Google paid roughly seventy-eight million dollars in property taxes to Caldwell County, North Carolina, and received about seventy-three million of it back under a rebate agreement. The county kept around five million.
That is not a scandal and it is not a secret. It is a negotiated economic development agreement, approved in public session, of a type that hundreds of jurisdictions have signed. What it illustrates is that the question of who pays for a data center has an answer, that the answer is written down in specific instruments, and that almost nobody reads them before forming a view.
Data center cost allocation is not one question. It is five, and they resolve through completely different mechanisms operating on completely different timescales. There is an electricity bill, a tax bill, a water bill, an infrastructure bill, and a bill that only arrives if the facility leaves early. Each has its own payer, its own paper, and its own failure mode, and conflating them is why the public argument produces so much heat and so little resolution.
Data center cost allocation bill one: electricity
The mechanism by which a data center raises somebody else’s electricity bill is indirect, which is why it took so long to become political.
A large new load tightens regional supply. Capacity auctions clear higher. Every customer in the zone pays the higher clearing price. No transfer from a residential customer to a data center appears anywhere in the accounting, because none occurs. What occurs is a price effect distributed across a market.
That indirectness is why data center cost allocation resisted regulation for as long as it did. A rate case is designed to allocate identifiable costs to identifiable customers, and a capacity price effect is neither. It took auctions clearing visibly short, and the resulting increases appearing on bills in the same news cycle as a project announcement, before commissions treated it as a rate design problem rather than a market outcome.
The regulatory response has been the fastest-moving development in American utility rate design in decades. As of mid-2026, roughly twenty-three states had approved at least one large load tariff, a dedicated rate class for very large customers designed to assign the costs of serving them directly rather than spreading those costs across the general rate base, with several more proposals pending.
The Database of Emerging Large-Load Tariffs, assembled by the Smart Electric Power Alliance and the North Carolina Clean Energy Technology Center, catalogs the resulting architecture, and a recognizable archetype has emerged from the filings. In the March 2026 public update, thirty-three of seventy-seven filings carried numeric minimum-bill requirements, averaging around eighty percent of contracted capacity, and thirty-seven included collateral requirements. That is a rate design archetype assembling itself in real time across dozens of jurisdictions, which is unusual, since utility rate structures normally change on a generational timescale and the regulatory apparatus was built for incremental load growth.
The analysis of what that archetype looks like across filings identifies the components clearly enough to list. Minimum billing, fixing monthly payment at a percentage of contracted load regardless of consumption. Extended contract terms aligned with the life of the infrastructure being built. Collateral, in letters of credit or cash. Exit fees. And provisions governing contract modification and capacity reassignment.
Virginia’s GS-5 rate class is the reference implementation, approved by the State Corporation Commission to take effect in January 2027. It applies to loads at or above twenty-five megawatts, requires fourteen-year contracts, and takes-or-pays a minimum of eighty-five percent of contracted transmission and distribution capacity and sixty percent of generation demand regardless of actual usage, with collateral reported at one and a half million dollars per megawatt. For a five-hundred-megawatt campus that is seven hundred and fifty million dollars posted before the first rack ships, which changes the capital structure of the project and pushes the financing question back onto the operator’s balance sheet.
What the tariff fight is actually about
The design parameters look technical and each one is a distributional decision, which is why proceedings that used to be uncontested now draw intervenors.
Minimum demand percentage is the central battleground. Oregon’s proceeding is representative: staff and a coalition supported a ninety percent minimum, the utility proposed eighty percent arguing that peer utilities sit near there and that setting it too high could deter siting, and the data center coalition agreed with the utility while noting that no other customer class faces such a requirement at all.
Every position in that dispute is defensible. A higher minimum shifts more risk onto the customer and protects other ratepayers. A lower minimum keeps the jurisdiction competitive and leaves more stranded asset exposure with the utility and therefore with everybody else. A higher minimum also has a second-order effect worth noting: it encourages customers to contract for less capacity than they might need, which improves the utility’s risk position and worsens its planning information, since the forecast it builds against becomes systematically conservative. And the observation that no other rate class faces a take-or-pay obligation is factually correct and is the strongest argument the industry has, though the response is that no other rate class arrives at five hundred megawatts requiring generation that would not otherwise be built.
