01 · Framework
Net financial position
Every existing data center risk tool we reviewed shows only costs: noise, air quality, property value depression. This is incomplete. Data centers also generate substantial tax revenue that flows to host counties as property tax reductions, school funding, and service quality. A household living far from a facility but inside the host county may receive thousands of dollars in annual benefit with almost no lived cost.
Datatox calculates both sides and presents the net annual position:
net = (tax_savings + school_benefit + service_benefit) − (electricity_cost + healthcare_cost + property_value_risk)
Each input scales with the household's actual circumstances — income, home value, household size, school-age children, and respiratory or cardiovascular sensitivity.
02 · Location tiers
Fourteen structural positions
Each location in the database is classified into one of 14 structural tiers based on host-county relationship, distance to nearest facility, and grid region. The tier determines the economic profile.
| Tier | Description | Example city |
|---|
| host_dense | Inside host county, <2mi from facility cluster | Ashburn, VA |
| host | Inside host county, 2-10mi from cluster | Leesburg, VA |
| host_distant | Inside host county, >10mi from facility | Purcellville, VA |
| fairfax_host | Fairfax County (smaller tax base than Loudoun) | Herndon, VA |
| adjacent_restrict | Adjacent county that restricts development | Warrenton, VA |
| stratos_host | Box Elder County, hosts Stratos project | Tremonton, UT |
| stratos_externality | Wasatch Front, downwind of Stratos | Salt Lake City, UT |
| distant_pjm | Outside direct exposure, on PJM grid | Charlottesville, VA |
03 · Benefits
How tax and service benefits are calculated
Property tax reduction is the central benefit in host counties, and it is the best-evidenced number in this model. Data centers made up about 42% of Loudoun County's local tax funding in 2026, up from 32.8% in 2024, paying roughly $894.5 million in property taxes in 2025 against a projected $1.14 billion in 2026.
Mangum Economics, for the Northern Virginia Technology Council (2026), estimates that without that base Loudoun's residential real property rate would have to rise from $0.805 to about $1.537 per $100 of assessed value — roughly a doubling, and about $5,800 a year for the typical homeowner. The rate difference, 0.732% of assessed value, is what this model uses for Loudoun-tier addresses. Loudoun has cut its rate in each of the last ten years, from $1.145 per $100 in 2016.
tax_savings = home_value × tax_rate_per_dollar_of_home_value
The rate varies by county:
| County | Rate | $5,800 example |
|---|
| Loudoun, VA | 0.81% of home value | $720K home → ~$5,800 |
| Prince William, VA | 0.78% | $720K home → ~$5,600 |
| Fairfax, VA | 0.35% | $720K home → ~$2,500 |
| Henrico, VA | 0.45% | $720K home → ~$3,200 |
| Box Elder, UT | 0.15% | $720K home → ~$1,080 |
Benefits require a host relationship. Tax, school and service benefits all flow from a data center paying property tax into your county. An address far from any facility is almost certainly in a county with no such base, and so receives essentially nothing. An earlier version of this model gave distant addresses in several states a benefit 15–30× larger than the equivalent curated Virginia tier, which meant much of the country was credited with a four-figure annual gain from facilities that were not there. Those tiers now match the curated scale.
School funding benefit is the in-kind value of higher per-pupil spending in host counties. Loudoun County Public Schools spends roughly $18,500 per pupil — among the highest in the U.S. — funded substantially by data center revenue. We treat the marginal benefit (over the state median) as ~$4,200 per public-school child per year.
Service quality benefit is a flat annual estimate for the value of higher-quality county services (parks, roads, public safety) funded by the commercial tax base. We use $800 for Loudoun, scaled proportionally elsewhere.
04 · Costs
How household costs are calculated
Water and sewer rate impact follows the same logic as electricity, at smaller scale. Municipal water systems recover largely fixed costs across their ratepayers, so a large new industrial draw shifts some of that burden onto households; arid regions carry a 1.6× multiplier. This is the least precise line in the ledger — local rate structures vary enormously — and it is reported as such.
