA0); the uplift is a documented approximation — county budgets grew at different rates, and the counties that gained a datacenter in that window grew fastest.

Two things to read carefully in the percentages. "Total local revenue" is a revenue figure, not a tax figure — it includes federal and state aid, so a tax measured against it looks smaller than it does against locally raised money alone. And the county government is not the main recipient: across Texas the median split of the property-tax levy is roughly 49% to school districts, 29% to the county, 9% to cities and 10% to special districts. A datacenter's largest single beneficiary is usually the school district.

For 608 counties across Texas, Ohio, Indiana, Nebraska and Mississippi the rate is observed, not modeled. It is built from each state's own published levy data — the Texas Comptroller's Reports of Property Value, Ohio's Abstract of the Tax Duplicate, Indiana's DLGF certified district rates, Nebraska's Certificate of Taxes Levied, Mississippi's Combined Millage Rate Schedules — and converted to an effective rate on market value using that state's own assessment ratio. Ohio publishes the commercial-class rate separately from residential; Texas requires reconstructing it from the overlapping county, school and special-district levies. Each state's method and its limits travel with the data.

Everywhere else the rate is a proxy: median real-estate taxes paid ÷ median home value (ACS 2023 5-year, tables B25103/B25077) — a residential rate standing in for a commercial one. Measuring it against real Texas rates across 249 counties shows the cost of that substitution: it runs about 1.22× the true rate in counties under 10,000 households and 0.86× in counties over 50,000 — the error changes direction with county size — and its spread is roughly four times wider in the small counties. Datacenters are built in small counties, so the proxy is least reliable exactly where this page is most often read.

Until August 2026 that proxy was scaled by a commercial-to-homestead classification ratio taken from the Lincoln Institute's 50-State Property Tax Comparison Study. That adjustment has been withdrawn in the 24 states where it came from a single large city. The ratio is computed inside one municipality — its commercial rate over its own homestead rate — while the model applied it to a county-wide residential rate for a property in neither; those two denominators diverge by 0.8× to 1.65×. Tested against real Texas rates it made the answer worse, not better. Statutory assessment classes, which are genuine classifications rather than effective-rate ratios, are unaffected.

No county-level rate can be exact. A county contains many overlapping levy districts, and in Texas the composite spread between the highest- and lowest-taxing district inside a single county runs a median of 26% and reaches 75%. Which school district a site lands in moves its bill by more than most of the modeling choices on this page. Where the source is district-level, that range is carried alongside the rate.

Fourteen states levy no tax on business tangible personal property (DE, HI, IL, IA, MN, NH, NJ, NY, ND, OH, PA, SD, plus Wisconsin since the 2024 repeal and Kansas for post-2006 equipment); equipment tax is zero there. Separate-rate jurisdictions use their actual rates: Virginia at Loudoun's $4.15/$100 on datacenter computer equipment with the county's own declining depreciation schedule (60%→5% floor, averaging ~32% of original cost over a 5-year refresh); DC $3.40/$100; Maryland's weighted-average local rate $3.17/$100; Kentucky's state + county tangible composite. Classification states apply their statutory TPP-to-residential assessment ratios (CO 26%/6.8%, MO 33.3%/19%, AL 20%/10%, and others; South Carolina's is 6.19× rather than the nominal 10.5/4 because Act 388 strips school operating millage out of the owner-occupied bill but not a datacenter's, and Rhode Island's is the statutory 1.5 ceiling rather than Providence's 3.3, Providence being one of the few municipalities the cap does not bind), and West Virginia's statutory salvage valuation prices qualifying servers at 5% of original cost. Everywhere else the fleet is valued at a steady-state 55% of original cost — the 5-year straight-line midpoint the Tax Foundation's "State Taxation of Data Centers" study uses. Every state row carries its statute citation.

Only ~55% of a datacenter's non-IT construction cost is real property (land and building shell); the rest — chillers, UPS and batteries, generators, switchgear — is assessed as tangible personal property in practice (Tax Foundation reference build ≈ 54/46; Texas appraisal districts schedule generators and UPS as business personal property; Virginia's statute places substations, UPS, electrical plant and air handlers in the datacenter TPP class, Va. Code §58.1-3506(A)(43)). The model taxes the real share at the county's real-property rate and the equipment share under the state's TPP rules with a slower ~65% steady-state factor (10–25 year asset lives). Observed anchor: Loudoun County collected $894.5M from datacenters in FY2025 — $670M of it (75%) from personal property (Loudoun OMB / Commissioner of the Revenue figures, via the Loudoun Coalition FY2026 budget review).

The model is more personal-property-heavy than that. Run on Loudoun's own rules it books about 9% of the bill as real property against the county's actual 25%, and the gap widens each year as shells finish and assess — Loudoun's FY2026 budget already puts personal property at 70% rather than 75%. Because both halves scale with capacity the comparison does not depend on knowing Loudoun's megawatts, which is what makes it a usable check. Treat the real-property line here as a floor, and the equipment line as carrying more of the total than it probably should.

38 states have enacted datacenter sales-tax exemption programs; where one is open to a new project, the equipment sales-tax line is $0 with the program named (minimum-investment thresholds shown). Where none applies, sales tax on equipment purchases (Tax Foundation Jan-2026 state + average local rates) is annualized over the 4-year GPU refresh cycle and shown as a separate estimate — it never enters the property-tax totals or budget percentages. Programs a new project cannot currently access are treated as absent: Arizona's application moratorium (July 2026–June 2029), Ohio's June 2026 pause, Nebraska's LB 901 repeal and executive order, and Massachusetts's not-yet-open certification.