The eligibility threshold is its own quiet fight. Utah legislated large loads at one hundred megawatts and above. Virginia’s GS-5 applies at twenty-five. Minnesota directed its commission to set a threshold. Where the line falls determines which facilities are covered and creates an obvious incentive to design just underneath it or to split a campus into separately metered parcels, which is the standard behaviour around any regulatory threshold in an industrial context.
Contract duration works the same way. Oregon staff recommended fifteen-year minimums for loads above twenty megawatts on the reasoning that stranded asset risk scales with load size. Longer contracts align customer commitment with the depreciation schedule of the assets built to serve them, which is precisely the point, and they also require a company to commit for three times the useful life of the hardware inside its building.
Pennsylvania’s Public Utility Commission adopted a model framework in April 2026 establishing that interconnection upgrade costs are recovered directly from large load customers rather than from the general rate base, with deposits and collateral sufficient to cover upgrade costs so that projects which do not proceed do not strand those costs on other customers.
Capacity reassignment provisions are the underrated innovation. Some tariffs permit a customer to reduce or reassign a portion of contracted capacity without penalty, require the utility to attempt reassignment beyond that threshold, and reduce or waive exit fees when the departing customer supplies a successor. That converts a binary default into a transferable position, which is a genuinely better instrument than either a hard lock-in or a free exit. One utility’s schedules permit reassigning or reducing up to twenty percent of contracted capacity without penalty under notice conditions, require the utility to attempt reassignment beyond that, and impose exit fees otherwise.
Bill two: taxes, and the instruments that hide the number
Property tax treatment is where the largest sums move and where the accounting is most opaque, and the opacity is structural rather than conspiratorial.
The straightforward instrument is an abatement, reducing the tax bill by some percentage for a set term. Arkansas law permits up to sixty-five percent abatement for as long as thirty years on projects financed through industrial development revenue bonds, and PILOT agreements reached for Google projects in the state carry the maximum sixty-five percent for thirty years on both real and personal property.
A payment in lieu of taxes agreement works differently and is easy to misread. The local government takes title to the property, which makes it public and therefore exempt, and leases it back to the company, which pays a negotiated fee instead of taxes. In one Ohio case a facility received a fifteen-year, seventy-five percent property tax abatement alongside a PILOT of five hundred thousand dollars annually.
Industrial revenue bonds are the third structure and produce the largest numbers. Dona Ana County, New Mexico approved approximately one hundred sixty-five billion dollars in industrial revenue bonds for a data center project, with associated abatements described as undisclosed and running as long as thirty years.
The aggregate figures where they exist are substantial. Virginia’s data center sales tax exemption reached roughly $1.6 billion annually. Georgia localities were estimated to lose $1.1 billion in 2026 and $1.4 billion in 2027 from state-awarded exemptions. Data centers owned by four large operators in Oregon received $616 million in property tax abatements between 2016 and 2025, with annual program costs rising several hundred percent across that period.
The measurement problem is the part that should trouble everybody regardless of position. Arkansas does not track the impact of PILOT agreements on property tax collections at the state level. In several states the cost estimates surfaced only because a legislator requested them or a records request produced them. Accounting standards require governments using generally accepted principles to disclose tax abatements in their financial reports, and compliance is uneven.
The case for the abatements, stated properly
The critique is easier to write than the defense, so the defense deserves its strongest form.
The counterfactual argument is the real one. If a facility would not have located in the jurisdiction without the abatement, then the abated revenue was never available to lose, and whatever the county collects, plus construction employment, plus utility revenue, plus any assessed value that does eventually appear, is a gain against a baseline of nothing. Virginia’s original exemption in 2008 carried a fiscal note estimating a forgone $2.8 million against a project the state was otherwise going to lose to North Carolina.