Jobs appear in the benefits ledger at $0, deliberately. A 100 MW campus supports 100–200 permanent staff, roughly 1.5 per MW; construction employs five to ten times that and then ends. The local income those jobs generate reaches households through the county tax base, which the property-tax line already counts — adding a dollar figure for jobs as well would count the same money twice. The row is shown with the real headcounts because it is the most-cited benefit at rezoning hearings and the most consistently overstated.
Electricity cost impact reflects elevated PJM (or Utah) grid prices driven by data center demand. We use a base annual cost per tier and scale upward modestly with household income, on the principle that higher-income households generally have larger homes and higher consumption.
electricity_cost = base + (income / 100,000) × scaling_factor
Healthcare risk premium reflects measurable health impact from elevated PM2.5 and NO2 emissions from backup diesel generators and on-site gas turbines. The Harvard EmPower Analytics group estimates 3.4–6.5 premature deaths per year per single hyperscale facility, with 1,300 deaths projected nationally by 2028.
For Datatox, we translate this into household-level expected costs using sensitivity adders:
| Sensitivity factor | Annual cost (host_dense tier) |
|---|
| Baseline per household member | $60 |
| Asthma diagnosis in household | +$950 |
| Cardiovascular condition | +$700 |
| Member aged 65+ | +$550 |
| Pregnancy | +$600 |
These adders are calibrated for the host_dense case. They are then scaled by an exposure factor derived from the address's own health composite, so a household with asthma forty miles from the nearest facility no longer carries the same $950 premium as one half a mile away. A composite of 6.5 — typical of a dense cluster — pays the full adder; the factor falls to a floor of 0.2, which preserves a residual premium for regional grid emissions that reach far beyond any cluster.
exposure_factor = clamp(0.2, 1.0, 0.2 + (health_composite / 6.5) × 0.8)
Property value risk is the most contested number in this model, and it has been cut sharply.
An earlier version annualised a "7–12% depression within half a mile" into about 0.85% of home value per year, which made it the largest single cost in the ledger. That figure describes the worst observed cases near established clusters; it is not the average, and treating it as one was wrong. Measured across the market, the current evidence is close to neutral:
| Study | Finding |
|---|
| George Mason, Center for Regional Analysis (2025) | Homes closer to data centers sold for higher prices — across detached, townhome and condo alike |
| Realtor.com (2026) | 42% appreciation near facilities vs 41% in surrounding areas |
| University of Rochester | Little measurable effect on nearby home prices |
| Separate GMU-led analysis | New data centers slowed local home-price growth |
The GMU price result is very likely confounded — data centers are sited where infrastructure, jobs and amenities already are, which is also where houses are expensive. That does not make it evidence of harm either.
property_risk = home_value × tier_rate × 0.25 × concentration
concentration = clamp(0.1, 1.0, health_composite / 6.5)
So the model retains a quarter of the original risk, concentrated where physical exposure is concentrated and close to zero elsewhere. That residual represents the two things the evidence does support: the slowed-appreciation effect near new construction, and the genuine near-field cases within half a mile of the largest clusters. Researchers expect the picture to worsen as facilities get bigger; if it does, this number should go back up.
05 · Projection model
10-year evolution of net position
The projection compounds the current annual net position with three drivers:
Facility expansion — host counties continue to add capacity. For host_dense tier, we model healthcare and property-value costs growing at 4% per year; for host_distant tier, costs grow more slowly (1.5%) as new facilities encroach.
Tax base maturation — benefits in host counties also grow as the tax base expands, but at a slower 2.5% per year, as new facilities pay reduced effective rates under economic development incentives.
Electricity rate growth — PJM capacity prices are growing at roughly 6% per year, more than 2x general inflation, with data center demand identified as the primary driver in PJM's own filings.
year_N_net = year_0_net × growth_factor_tier(N)
06 · Datatox score
Two scores, one formula
Datatox now reports two related scores on each report: the Location Score (fixed by your address) and the Personal Risk Score (which adjusts for household sensitivities).
