The Thirsty Grid

Executive Summary

Global data centre electricity consumption reached approximately 415 TWh in 2024 — about 1.5% of world electricity use — and is on track to reach roughly 945 TWh by 2030 under the International Energy Agency's base case, a demand trajectory growing four times faster than electricity consumption from every other sector combined. The United States and China alone account for nearly 80% of that growth. This is not a story about chips getting smarter. It is a story about turbines, transformers, transmission lines and cooling towers — the oldest kind of industrial capital — being built faster than the grids and watersheds around them can absorb.

Two resources are the binding constraints: electricity and water. In Virginia's "Data Centre Alley," data centres already draw roughly 25% of the state's electricity, a share that has helped push year-over-year retail electricity prices in the Washington D.C. area up 23.7% in twelve months. In Bengaluru, India, roughly 30 data centres sit inside a city where 6,900 government borewells have already run dry. In Uruguay and Chile, proposed hyperscale campuses have been paused or contested by residents living through multi-year droughts. In the Netherlands, opposition to a Microsoft data centre escalated to an acid attack on construction equipment in July 2026. These are not hypothetical tail risks. They are the visible edge of a resource-allocation problem that is already underway.

This paper argues that the economics of the AI build-out cannot be separated from the economics of the grids and rivers that feed it, quantifies the scale of the electricity and water problem with the best available public data, examines six real geographies where the strain has already surfaced, reviews what has actually worked when governments and companies responded (Singapore's efficiency-linked capacity allocation, Cornell's finding that strategic siting alone can cut AI's water footprint by up to 86%, and the emerging wave of nuclear power-purchase agreements), and proposes an eight-pillar solutions framework and an original composite metric — the Grid-Water Stress Score — for comparing data centre siting decisions on both resource dimensions at once, not electricity alone.

How to Read This Paper

NGE's standard practice is to verify before publishing, name what could not be verified, and correct errors in public when found. This paper mixes three different kinds of claims, and they should be read differently.

Established, citable research and data: the IEA's electricity-demand figures (Part II), the Cornell/Fengqi You water-footprint modelling (Part IV and Part VII), the Bengaluru borewell and Cauvery-allocation figures (Part V), the Virginia electricity-share and ratepayer figures (Part III), and the nuclear power-purchase-agreement terms (Part VII) are drawn from named public sources — the IEA's Energy and AI and Electricity 2026 reports, peer-reviewed and university research, utility and regulatory filings, and contemporaneous reporting from named outlets — and are cited in the References section. Where a figure is a range or an estimate rather than a single measured number, this paper says so explicitly rather than presenting a midpoint as precise fact.

This paper's own framing and synthesis: the grouping of six geographies into "three grids" and "three basins," the characterisation of any one company's strategy as representative of an industry trend, and the narrative connecting separate data points into a single argument are this paper's own editorial construction, not a finding attributed to any cited source.

This paper's own proposal: the eight-pillar solutions framework in Part VII and the Grid-Water Stress Score (GWSS) introduced in Part X are original to this paper. Neither is an existing regulatory standard, an industry benchmark, or the output of a third-party model. The worked GWSS example is illustrative, built to demonstrate the mechanic, not a validated measurement of any real facility.

Part I — The New Industrial Base: Data Centres as Physical Infrastructure

Observation: artificial intelligence is routinely discussed as a software phenomenon — a matter of algorithms, parameters, and model weights. Evidence: yet every one of those algorithms runs on physical hardware that draws real electricity from a real grid and expels real heat that must be removed, in most cases, with real water. A single large frontier-model training run can consume electricity comparable to the annual usage of several thousand households, and inference — the ongoing cost of simply answering user queries after a model is deployed — now accounts for a majority of AI's total compute demand as adoption scales, because a model is trained once but queried billions of times. Insight: AI's growth curve is therefore not a software curve at all; it is an industrial capital-expenditure curve, resembling the build-out of railways, telegraph networks, or the electrification of manufacturing more than it resembles a normal software product cycle. Implication: the constraint on how much "intelligence" can be deployed is not, in the near term, computer science. It is how quickly turbines can be manufactured, transmission lines permitted, and water rights allocated — timelines measured in years, against a compute demand curve measured in months.

This reframing matters for how the risk should be assessed. A software bottleneck is solved by better code. An infrastructure bottleneck is solved by capital, permits, physical construction, and — critically — by decisions about who gets first claim on a finite regional supply of electricity and water when a hyperscale campus and a residential neighbourhood are drawing from the same grid and the same river at the same time.