The service-demand argument is also legitimate. A data center generates minimal traffic, few emergency calls, no students, and no demand on the largest line items in a county budget. Bartow County, Georgia’s policy statement makes exactly this case: substantial revenue against comparatively little service demand, with an explicit intent to increase homestead exemptions as the revenue arrives. Whether that intent survives a change of commissioner is a different question, and the history of resource-revenue windfalls being absorbed rather than distributed is not encouraging on that point.
The depreciation point is the technical one that both sides underuse. Where equipment is assessed at a percentage of fair market value with statutory depreciation applied, the tax digest impact is a function of depreciated value rather than gross capital cost, which means the headline investment figure and the eventual assessment are very different numbers, and the assessment declines every year as the servers age. The equipment inside is also replaced on a cycle shorter than most abatement terms, which means the digest is a function of reinvestment as much as of the original build. That depreciation curve is also why the useful-life assumption fight in the financing layer has a fiscal consequence nobody discusses: a shorter economic life for the equipment means a faster-declining tax digest for the county that hosts it.
The honest counter is the timing. During an abatement period running ten to thirty years, the community provides road maintenance, emergency response capacity, and utility infrastructure to a facility that is not yet contributing proportionally, and the service demands arrive before the revenue does, which is the same sequencing problem that afflicts any large industrial facility with a long ramp. Whether that sequencing is a subsidy or an investment depends on what happens at the end of the term, which nobody in the room when it is signed will still be in office to see.
Bill three: water, and the rate structure underneath it
Water is the smallest of the five bills in dollar terms and frequently the largest politically, and the reason is that the rate structure makes the cost invisible.
Municipal water systems are largely fixed-cost businesses. Treatment plants, mains, and pumping stations cost what they cost regardless of throughput. A large new customer improves the utilization of that fixed base, which is genuinely good for the system’s economics and can lower unit costs for everybody on the network, and a volumetric rate that reflects average cost therefore undercharges relative to the capacity the customer requires at peak.
Where the withdrawal is from groundwater, the cost is not on any bill at all, because a permit to withdraw is not a purchase. An aquifer drawn down faster than it recharges is a stock being depleted, and the depletion shows up as a cost to future users rather than to the current one. Prior appropriation states add a market layer, since a new entrant holds a junior right and can only obtain seniority by purchasing it, generally from an agricultural holder, which converts a public resource question into a private transaction the county has no formal role in. The same structure governs mineral and extraction rights and produces the same local ambivalence.
Reclaimed water is the mitigation with the clearest economics, since it uses a resource that had no competing use and improves the utilization of treatment infrastructure. It also requires the distribution network to exist, which is a capital project somebody has to fund, and the funding question is a rate case. Purple pipe extensions are expensive per mile and only pencil where a large anchor customer justifies them, which means the data center is frequently the reason the reclaimed system exists at all, and that is a genuine public benefit that the water accounting rarely credits.
Bill four: infrastructure, and who owns what afterward
Substations, transmission upgrades, road improvements, and water main extensions are capital projects with thirty-to-forty-year cost recovery, built for a customer whose relationship may be much shorter.
The Pennsylvania framework’s answer, now increasingly standard, is that interconnection upgrade costs are recovered directly from the large load customer with deposits and collateral sufficient to cover them. That resolves the funding question and leaves the ownership question, because the utility owns the substation afterward regardless of who paid, and the asset either serves a successor or does not. Where it does, the community has acquired grid capacity it did not pay for, which is the best case and is real. Where it does not, the community has acquired a substation serving nothing, which is the stranded infrastructure pattern that has emptied out industrial towns before.
California’s proceeding surfaced the refund issue that follows. Where a customer funds infrastructure and the utility later serves others from it, some portion is conventionally refunded. The California commission provisionally found a maximum refund of seventy-five percent of total capital expenditure appropriate, reasoning that large load customers present unique stranded cost risks, already receive favourable energy rates, and require investment at a scale that justifies limiting refunds.
Underneath all of it sits the equipment lead time problem, which changes the economics of any upgrade because a transformer ordered for one customer and delivered four years later may arrive for a project that no longer exists.
Bill five: the stranded cost, and the one nobody has paid yet
This is the bill that has not arrived, and it is the reason every tariff proceeding in the country now contains the phrase stranded asset risk.