Location Score (0–100)
Every city in the database runs through the same unified formula. There are no tier-specific score functions — the differences between cities come from their structural facts and four tier-level weights.
Location Score = Base (25) + Direct Proximity (0–35) + Cluster Density (0–20) + Jurisdictional Penalty (0–15) + Regional Risk (0–20) − Tax Base Offset (0 to −15)
Direct Proximity is hyperbolic in distance: 35 / (1 + miles / 2). Half points at 2 miles, quarter points at 6 miles, near zero by 30 miles.
Cluster Density is min(20, facility_count_within_2mi × 1.5). Each nearby facility adds 1.5 points, capped at 20.
Jurisdictional Penalty reflects bearing externalities without local benefit: 0 for host counties, 15 for adjacent counties that absorb regional costs, 8 for regions with cross-state effects, 6 for protected zones receiving spillover.
Regional Risk is a per-tier constant capturing water, air, or grid risk independent of distance (e.g., West Texas water stress, Great Salt Lake watershed, Northern Virginia grid demand).
Tax Base Offset subtracts from the score in proportion to host-county fiscal benefit, capped at −15.
Soft ceiling. The components can sum past 100 at the densest addresses. Clamping there flattened the top of the scale — Ashburn, Sterling and Manassas all returned exactly 100, which is useless precisely where the differences matter. Scores above 85 are now compressed asymptotically instead, so ordering survives and the scale still tops out at 100.
score = raw raw ≤ 85
score = 85 + 15 × (1 − e^−(raw − 85)/25) raw > 85
Personal Risk Score (0–100)
Personal Risk layers household sensitivities onto the Location Score. Each sensitivity adder is weighted, then multiplied by an exposure scalar so distant locations don't dramatically amplify household risk.
Personal Risk = Location Score + (Asthma × 10 + Cardio × 7 + Elderly × 5 + Pregnancy × 6 + Household-size × 0.5) × exposure_scalar
exposure_scalar = min(1, max(0.15, Location Score / 75))
The scalar means a household with asthma at an Ashburn address (Location Score ≈ 70) gets close to the full +10 adder, while the same household in distant rural Texas gets only +1.5. This reflects epidemiology: health risk from data-center emissions is concentrated near sources.
Bands
| Range | Band | Interpretation |
|---|
| 0–17 | Minimal concern | No direct exposure; only diffuse grid effects |
| 18–29 | Low concern | Distant regional effects, no direct impact |
| 30–44 | Low-moderate concern | Diffuse regional effects + grid exposure |
| 45–59 | Moderate concern | Real direct or jurisdictional exposure |
| 60–74 | Moderate-high concern | Substantial direct exposure or major regional risk |
| 75–100 | High concern | Dense cluster within close range |
07 · Capacity model
Why facility size matters
The original Datatox Score treated every nearby data center as one count. A 30 MW colocation building and an 1,200 MW hyperscale facility counted the same. This was a real flaw — capacity drives essentially every direct health impact, and the model now reflects that.
Each city in the database has an estimated mwNearby figure representing total IT capacity within roughly 5 miles, distance-weighted. For cities with named major projects (Stargate, HyperGrid, Stratos, Project Sail, etc.), we use published capacity figures attenuated by distance. For cities without named projects, we use tier-typical capacity per facility multiplied by the count of nearby facilities.