Part II — The Power Equation: Quantifying Global Electricity Demand

2.1 The Base Numbers

According to the International Energy Agency's Energy and AI analysis, global data centre electricity consumption stood at approximately 415 TWh in 2024, roughly 1.5% of global electricity use, having grown at about 12% per year over the preceding five years. Under the IEA's base case, that figure is projected to reach around 945 TWh by 2030 — just under 3% of global electricity consumption — implying a forward growth rate of roughly 15% per year, more than four times the growth rate of electricity demand from every other sector of the economy combined.

Source: IEA, Energy and AI and Electricity 2026. Even the IEA's own sensitivity range spans roughly 700–1,700 TWh by the mid-2030s — a gap wider than the current electricity consumption of Japan — which is itself a data point: the honest state of forecasting in this sector is a wide band, not a precise number, because it depends on variables (chip efficiency gains, model architecture shifts, the pace of physical build-out) that have moved faster than forecasters in almost every year since 2022.

2.2 Regional Concentration

The United States and China together account for nearly 80% of global data centre electricity growth to 2030. Per-capita data centre electricity use in the United States is projected to rise from roughly 540 kWh (2024) to over 1,200 kWh (2030) — more than the total annual household electricity consumption of many lower-income countries, attributable to data centres alone, before counting anything else the grid serves.

Part III — Geography of Power: Three Grids Under Strain

3.1 Northern Virginia, United States — "Data Centre Alley"

Loudoun County and its surrounding jurisdictions in Northern Virginia host the largest concentration of data centres on Earth, with an estimated 70% of global internet traffic passing through the region at some point. Data centres now consume approximately 25% of Virginia's total electricity — nearly double the 15% share reported in neighbouring North Carolina, and a level of grid dependence on a single commercial customer class with few historical parallels outside heavy industry.

The trajectory of demand growth has surprised the region's own utility. In 2021, Dominion Energy projected annual electricity demand growth in its Virginia service territory of roughly 1%. By 2026, that projection had been revised upward to over 5% annually — a five-fold increase in the underlying growth assumption within five years, driven almost entirely by data centre interconnection requests. Dominion has since stated publicly that it cannot fully meet current data centre power demand on its existing timeline, and is undertaking an infrastructure expansion of roughly $8 billion, of which Virginia ratepayers — not the data centre operators driving the need — are projected to bear approximately 55% through rate base charges.

3.2 Ireland — The Grid That Said No, Then Yes With Conditions

Ireland's data centre sector grew large enough, relative to the size of its national grid, that the national transmission operator EirGrid and the energy regulator CRU effectively froze new large-load grid connections in the Dublin region for several years — a moratorium driven by the same underlying dynamic as Virginia's, at national rather than regional scale, because Ireland's total grid is small enough that a single sector's growth could threaten system-wide reliability. In late 2025, the CRU published its final policy ending the connections freeze, but only under new conditions: new data centres are now required, under the announced policy, to be capable of generating power back into the national grid, effectively converting the largest new loads into partial grid assets rather than pure consumers, and to demonstrate demand-flexibility capability during system stress events.

Ireland is the clearest existing precedent for a regulator using connection policy itself — not just pricing — as the tool for aligning data centre growth with grid capacity, and its 2025 policy reset is referenced directly in the solutions framework in Part VII.

3.3 Singapore — Efficiency as the Price of Entry

Singapore banned new data centre construction outright in 2019 once the sector reached roughly 7% of national electricity consumption on a small, land-constrained island grid with no cross-border transmission buffer of the kind larger countries can rely on. The moratorium lifted only gradually: a roughly 60 MW pilot in 2022, an 80 MW allocation round in 2023, and — in the current 2026 "DC-CFA2" call for applications — a 200 MW minimum tranche with a possible expansion via a dedicated 700 MW low-carbon data centre park on Jurong Island.

Access to that new capacity is not unconditional. Applicants must meet a Power Usage Effectiveness (PUE) requirement of 1.25 at full IT load — among the most stringent efficiency thresholds in the Asia-Pacific region — hold BCA-IMDA Green Mark Platinum certification, and source at least 50% of power from an approved list of green-energy pathways (biomethane, low-carbon hydrogen and ammonia, advanced solar, or fuel cells paired with carbon capture); generic renewable energy certificates purchased without direct procurement do not qualify. Singapore is discussed in detail as a "what worked" case in Part VII.