The structure is straightforward and uncomfortable. A utility builds generation, transmission, and substation capacity for a data center, financed over decades and recovered through rates. The customer signs for five, ten, or fourteen years. The silicon inside the building has an economic life of four to six. If the load leaves, reduces, or never fully materializes after the assets are built, the assets remain and somebody pays for them, and that somebody is the remaining ratepayers.
The protections regulators are assembling against exactly this scenario are extensive and recent, and the historical precedent is why they are treating this seriously rather than theoretically. Utility commissions have been through stranded cost episodes before, in the aftermath of nuclear construction programs and again during restructuring, and the resolutions were expensive and politically brutal. The institutional memory is real, and it is why commission staff in these proceedings are frequently more conservative than either the utility or the intervenors.
The protections now standard in tariffs are all addressed to this single risk. Minimum billing creates a revenue floor independent of consumption. Extended contract terms align commitment with asset life. Collateral covers unpaid obligations. Exit fees penalize early departure. Capacity reassignment creates an alternative to default. Each is an attempt to make a customer whose planning horizon is short behave like one whose horizon matches the infrastructure.
Whether they are sufficient is unknown, because none has been tested by an actual departure at scale. A fourteen-year contract with collateral at one and a half million dollars per megawatt looks robust against a five-hundred-megawatt facility walking away, and considerably less robust against a general repricing in which many facilities reduce simultaneously and the utility cannot reassign capacity because nobody wants it.
That is the tail risk in data center cost allocation, and it is correlated rather than idiosyncratic, which is exactly the property that makes collateral requirements less protective than they appear.
The verification problem
A structural difficulty runs underneath all five bills and deserves naming, because it limits what any of this analysis can establish.
Cost shifting is extremely difficult to verify. Utility cost allocation methodologies are complex, contested, and jurisdiction-specific. Whether a particular tariff fully assigns incremental costs depends on assumptions about how those costs are measured, which are precisely the assumptions being litigated. Harvard researchers examining the question have noted that in many markets verification is close to impossible with publicly available information.
The same applies to fiscal impact. Studies commissioned by industry find substantial net benefits, with one national assessment putting the sector’s contribution above two trillion dollars. Studies commissioned by opponents find substantial net costs. Both are typically methodologically defensible, because the result depends on the counterfactual assumed, and the counterfactual is unobservable.
Non-disclosure agreements compound it. In several documented cases, commissioners approved abatement agreements while under NDA and without accompanying economic impact assessments or cost-benefit analyses. The same confidentiality practice governs water and power figures, which means a single project can present three separate unverifiable numbers to the same board. A decision made on information the decision-maker could not share and the public could not review is not necessarily a bad decision, and it is one nobody can audit. Legislative efforts to prohibit officials from signing such agreements have been introduced in several states and have generally stalled against the argument that confidentiality is required to compete for projects, which is an argument with real force and no way to test it.
Which means the honest position on most specific claims in this area is that the number is contested and the methodology is where the argument actually lives. Anyone presenting a clean figure for what a data center costs or contributes is presenting a modeled result and usually not the assumptions. That is the same discipline the critical minerals literature requires of any demand forecast, and for the same reason: the number is downstream of a model, and the model is downstream of an interest.
What a resident can actually check
The information asymmetry is real and it is narrower than it looks, because most of these instruments generate a public record even when the negotiation did not.
The abatement agreement itself is usually a matter of record. If a deal required a vote by a city council, county board, or industrial development authority, that vote appears in minutes, agenda packets, and resolutions, typically published or available by request, and those documents frequently contain the executed terms.
Where the incentive was created by statute rather than negotiated case by case, the legislation and its fiscal note are public, and the agency administering it usually reports aggregate usage to the legislature.
State open records laws reach the rest. Every state has one, and the executed incentive agreement between an economic development agency and an operator is generally a public record even where the negotiation was confidential.
Utility filings are the most useful and least used source, and the technical constraints they document frequently explain project delays that get attributed to politics. A tariff proceeding is a public docket containing the utility’s cost justification, the intervenor testimony disputing it, and the commission’s reasoning, which together constitute a far more rigorous examination of the cost allocation question than any news coverage of it.