Published capacity figures used
| Project | Location | Capacity (MW) | Status |
|---|
| HyperGrid | Amarillo, TX | 11,000 | Planned |
| Stratos | Box Elder, UT | 9,000 | Approved |
| Data City | Laredo, TX | 2,000+ (phase 1) | Planned |
| Tract Caldwell Valley | Austin-San Antonio, TX | 2,000 (500 MW approved) | Approved |
| OpenAI Stargate | Abilene, TX | 1,200 | Operational |
| Project Sail | Coweta, GA | 900 | Approved |
| Fort Bliss (Carlyle/CyrusOne) | El Paso, TX | ~500 | Planned |
| DC BLOX Atlanta East | Conyers, GA | 216 | Under construction |
Capacity exposure scalar
Raw MW is converted to a 0-1 scalar using a logarithmic curve, because the per-resident impact of capacity is roughly logarithmic — doubling capacity doesn't double the impact at any given distance, but capacity does increase impact along a log-scale.
capacity_exposure_scalar = min(1, max(0, log₁₀(MW) / 4))
Examples:
10 MW → 0.25
100 MW → 0.50
1,000 MW → 0.75
10,000 MW → 1.00 (capped)
This scalar then weights the Direct Proximity and Cluster Density components of the Location Score, so being 1 mile from Stargate (capacity scalar 0.78) contributes much more to the score than being 1 mile from a 30 MW colo (capacity scalar 0.37).
08 · Health factors
Eight exposure categories
The Location Score is a composite — but composites can hide which specific risk is driving the number. The "Show calculation" panel breaks down eight distinct exposure categories, each scored 0-10 with its own analysis.
Three of these were added after the first version: vibration, construction phase and night lighting. The first was previously folded into the acoustic factor, which describes a different physical phenomenon; the second was missing entirely despite being, for most buyers, the most immediately relevant fact about an approved-but-unbuilt site.
1. Air pollution (42% weight)
PM2.5, NOx, and ozone precursors from backup diesel generators and on-site gas turbines. The dominant health concern at most addresses, and the one with the clearest household-level evidence — which is why it carries the largest weight in the composite.
Roughly 0.55 backup generators per MW of IT capacity. Standby capacity runs to about 1.44 MW per MW of IT load once you account for a PUE near 1.25 and N+1 redundancy; at the 2.5 MW units common in hyperscale builds that is a little over half a generator per IT megawatt. Each unit runs 50–150 hours a year in testing, and a 2 MW unit at full load emits roughly 45 kg of NOx and 1.3 kg of particulates per hour.
air_exposure = capacity_scalar × plume_decay(distance) × 10
plume_decay(d) = 1 / (1 + (d / 0.75)^1.5)
0.5 mi → 0.65 2 mi → 0.19 10 mi → 0.02
1.0 mi → 0.39 5 mi → 0.06 20 mi → 0.01
Ground-level concentration downwind of a low stack peaks a few hundred metres out and then falls off faster than 1/d. The kernel above reproduces that shape. An earlier version of this model used a gentler 1/(1 + d/2) curve, which implied an address eight miles away still sat at a fifth of the fence-line concentration — far too high. Sources: Buonocore et al. (Harvard EmPower); Sarnat et al., near-source impacts of diesel backup generators in urban environments (Atmospheric Environment, 2015).
2. Water stress (13% weight)
Cooling water demand against local supply. Roughly 2,350 gallons per MW per day. That figure comes from LBNL's 2024 US Data Center Energy Usage Report: 17.4 billion gallons of direct onsite consumption in 2023, a fleet water-use effectiveness near 0.37 L/kWh, which works out to about 2,350 gallons for each megawatt-day of IT load. Evaporatively cooled hyperscale sites run closer to 5,300; air-cooled and closed-loop designs use a small fraction of it.
Arid regions — Utah, West Texas, the Southern California interior — carry a 1.5× multiplier reflecting limited aquifer recharge. The humid eastern Texas metros are excluded from that multiplier. Water stress is scored without a distance term, because watershed drawdown is a regional effect rather than something that fades over a few miles. For Amarillo and the HyperGrid proposal this category dominates, given Ogallala Aquifer fragility.
3. Acoustic exposure (24% weight)
Cooling fan walls (constant), generator testing (periodic), transformer hum (continuous, low-frequency). This is now modelled in decibels rather than as an arbitrary curve, because decibels are what determines whether you can sleep.