Part IV — The Hidden Water Footprint

4.1 Why Water, Not Just Electricity

Observation: public debate about AI's environmental cost has focused almost entirely on electricity and carbon. Evidence: yet the dominant cooling technology for large data centres — evaporative cooling — converts water directly into water vapour to remove server heat, meaning the water is genuinely consumed, not merely borrowed and returned, in a way electricity from a river-cooled power plant is not. U.S. data centre cooling systems alone withdrew an estimated 66 billion litres of water in 2023 (still under 1% of total U.S. water consumption at that scale), a figure research groups led by Cornell's Fengqi You project could rise to between 731 billion and 1,125 billion litres annually by 2030 once electricity-generation water use is included — the higher end of that range roughly equal to New York City's entire annual drinking-water supply. Insight: water and electricity are not two separate environmental costs of AI; they are coupled, because most electricity generation is itself water-intensive (thermoelectric power plants use water for cooling too), so a data centre's true water footprint includes both the water it uses directly for cooling and the water embedded in the electricity it draws. Implication: any solutions framework that addresses electricity sourcing without addressing water, or vice versa, is solving only half of the resource-allocation problem — which is why Part VII's framework and Part X's proposed metric treat both dimensions jointly rather than separately.

4.2 Per-Query Estimates — and Why They Keep Changing

Early per-query water estimates, including Shaolei Ren's widely cited 2023 University of California, Riverside research, put the cost of a short AI-generated exchange at roughly 500 millilitres of water per email-length draft using a GPT-4-class model under specific cooling and grid assumptions. Google's own 2025 measurement of its Gemini system put a median text query at roughly five drops of water — several orders of magnitude lower. Both figures can be true simultaneously: they describe different models, different data centre cooling architectures, different regional grids, and different query lengths, and neither is a universal constant for "AI water use." This paper treats per-query estimates as illustrative of the mechanism, not as a stable statistic to be quoted without its underlying assumptions — a caution this paper extends to its own use of any single per-query number elsewhere in this text.

4.3 Cooling Technology Choices — Why the Method Matters as Much as the Climate

Observation: most public commentary on data centre water use talks about "cooling" as a single undifferentiated activity. Evidence: in practice, three materially different cooling architectures dominate the industry, with very different water consequences. Open evaporative cooling (cooling towers) is the most water-intensive: it deliberately evaporates water to remove heat and is the technology behind most of the withdrawal figures in Part IV. Direct air cooling (drawing outside air, common in cool climates such as the Nordics or the Pacific Northwest) uses little to no water but only works efficiently below a given ambient temperature threshold, roughly 20-25°C depending on design, which rules it out for much of the tropical and desert geography this paper's case studies are drawn from. Closed-loop liquid cooling and immersion cooling (submerging servers directly in a dielectric fluid) recirculate a fixed volume of coolant with minimal ongoing water loss, at a materially higher upfront capital cost and, for immersion in particular, a still-developing maintenance and supply chain. Insight: the honest picture is that no cooling technology is universally superior — each is a different trade against local climate, water availability, and capital budget — which is precisely why Part VII treats siting and cooling-technology choice as two separate, complementary levers rather than a single fix. Implication: a data centre in a water-stressed, hot climate faces the least favourable version of this trade-off on every axis simultaneously, which is exactly the combination present in Bengaluru, Chennai, and much of the Gulf — making those geographies the highest-priority candidates for closed-loop retrofit, not the lowest.

Part V — Geography of Water: Three Basins Under Stress

5.1 Bengaluru, India

India's data centre capacity grew from roughly 375 MW in 2020 to nearly 1,500 MW by 2025 — a fourfold increase in five years — and national data centre water consumption is estimated at approximately 150 billion litres annually, with research projecting that figure to more than double by the end of the decade. A single 20 MW facility can require on the order of 1.4 million litres of water per day, comparable to the daily water needs of roughly 27,000 urban households.

Bengaluru concentrates the risk. The city hosts roughly 30 data centres, drawing on a municipal water system already under severe strain: the city relies on approximately 2,000 million litres per day from the Cauvery river against citywide demand that can reach 3,000 million litres per day, and roughly 6,900 government-operated borewells in the city have already run dry. India ranks second only to Indonesia among Asian countries in the share of data centres sited in regions classified as "extremely high" water stress, and Mumbai and Chennai — which together host roughly 70% of India's data centre capacity — both face documented water stress of their own, meaning the country's data centre build-out is disproportionately concentrated in exactly the geographies least able to absorb new demand.