And accounting standards require governments using generally accepted principles to disclose forgone revenue from tax abatements in their annual financial reports. Compliance varies, and where it exists the number is in the notes.
None of that resolves the counterfactual problem. It does mean that the specific terms of a specific deal are usually knowable, and that most public argument about data center cost allocation proceeds without anybody having read them.
The claims that do not hold up
An audit, because the confident assertions run in both directions.
Data center cost allocation is a solved problem in states with large load tariffs overstates instruments that mostly take effect in 2027 and have never been tested by a departure.
Data centers pay nothing in taxes is false. Abatements are partial and time-limited in most jurisdictions, sales tax exemptions typically cover equipment rather than everything, and utility taxes and payroll taxes are unaffected.
Data centers pay their own way is equally unsupported as a general claim, since it depends entirely on the specific instrument, the abatement term, and the tariff in force, all of which vary enormously by jurisdiction.
Ratepayers are subsidizing data centers is the strongest form of a claim that is genuinely hard to verify, and the direction is plausible while the magnitude is contested.
Large load tariffs solve the problem overstates instruments that are two years old, mostly untested, and varying widely in how much risk they actually transfer.
Tax abatements are always a giveaway ignores the counterfactual question, which is the only question that matters and the one nobody can answer.
The jobs justify the incentives is difficult to sustain at the ratios involved, and most serious economic development arguments have shifted to the tax base and utility revenue rather than employment. Construction employment is genuinely substantial and genuinely temporary, and the specialized trades involved are in national shortage, which means a project frequently imports its workforce rather than hiring locally.
Data centers do not use public services understates road wear during construction, the heavy-haul permits required to move transformers and turbines, emergency response capability that must be maintained for a high-value facility, and the water and power infrastructure that is public in most jurisdictions.
Communities can just say no is true and incomplete, since state-level exemptions frequently bind localities that had no vote on them, which is the specific grievance driving several state legislative fights. A county that never voted on a state sales tax exemption still absorbs the service demand of the facility the exemption attracted, and that mismatch between who granted the incentive and who bears the cost is the structural complaint underneath a great deal of otherwise inarticulate local anger.
What data center cost allocation is actually telling us
Assemble the five bills and the pattern is that each one is a mechanism for deciding who bears a risk, and the risks are what actually differ.
The electricity bill allocates the risk that supply tightens. The tax bill allocates the risk that a facility does not deliver the promised base. The water bill allocates the risk that a resource depletes. The infrastructure bill allocates the risk that an asset built for one customer serves nobody. And the stranded cost bill allocates the risk that all of the above happen at once because the demand did not hold.
That last correlation is the thing most of the instruments handle poorly. Collateral, exit fees, and minimum billing all protect against an individual customer failing. None of them protects against a sector-wide repricing in which many customers reduce simultaneously, capacity cannot be reassigned because demand has fallen everywhere, and the utility holds assets built for a load that no longer exists. Financial instruments designed for idiosyncratic risk perform badly against correlated risk, which is precisely the failure mode the asset-backed lending against depreciating hardware exhibits one layer up the capital stack, which is a lesson that gets relearned expensively about once a decade, most recently in commodity markets where every producer hedged against their own idiosyncratic risk and none against the cycle.
Which suggests the useful question for anybody evaluating a specific project is not whether data centers pay their fair share, because that phrase does not designate anything checkable. It is: which instrument governs each of the five bills here, what does it assume, and what happens under it if the load reduces by half in year six.
The ten-lecture briefing on how AI data centers work runs the physics, the money, and the politics in sequence because the allocation question sits downstream of all of them. The thermal density set the load, the load required infrastructure, the infrastructure required financing over decades, and the financing has to be recovered from somebody across a period longer than anybody involved can forecast. Every instrument described here is an attempt to write down, in advance, who that somebody is under conditions nobody can specify.
A county kept five million dollars out of seventy-eight. Whether that was a good deal depends entirely on what the county would have collected from an empty field, and nobody in that room knew, and nobody knows now.
Leave a Reply