L(d) = L_ref − 20·log₁₀(d / 300 ft) − 1.2·d [dBA]
L_ref = 42 + 10·log₁₀(MW) continuous, at the property line
L_ref + 16 during generator testing
excess = L(d) − 38 dBA night-time ambient
score = max(excess_continuous, 0.6 × excess_testing) ÷ 3
The first term is spherical geometric spreading — 6 dB per doubling of distance — and the second is ground and atmospheric absorption. Sound power rises about 3 dB per doubling of capacity. Generator testing is roughly 16 dB louder than steady-state operation but runs only 50–150 hours a year per unit, so it is weighted to 60% rather than treated as continuous; it is still frequently the loudest thing a neighbour actually notices.
What the score measures is the excess over a quiet night, not the absolute level, since that is what the sleep-disturbance literature is built around. Low-frequency hum is the persistent complaint in Loudoun County and it is precisely what standard A-weighted ordinances are worst at capturing. Sources: WHO Environmental Noise Guidelines (2018); Piedmont Environmental Council documentation of resident complaints.
4. Electromagnetic / grid infrastructure (5% weight)
Substations, transmission lines, and transformer EMF. WHO's IARC classifies extremely-low-frequency magnetic fields as Group 2B (possibly carcinogenic), but at residential distances levels are typically well below precautionary thresholds. We weight this category low — 5% — because the science is genuinely uncertain compared to air and water. We include it because the visible HV infrastructure does affect property values and viewsheds even when health effects are unproven.
5. Heat island (16% weight)
Waste heat is reckoned against the whole facility draw, not just the IT load: at a PUE of about 1.25, a 1,000 MW campus rejects roughly 1,250 MW of heat. This category previously carried 25% of the composite, which overweighted it relative to the epidemiology — heat island is real and well documented, but the household health burden near these facilities is dominated by air and noise. It now carries 16%.
Each MW of IT load rejects roughly 1 MW of waste heat to ambient. At hyperscale density, this creates measurable local temperature elevation. Rob Davies (Utah State) modeled Stratos as producing up to 5°F daytime and 28°F nighttime temperature increases within the facility's immediate footprint — unprecedented at industrial scale. For Northern Virginia clusters the effect is smaller but documented; for HyperGrid (if built at full scale) it would be the largest single industrial heat source in North America.
6. Low-frequency vibration (10% weight)
Separated from the acoustic factor because it is a different phenomenon with different physics. Low-frequency energy — the bottom octave bands, where large slow fans, chillers and generators put most of their output — suffers far less atmospheric and barrier attenuation than the mid frequencies an A-weighted meter is tuned to. We model it at 0.35 dB per mile against 1.2 for A-weighted sound, with a reference level 4 dB above the continuous broadband figure.
This is the gap that produces the most persistent complaints. Facilities measured at 40–59 dBA on residential property — comfortably inside their permits — still generate reports of sleep disruption, difficulty concentrating, and what one resident described as "an internal organ vibration." Standard noise ordinances are written around A-weighting and are close to blind to it. A household is likelier to describe this exposure as felt than heard.
7. Construction phase (9% weight)
Time-limited, and absent from the operational factors entirely, but often the most decision-relevant thing on the page: an approved site next door means haul-truck routes, dust, pile driving and road damage for years before any of the operational effects begin.
construction = capacity_scalar(active_MW + 0.4 × planned_MW)
× 1/(1 + (d/1.5)²) × 10
months ≈ 18 + 9·log₁₀(MW), bounded to 18–60
peak workers ≈ 10 per MW
A 100 MW campus runs 18–36 months with 800–1,200 workers at peak. Capacity that is merely planned — not approved, not started — is discounted to 40%, because a meaningful share of announced projects never break ground. The distance kernel is tight: haul routes and site noise are a local phenomenon in a way that grid effects are not.
8. Night lighting (4% weight)
Security and parking illumination runs continuously, frequently in places that were previously dark. Direct illumination falls off quickly; sky glow carries further. Weighted low because the health evidence at residential distances is modest, but included because it materially changes what a rural property is like after dark — and because it is one of the effects residents raise most often that no risk tool reports.