5.2 Uruguay and Chile — Drought Meets Data

Both countries experienced multi-year droughts through the mid-2020s severe enough to threaten municipal drinking-water supply, and both saw proposed hyperscale data centre campuses become flashpoints for public opposition on that basis — most prominently a large proposed Google facility in Uruguay, where residents and civil-society groups organised specifically around the optics and substance of allocating scarce municipal water to a foreign-owned server farm during an active drought emergency. The Uruguayan and Chilean cases are documented in contemporaneous regional reporting as part of a broader pattern research groups describe as the data centre fight "going global" — moving from a niche planning objection into mainstream local politics in water-stressed middle-income economies.

5.3 The Netherlands — When Opposition Turns Physical

Dutch opposition to data centre expansion has been driven as much by land use and electricity grid congestion (the Netherlands has among the most constrained transmission grids in Western Europe) as by water directly, but it illustrates how quickly local resistance can escalate. In July 2026, activists opposed to a Microsoft data centre project were reported to have thrown acid at construction equipment at the site — an extreme, isolated act, but one that registered internationally as a signal that data centre siting conflicts in wealthy, high-regulation European economies had moved beyond planning-committee objections into direct action.

Part VI — The Grid Bottleneck and the Equity Question

Across all three power geographies in Part III, the same structural pattern recurs: a utility's demand forecast is revised sharply upward by data centre interconnection requests, a large capital programme is approved to serve that demand, and the cost of that capital programme is recovered — in whole or in large part — through the standard rate base charged to all customers, residential and commercial alike, under the traditional utility regulatory model built for an era of predictable, broadly-shared demand growth.

This produces what several U.S. state regulators and consumer advocates have begun calling a large-load cost-allocation problem: the beneficiary of new infrastructure (a data centre operator, often paying a negotiated large-load tariff below the average system cost) is not the party bearing the majority of the infrastructure's financing cost (the general ratepayer). Diesel backup generator emissions from data centre campuses have also been found, in independent monitoring, to run as much as 662% higher than what facilities self-report in permit filings — a gap that matters because backup generation is typically permitted and regulated with the assumption of infrequent, short-duration use, an assumption strained by grids that increasingly rely on data centres' own backup capacity as informal peaking reserve.

The pattern across all three: the regulatory response arrives only after demand has already outrun the grid's planning assumptions, never ahead of it. This is the single clearest argument, developed further in Part VII, for treating resource-intensity disclosure and siting policy as a precondition for new data centre capacity rather than a retrofit applied once strain is already visible.

6.1 The Diesel Backstop Problem

Large data centres typically hold diesel generator capacity sized to carry the entire facility through an extended grid outage — often for 48 to 96 hours — because a single unplanned shutdown of a hyperscale campus is commercially unacceptable to its tenants. That backup capacity is usually permitted on the regulatory assumption that it will run only rarely, for short test cycles and genuine emergencies. Independent monitoring cited in Part VI found actual diesel-generator emissions running as much as 662% above what facilities self-report — consistent with a pattern in which backup generators are being used more frequently than their permits assume, including, in some documented cases, as informal peaking capacity during grid stress events rather than true emergency backup. This is a second, distinct pollution channel from the electricity-sourcing question in Part II: even a data centre supplied entirely by clean grid power still typically holds a fossil-fuel-burning backup fleet sized for total independence from that grid.

6.2 Land Use, Noise, and the Erosion of Local Consent

Beyond electricity and water, hyperscale campuses consume large tracts of land — often hundreds of acres — and operate industrial-scale cooling fans and transformers that generate continuous low-frequency noise, a documented source of neighbour complaints in Northern Virginia, Ireland, and the Netherlands alike. Local governments frequently approve individual facilities as isolated planning applications, without a cumulative-impact assessment of what the fifth or tenth data centre in a given county does to the local grid, water table, and quality of life collectively — a structural blind spot rather than any single facility's fault, and one that helps explain why opposition in Part V's case studies tends to escalate suddenly rather than build gradually: residents often do not perceive the cumulative scale until a threshold is crossed.

