Data-center tax incentives: what they cost, what they bring, and who pays
Should a state keep, tighten, or end its data-center tax incentives? Bloom shows what each choice costs over 10 years, who pays (down to $/household), and how sure we are — a back-tested engine that informs the decision, not makes it.
Why a model, not a spreadsheet
In 2008 Virginia estimated its exemption would cost $1.54M a year. It became $1.9 billion — off by 125,000%. A single best-guess number fails badly here; Bloom gives a range that actually contained the real cost.
Virginia data-center sales-tax exemption: official 2008 estimate vs FY2025 actual. Georgia, Ohio, and Texas missed by 665–1,085%.
The decision
Keep it, tighten it, or attach flexible-power conditions — move any control and watch who gains and who pays update live.
Try a "what-if"
What each choice costs over 10 years
The fiscal cost (foregone revenue), built from the per-year base our back-test validates, compounded over 10 years under stated assumptions. Keep the incentive (do nothing) vs tighten it vs attach flexible-power conditions. The line is the middle estimate; the band is the likely range. Wider power/health costs are shown separately as an illustrative layer, not stacked here.
Who gains and who pays
Every dollar lands on someone. Green bars are groups that come out ahead; red bars are groups that bear the cost.
Who's affected
Who actually feels this — which groups pay, a fairness check for lower-income households, and the honest jobs picture.
Cost channels & opportunity cost
Fiscal is the validated headline; Power/Health/overlays are an illustrative wider-cost layer — not back-tested. Scales with the scenario. costbenefitcontext
Two-sided ledger — cost vs benefit
Industry's side, counted: the public cost set against the estimated economic benefits (jobs, construction, tax base). Benefits are illustrative, not back-tested like the cost.
Equity & environmental justice
How the electricity-bill increase lands across income groups (lowest fifth → highest fifth). We flag it as unfair when the lowest-income group is hit more than 1.25× the average.
Megadeals — the extreme case
Your current scenario sized against named real-world deals (worst-case stress tests).
| Deal | Incentive | Context |
|---|---|---|
| Amazon — Indiana | up to $8.0B | > Indiana's entire HHS budget ($5.7B) |
| Amazon — Oregon | $1.0B | = 11 years of the county budget |
| Big 4 hyperscalers | $360B capex (2025) | ~50% on-time · queues overstated 3–5× |
Rigor & trust
How we know the numbers hold up — ranges tested against what really happened, what drives them, every source, and how the tool is governed.
Validation — official estimate vs actual
How we tested it: predict each state's most recent year from earlier years only, then check the real cost against our range.
Sources — every number is checkable
Each figure in Bloom traces to a public source. Click to open the original.
What drives the cost most
The inputs that change the 10-year cost the most, biggest first — so you know which assumptions matter.
Governance & guardrails
Who's accountable for the tool, its non-partisan charter, and why you can't export a one-sided "cost-only" chart.
| Intended use | US state legislative / regulatory fiscal analysis |
| Out of scope (named) | siting · facility profitability · federal levers (FERC/ITC) · tribal/rural-specific impacts · non-US jurisdictions |
| Transferability | method ports to other jurisdictions via a country/currency/grid config (this build is the US instance) |
| Alignment | supports SDG-tracked public-finance transparency & subsidy accountability |
| Accountable owner | named PUC / budget economist (RACI) |
| Known bias | EPRI dependence; structural-break under-prediction |
| Review cycle | quarterly recalibration (EIA / PJM) |
"Grid-and-fiscal analysis under current law, not advocacy." Best / expected / worst shown with equal prominence; every number cites a source.
Ask the data
A built-in assistant answers in plain language from this run's numbers, cites its sources, and won't recommend a policy — that's your call.
Docked at the bottom of every tab — ask anytime, or tap a suggestion to begin.
Every reader, answered
The kind of reader, the question they bring, and where Bloom answers it.
| Reader | What they ask | Where to look |
|---|---|---|
| Fiscal analysts | defensible range + sourced assumptions? | Overview · Decision |
| PUC / advocates | survives cross-examination? | Decision · Who's affected |
| Legislators | what happens to my district? | Winner / loser |
| EJ / community | who actually pays? | Equity + health overlay |
| Grid planners | physically credible? | Flexibility · grid dynamics |
| Industry / econ-dev | are benefits counted? | Dual-sided + jobs reality |
| Academics | back-test + failure mode? | Rigor & trust |
| Responsible AI | accountable + non-partisan? | Governance + lock |
| Journalists | one stat, one chart? | Overview hero |
| Hackathon judges | reasoning + impact + RAI + demo? | all · Ask the data |
Open-the-file demo using a seeded model that mirrors the validated Python engine; when the live engine is connected the numbers come straight from it. The back-test is real (see Rigor & trust). Key external figures are sourced — re-confirm before publication.