Composite
composite_health = (air × 0.34) + (noise × 0.18) + (vibration × 0.10)
+ (heat × 0.12) + (water × 0.10) + (construction × 0.09)
+ (light × 0.04) + (emf × 0.03)
The weights still favour the exposures with direct, documented household-level pathways. Air leads because it carries the clearest evidence. Noise and vibration together account for 28% — more than noise alone did before — which reflects that they are the two things residents near these facilities actually report. Heat island and water stress are real but act more slowly and more diffusely. Construction is weighted for how sharply it lands while it lasts, discounted for the fact that it ends. Electromagnetic exposure stays lowest because the evidence at residential distances remains genuinely weak.
Adding these three shifted every score. That is the intended consequence of a more complete model, not a calibration drift: an address next to an approved-but-unbuilt campus now scores differently from an identical address next to a finished one, which is a distinction the previous version could not make.
The composite is reported as a separate metric in the score detail panel — it's not the same as the Location Score, which also includes jurisdictional and tax-base factors. The composite is purely the lived-health summary.
09 · Limitations
What this is and isn't
Datatox is a research instrument, not a financial or legal product. Every number here is a calibrated Fermi estimate based on the best publicly available data — but data center economics is fast-moving and locally specific. Key limitations:
City-level resolution masks within-city variation. A house 0.3 miles from a facility cluster and one 1.4 miles away in the same city face materially different conditions. The current tool treats them as equivalent.
Capacity estimates are based on published major-project figures plus tier-typical defaults — they are not parcel-resolution facility-by-facility audits. Real per-address exposure can differ from these estimates, particularly in mixed-tier areas.
Tax-base assumptions rest on county-level totals and median home values. The actual per-household share depends on assessment ratios and exemptions specific to your parcel.
Healthcare cost estimates and the 5-factor health analysis are population-level extrapolations from epidemiological studies. They do not predict individual outcomes.
Projections assume current siting and policy trends continue. Major regulatory changes — moratoria, federal preemption, or aggressive setback rules — would materially alter the outcomes.
Effects we know about and do not score
Some real effects are left out because they cannot be estimated from an address without site-specific data. Naming them is more useful than modelling them badly:
Stormwater and flood risk. Tens of acres of new impervious surface changes runoff. Whether that reaches a given property depends on local drainage and topography that no national model can see.
Cooling-tower discharge. Blowdown carries biocides, anti-scalants and high dissolved solids into municipal sewers or surface water. Permitted and monitored, but locally variable.
Fuel and battery storage. Large diesel inventories and grid-scale battery installations carry fire and groundwater-contamination risk, and place specialised demands on volunteer fire departments in rural host counties.
Viewshed. Windowless structures of 30–60 feet, security fencing and new transmission towers affect outlook and, indirectly, value. Partly captured in the property-value line; not separately scored.
Grid reliability. Rapid load growth raises questions about local reliability that are real but not resolvable per-address.
Farmland conversion. A community-level effect rather than a household one, and outside what this tool measures.
10 · Every number
Every constant in the model
A research instrument should not contain a number you cannot trace. This is every constant the model uses, what it is, and where it comes from. Where a figure is an estimate rather than a measurement, it says so.