6.3 Tax Incentives and the Local Fiscal Trade-Off

Most jurisdictions courting data centre investment offer significant tax abatements — commonly on property tax and sales tax for equipment purchases — to win the facility over a competing location. These abatements are a real, quantifiable cost to local government revenue at exactly the moment that same government may be asked to fund grid and water infrastructure upgrades to serve the facility, as documented in Virginia's rate-base dynamics in Part III. The jobs argument commonly used to justify these abatements should be read carefully: hyperscale data centres are among the most capital-intensive, least labour-intensive major industrial facilities built today — a large campus typically employs several dozen to a few hundred permanent operational staff post-construction, a materially smaller multiplier per dollar of tax abatement than the manufacturing or logistics investments such incentive programmes were originally designed to attract.

What Real-World Interventions Have Actually Shown

NGE's house standard is to ground policy proposals in interventions that have actually been tried, not only in what sounds plausible. Four cases meet that bar.

Singapore's Efficiency-Gated Capacity Allocation

Singapore's current DC-CFA2 framework is the most concrete real-world example of a government using new-capacity access itself, rather than a tax or a voluntary pledge, as the enforcement mechanism for efficiency. Requiring a 1.25 PUE, Green Mark Platinum certification, and 50% green-energy procurement as conditions of receiving any of the 200 MW (and potentially 700 MW) newly released has produced, by industry accounts, some of the most efficient large-scale data centre construction pipelines anywhere in Asia — because operators who cannot meet the bar simply do not receive an allocation, removing the free-rider problem inherent in voluntary sustainability commitments.

Cornell's Siting-Based Water Footprint Reduction

Research led by Fengqi You's group at Cornell, published via Nature Sustainability-linked analysis, modelled the effect of strategic geographic siting — locating new data centres in lower-drought-risk regions with abundant renewable generation, such as Montana, Nebraska, South Dakota, and parts of Texas outside its most stressed aquifers, rather than defaulting to already-crowded hubs — and found that siting strategy alone, without any change to cooling technology, could reduce AI's projected future water footprint by up to 86%. This is the single strongest piece of evidence in this paper's entire research base for a specific, quantified solution: where a data centre is built matters at least as much as how it is built.

The Nuclear Power-Purchase Wave

Since 2024, major technology companies have signed a series of long-duration power-purchase agreements tied to nuclear generation, treating grid decarbonisation and 24/7 reliability as a single problem rather than two. Microsoft's agreement with Constellation Energy to restart the Crane Clean Energy Center (formerly Three Mile Island Unit 1) secures 837 MW over 20 years. Amazon Web Services' agreement with Talen Energy secures 1,920 MW from the Susquehanna Steam Electric Station in Pennsylvania through 2042. Meta has committed to purchase the expected output of 6 GW of new nuclear capacity and separately holds an option on a 300 MW small modular reactor through a Vistra-linked agreement. Across these and comparable deals, the industry-tracked total for tech-sector nuclear commitments between 2024 and 2026 runs to roughly 2.7 GW of restarted, extended, or newly committed nuclear capacity — the first time in decades that private commercial demand, rather than public utility planning, has been the primary driver of new nuclear investment in the United States.

Ireland's Bidirectional Grid Requirement

Ireland's 2025 policy reset — requiring new data centres to be capable of exporting power back to the grid and demonstrating demand flexibility during stress events — represents a genuinely novel regulatory instrument: rather than simply rationing connections, it converts the largest new loads into partial grid assets. It is too new, as of this paper's publication, for outcome data to exist; it is included here as a mechanism worth tracking rather than a proven result, in the same epistemic tier as the SMR caveat above.

Odense, Denmark: Waste Heat as District Heating

Since late 2019, surplus heat recovered from a Meta data centre in Odense has been raised to 70-75°C using large ammonia heat pumps and fed directly into the city's district heating network — recovering approximately 215,000 MWh of surplus heat annually, with total heat output capacity of roughly 45 MW, and warming more than 12,000 homes as of 2023. The servers themselves are powered entirely by wind energy, meaning the same electricity is effectively used twice: once for computation, once for heat. This is the clearest existing proof that a data centre's waste heat — treated everywhere else in this paper as pure loss — can be converted into a second useful output, provided a district heating network already exists nearby to receive it.