Defaults are $11M/MW for the facility (Turner & Townsend 2025 US index: $9.5–13.3 per IT-load watt, +7–10% for liquid-cooled AI builds) and $30M/MW for GPUs and IT equipment (GB200-class racks at ~$3–3.5M per ~120 kW imply $25–38M/MW; announced frontier campuses run $35–45M/MW), on a critical-IT MW basis.

These are construction and purchase costs, and the model uses them directly as the assessment base. That is right for equipment — Virginia, Texas and Nevada all assess business personal property as a percentage of original cost — but real property is assessed at fair market value, which need not equal what a building cost to build. No cost-to-market adjustment is applied, and its direction is jurisdiction-specific: a cost-approach assessment net of obsolescence lands below cost, while an income-approach valuation of a leased shell can land above it.

The Tax Foundation's $1B / 32 MW reference facility was previously cited here as a cross-check on these defaults. It is not one. Decomposed on the same basis it implies $13.1M/MW of facility and $18.2M/MW of servers, so these figures sit 16% below it on the building and 65% above it on the equipment. The divergence is deliberate — that reference is a generic datacenter and AI builds are equipment-heavy — but a figure the model departs from in both directions cannot validate it. Real deals almost always carry negotiated abatements — the abatement slider defaults to 0% so the headline shows full statutory tax, and where a state has a clean statutory program (Nevada's 75% TPP abatement, Connecticut's qualified-data-center exemption, New Mexico's IRB structure, Mississippi's fee-in-lieu, South Carolina's FILOT) a one-click control applies its typical percentage, always labeled as the program default rather than any specific deal.

A datacenter's power bill carries its own tax stack: state sales/use tax on industrial electricity where it applies (many states exempt utility-delivered power or manufacturing use — datacenters usually don't qualify as manufacturing), utility gross-receipts or per-kWh excises embedded in rates (Pennsylvania's 59-mill gross receipts tax, Washington's 3.87% public utility tax, Illinois' electricity excise), datacenter-specific provisions (Tennessee's reduced 1.5% rate; Wisconsin's certified-datacenter electricity exemption; Minnesota's 2025 repeal; Iowa's 2025 cap) — and, first in the nation, Virginia's $0.011/kWh datacenter electricity consumption tax (July 2026–June 2028, capped $600M/yr statewide). Consumption is modeled as critical-IT MW × 8,760 h × PUE 1.25 × the state's EIA industrial price. These are predominantly state taxes, so they appear as a separate context line and never enter the county budget percentages. Local utility franchise fees are out of scope — no national dataset exists.

Employment and pay figures marked observed come from BLS QCEW (NAICS 518210, 2024 annual averages; county rows where disclosed, state fallback where suppressed) — e.g. Loudoun's computing-infrastructure industry pays $212,811 on average, 2.5× the county's all-industries wage. Construction and operations headcounts for a hypothetical build are estimates from published national coefficients (0.7–2.0 construction workers and 0.15–0.35 permanent FTE per critical-IT MW; PwC/IMPLAN and US Chamber/RIMS II studies) — county-specific multipliers are paid BEA products and are not used. The two layers carry distinct badges and never blend.

Where a county's own audited financial statements (ACFR statistical sections), assessor top-taxpayer reports, or bond official statements publish principal-taxpayer tables, they appear verbatim with the source linked — the highest-trust tier on the page. Summing the identified datacenter entities over the county's taxable base gives an observed floor on the datacenter share: top-10 lists truncate, only identified entities count, and Virginia tables cover real property only (computer-equipment personal property — the larger half of datacenter tax there — is extra). The homeowner counterfactual applies fixed-levy reallocation to that floor: if the datacenter share s disappeared and local governments collected the same total, every remaining bill rises by s/(1−s), applied to the county's median owner-occupied bill (ACS 2023 5-year, B25103). Because the expression only grows with s, a floor share yields a floor counterfactual — the figures read "at least." Counties whose published tables use a non-ratioable basis (taxes-paid rankings, abatement-inflated account values) show the table but no share.

Datacenters pair enormous assessed value with near-zero service demand — no students, minimal police, fire, and road load per tax dollar. Cost-of-community-services studies (American Farmland Trust meta-analysis) find commercial/industrial property consumes ≈ $0.30 of services per $1.00 of tax paid versus ≈ $1.15 for residential — so the realistic alternative land use is typically a net fiscal drain.

• Connecticut is greyed out: CT reports under 2022 planning-region FIPS that do not nest inside the county boundaries the map draws. Population-weighted apportionment is a tracked follow-up.
• New York City's five boroughs are one consolidated government and are modeled as one area.
• The Census of Governments carries stub records for some county governments — Montezuma County, CO has a single revenue line in the whole file — so 165 counties show no "share of county government revenue" rather than a ratio computed against an incomplete denominator. Vermont's counties and a handful of others are suppressed for the related reason that their county government is not a meaningful fiscal entity.
• Hawaii's K-12 line is $0 locally — its schools are state-run; school equivalents there are not meaningful.
• Some states offset local school-revenue gains with reduced state aid or recapture; school figures are gross of that offset (a per-state retention model is a tracked follow-up).
• Estimates are flagged: any figure marked estimate names its anchor inputs. Observed government data is never blended with modeled figures.