Exposure model
| Constant | Value | Basis |
|---|
| Air plume kernel | 1/(1+(d/0.75)^1.5) | Near-source diesel dispersion shape (Sarnat et al. 2015) |
| Backup generators per MW | 0.55 | PUE 1.25 × N+1 redundancy ÷ 2.5 MW units — derived |
| Generator run hours | 50–150 /yr | Typical monthly testing schedules |
| Water intensity | 2,350 gal/MW-day | LBNL 2024: 0.37 L/kWh fleet WUE |
| Arid multiplier | 1.5× | Judgement — limited aquifer recharge |
| Noise reference | 42 + 10·log₁₀(MW) dBA at 300 ft | Calibrated to reported Loudoun property-line levels |
| Noise absorption | 1.2 dB/mile | Ground + atmospheric, A-weighted |
| Night ambient | 38 dBA | Quiet suburban/rural night |
| Generator test premium | +16 dB, weighted 0.6 | Intermittent — 50–150 h/yr |
| Low-frequency absorption | 0.35 dB/mile | LF attenuates far less than A-weighted |
| Facility PUE | 1.25 | Modern hyperscale typical |
| Construction duration | 18 + 9·log₁₀(MW) months | 18–36 months for 100 MW — fitted |
| Construction workforce | 10 per MW at peak | 800–1,200 workers per 100 MW campus |
| Planned-capacity discount | 0.4× | Judgement — many announced projects never start |
| Capacity scalar | log₁₀(MW)/4, capped at 1 | Impact is roughly logarithmic in capacity |
| Composite weights | .34 / .18 / .10 / .12 / .10 / .09 / .04 / .03 | Weighted to strength of household-level evidence |
Score
| Constant | Value | Basis |
|---|
| Base score | 15 | Floor — grid-level effects reach everywhere |
| Proximity component | 0–40 | 40% structural, 60% capacity-weighted |
| Cluster density | 0–25 | 1.2 per facility within 2 mi, capacity-weighted |
| Regional saturation | 0–20 | Total capacity within 15 mi |
| Jurisdictional penalty | 0–12 | Costs borne without the local tax benefit |
| Regional risk | 0–20 | Per-tier constant — water, air, grid pressure |
| Tax base offset | 0 to −12 | Proportional to host-county fiscal benefit |
| Soft ceiling | knee 85, scale 25 | Preserves ranking above 85 instead of clamping |
| Personal risk adders | asthma 10, cardio 7, pregnancy 6, elderly 5, noise-sensitive 5, children 4 | Ordered by strength of the PM2.5 morbidity literature — magnitudes are judgement |
| Personal exposure scalar | clamp(0.15, 1, score/75) | Health risk concentrates near sources |
Money
| Constant | Value | Basis |
|---|
| Loudoun tax saving | 0.732% of home value | NVTC / Mangum 2026: rate would rise $0.805 → $1.537 per $100 |
| School benefit | $4,200 per public-school child | Host-county per-pupil spending over state median |
| County service benefit | $800 (Loudoun), scaled elsewhere | Estimate — in-kind value of service quality |
| Electricity, PJM | $372/household/yr | Monitoring Analytics: $9.3bn of 2025/26 capacity cost ÷ ~25M households |
| Electricity, other grids | $60–300/yr | Scaled by data-center penetration and market structure |
| Water & sewer | up to ~$90/yr | Estimate — rate-base shifting; least precise line here |
| Health adders | asthma $950, cardio $700, pregnancy $600, 65+ $550, $60/member | Harvard EmPower morbidity costs, calibrated for host_dense |
| Health exposure factor | clamp(0.2, 1, composite/6.5) | 6.5 is the dense-cluster composite — the tier the adders were set for |
| Property value risk | tier rate × 0.25 × concentration | Discounted hard — see §04. Evidence is close to neutral |
| Jobs | 1.5 permanent per MW | 100–200 staff per 100 MW campus; valued at $0 to avoid double-counting |
| Projection growth | costs 0.5–6%/yr, electricity 4–6.5%/yr | PJM capacity trend and tier-specific build-out pace |
Which of these would move the answer most
If you only check three numbers, check these. The property-value line is the most contested and was recently cut by 75%; if the pessimistic reading turns out right it should go back up, and it would again dominate the ledger. The tax-saving rate is the best-evidenced figure here and the largest benefit, so it carries the positive side almost single-handedly in host counties. The health adders are extrapolations from population studies to households, which is the weakest inferential step anywhere in the model.
11 · Sources
Primary research and data inputs
- Facility data Named project specifications and coordinates from Cleanview's US Data Center Tracker and public filings. National counts are drawn from independent census trackers rather than any single list: roughly 4,800 tracked facilities as of July 2026 (3,687 operating, 229 under construction, 857 planned), against 4,423 reported by Statista in April 2026. The spread is definitional — building versus campus, and where enterprise server rooms stop counting.