Google DeepMind's AI-Optimised Cooling

Google's DeepMind team demonstrated as early as 2016 that a machine-learning system, given real-time telemetry from a data centre's cooling infrastructure, could learn to control pumps, chillers and cooling towers more efficiently than static, rule-based control systems — cutting the energy used for cooling specifically by roughly 40% at the facilities where it was deployed. This is one of the oldest results cited in this paper, and it predates the current AI-driven data centre boom entirely — an irony worth naming directly: some of the clearest evidence that AI can reduce a data centre's own resource footprint comes from before the current wave of AI-driven data centre construction began. The technique has since been generalised across Google's fleet and echoed by comparable AI-driven building-management systems from other operators and third-party vendors.

AWS's Recycled-Water Cooling Programme

Amazon Web Services has expanded its use of recycled (non-potable, treated) water for evaporative cooling from 24 to more than 120 U.S. data centre locations, a programme the company states will preserve over 530 million gallons of fresh drinking water annually — roughly 800 Olympic swimming pools' worth — once fully deployed. Against a company-wide "water positive by 2030" target, AWS reported reaching 53% of that goal by the end of 2024. The mechanism is straightforward and directly addresses the core tension in Part IV: evaporative cooling still consumes water, but it does not need to consume drinking-quality water, and substituting treated wastewater or reclaimed water for potable supply removes the facility from direct competition with municipal and agricultural users for the same protected resource — the exact conflict documented in the Bengaluru and Uruguay cases in Part V.

Part VII — The Solutions Framework: Eight Pillars

7.1 Pillar One: Mandatory Water and Power Usage Disclosure

Require facility-level, not just corporate-aggregate, disclosure of PUE (Power Usage Effectiveness) and WUE (Water Usage Effectiveness) for any data centre above a defined load threshold, reported to the local grid or water regulator and made public. Singapore's regime shows disclosure paired with enforcement changes behaviour; disclosure alone, without the gating mechanism in Pillar Two, is a necessary but insufficient first step. In practice, this means standardising a single reporting template — most jurisdictions currently accept whatever format an operator's own sustainability report uses, which makes cross-site and cross-company comparison close to impossible for regulators and researchers alike.

7.2 Pillar Two: Efficiency- and Resource-Gated Capacity Allocation

Following Singapore's model directly: make access to new grid or water allocation conditional on meeting a defined efficiency threshold and a minimum share of low-water or closed-loop cooling and clean generation, rather than treating efficiency as a voluntary best practice pursued only by cost-motivated operators. The mechanism that makes this work is sequencing: efficiency requirements attached before a permit or interconnection is granted change behaviour; the same requirements published as guidance after construction has already begun do not.

7.3 Pillar Three: Siting Incentives Tied to Regional Water Stress

Applying Cornell's 86% finding directly as policy: national or state-level incentives (accelerated permitting, tax treatment, or preferential interconnection queue position) for data centres sited in verified low-water-stress, high-renewable-availability regions, and correspondingly slower or more conditional approval pathways in regions already classified as high or extremely high water stress, mirroring the classification India itself already uses to flag Bengaluru, Mumbai and Chennai. Part VIII sets out the practical mechanics of how a specific site would be evaluated against this pillar.

7.4 Pillar Four: Closed-Loop Cooling, Water Reuse, and Waste-Heat Offtake Targets

Set phased targets for new facilities to adopt closed-loop liquid cooling or direct air cooling in place of open evaporative cooling, particularly in any region already flagged under Pillar Three. Where evaporative cooling remains the most practical choice for climate reasons, require substitution of recycled or reclaimed water for potable supply as the default — AWS's expansion to over 120 U.S. locations, preserving an estimated 530 million gallons of fresh water annually toward a 2030 water-positive target, demonstrates this is achievable at scale, not merely a pilot-project gesture. Where a district heating network or a large industrial heat user already exists nearby, require an engineering feasibility study of waste-heat offtake before construction — Odense's 215,000 MWh of annually recovered heat, warming over 12,000 homes, shows the economics can work when the receiving infrastructure is already in place.

7.5 Pillar Five: Large-Load Cost Allocation Reform

Reform utility rate-base treatment so that data centre operators bear a share of new transmission and generation capital costs proportional to their actual contribution to the demand growth that necessitated it, rather than the current pattern — visible directly in the Virginia case in Part III — in which residential ratepayers fund the majority of infrastructure built to serve a single commercial customer class. This should extend to backup diesel generation: permitting authorities should require metered, auditable run-time reporting for all backup generators above a defined capacity threshold, directly responding to the 662% self-reporting gap documented in Part VI.