- Exposure Pew Research Center (April 2026) — 38% of Americans live within five miles of an operational data center, rising to 42% including planned sites; 67% of planned facilities are headed to rural areas. Environmental Data & Governance Initiative — roughly 4 million people live within one mile of an EPA-regulated data center, disproportionately communities of color.
- Capacity Goldman Sachs — US data center power demand rising from 31 GW in 2025 to 41 GW in 2026. BloombergNEF — 118 GW of installed US capacity forecast by 2030, 194 GW by 2035. Announced project pipelines run several times higher than either figure; most announced capacity is never built.
- Health Harvard EmPower Analytics — "Health impacts of data center backup generation in Northern Virginia" (2025).
- Fiscal Northern Virginia Technology Council — "Data Centers and Northern Virginia's Tax Base" (2026 update).
- Property JLARC — Virginia data center development impact assessment (2024).
- Climate Frontiers in Climate — modeling of the proposed Stratos project's heat island and watershed effects (2026).
- Grid PJM Interconnection — Capacity Performance auction results, 2024–2026.
- Water Utah Division of Water Resources — Great Salt Lake watershed monitoring data.
- Permits Loudoun County, Prince William County, Fairfax County planning department filings, 2020–2026.
- Geocoding US Census Bureau 2023 Gazetteer — place file (32,126 places), ZCTA file (33,791 ZIP areas) and the 2020 ZCTA-to-county relationship file, embedded in the page and matched locally. Street-level lookup uses Photon (OpenStreetMap data, ODbL) because nationwide street data runs to gigabytes and cannot ship with a page; it is the only request this site makes, and only when you type a house number.
- Site map OpenStreetMap contributors (ODbL) — 912 additional mapped data centre sites, location only. Shown for awareness; excluded from all scored quantities because OpenStreetMap does not record capacity.
- Water Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report. Direct onsite water use of 17.4 billion gallons in 2023, a fleet water-use effectiveness near 0.37 L/kWh (~2,350 gal per MW-day).
- Dispersion Sarnat et al. — near-source impacts of diesel backup generators in urban environments (Atmospheric Environment, 2015), used to shape the air-quality distance kernel.
- Noise WHO Environmental Noise Guidelines for the European Region (2018) — night-time sleep-disturbance thresholds used as the acoustic reference.
- Electricity Monitoring Analytics (PJM's independent market monitor) — data centers responsible for 63% of the 2025/26 capacity price increase, $9.3 billion recovered from customers in one year. NRDC — projected ~$70/month per PJM household by 2028. PJM capacity prices rose from $28.92 to $329.17 per MW-day between 2024/25 and 2026/27.
- Property George Mason University Center for Regional Analysis (2025) — homes nearer data centers sold for higher prices in Northern Virginia. Realtor.com (2026) — 42% appreciation near facilities vs 41% surrounding. University of Rochester — little measurable effect. A separate GMU-led analysis found new data centers slowed local price growth. These findings are why the property-value line was cut by 75%.
- Fiscal Mangum Economics for the Northern Virginia Technology Council (2026) — Loudoun residential rate would rise from $0.805 to $1.537 per $100 without data center revenue; data centers were 42% of local tax funding in 2026.
Address coverage is national: 33,791 ZIP areas and 32,104 places across all 50 states, DC, Puerto Rico, the US Virgin Islands, Guam, American Samoa and the Northern Mariana Islands. Datatox scores against ~145 facilities with published capacity, weighted toward the large and named projects, out of roughly 4,800 tracked nationally. A further 912 sites are drawn from OpenStreetMap for awareness only. This covers the lion's share of capacity (the largest hyperscale and named projects), but smaller colocation buildings under ~30 MW are not individually mapped. Address-level analysis uses real distance to all 145 mapped sites; for addresses far from any mapped site, the tool falls back to city-level defaults.