7.6 Pillar Six: Clean Firm Power Procurement Standards

Encourage or require new large-load data centre capacity to be matched with firm, low-carbon generation — nuclear power-purchase agreements being the most advanced current example — rather than allowing new demand to be served predominantly by fossil peaking capacity, which several U.S. states have already had to lean on when data centre demand outran planned clean-generation timelines.

7.7 Pillar Seven: Bidirectional Grid Participation

Following Ireland's 2025 policy directly: require new large-load facilities above a defined threshold to demonstrate power-export or demand-flexibility capability, converting them from pure consumers into partial grid assets that can absorb rather than worsen peak-demand stress events.

7.8 Pillar Eight: Community Benefit and Local Water Rights Protection

Establish enforceable local water-rights protections — including a defined minimum reserve for municipal and agricultural use before any new industrial water allocation — in any jurisdiction hosting or considering hosting large data centre capacity, directly addressing the mechanism behind the Uruguay, Chile, and Bengaluru cases in Part V, where the absence of such a protected minimum was the proximate cause of public conflict. This pillar should also require a cumulative-impact assessment at the county or municipal level once a defined number of facilities or aggregate load threshold is reached in one area, closing the "isolated planning application" blind spot identified in Part VI.

Part VIII — How to Choose a Site: A Practical Siting Methodology

Every geography examined in Parts III and V shares a common root cause: the siting decision was made — implicitly or explicitly — before the resource constraints it would collide with were fully weighed. This section sets out, step by step, the due-diligence sequence this paper argues a developer, utility, or regulator should actually work through before committing to a site, translating the eight pillars above into an ordered checklist rather than a list of principles.

8.1 Step One — Map Regional Water Stress Before Anything Else

Screen every candidate site against a recognised water-stress classification (the kind of "extremely high water stress" designation already used to flag Bengaluru, Mumbai, and Chennai in Part V). A site inside a basin already classified as highly or extremely water-stressed should require an explicitly higher burden of proof — demonstrated closed-loop or air cooling, or a confirmed non-potable water source — before proceeding to the next step at all. This is the single highest-leverage screen available, given Cornell's finding that siting choice alone can move the water outcome by up to 86%.

8.2 Step Two — Check Grid Headroom and Interconnection Queue Position, Not Just Nameplate Capacity

A region's advertised generation capacity is a poor proxy for whether it can actually serve a new large load quickly. The more useful question, illustrated directly by Virginia's experience in Part III, is the utility's current interconnection queue: how many other large-load projects are already ahead of this one, and how far has the utility's own demand forecast already drifted from its planning assumptions. A region where the queue is short and the forecast has been stable is a materially safer bet than one with high nameplate capacity but a queue backlogged by years.

8.3 Step Three — Match Climate to Cooling Technology, Not the Reverse

Rather than choosing a cooling technology and then forcing it onto a site, work in the other direction: a cool, temperate climate (Nordic countries, the Pacific Northwest, parts of Canada) makes direct air cooling viable and removes most of the water question entirely; a hot, humid, or water-stressed climate should trigger a presumption in favour of closed-loop or immersion cooling from the outset, accepting the higher capital cost as the price of operating in that geography, rather than defaulting to open evaporative cooling because it is cheaper to build.

8.4 Step Four — Assess Clean Firm Power Proximity

Favour sites within economic transmission distance of existing or committed firm, low-carbon generation — hydro, nuclear, or geothermal — over sites that would need new fossil peaking capacity built specifically to serve the facility. The nuclear power-purchase agreements catalogued in Part VII show this is now a live, executable option for large operators, not a theoretical ideal; the same logic applies at smaller scale to hydro-rich regions.

8.5 Step Five — Check for Waste-Heat Offtake Potential

Before finalising a site, assess whether a district heating network, greenhouse cluster, or other large heat user exists within economic pipeline distance. Where one does, as in Odense, the facility's waste heat becomes a second revenue stream and a genuine community benefit rather than pure thermal loss — and that benefit can materially improve the facility's standing with local authorities and residents during the approval process itself.

8.6 Step Six — Assess Disaster, Seismic, and Climate-Change Exposure

Screen for flood risk, seismic activity, wildfire exposure, and long-run climate projections for the site's operating lifetime (typically 20-plus years) — a data centre's own uninterruptible-power and cooling infrastructure is designed for grid failure, not for the facility's physical foundations being underwater or its access roads being cut, and a facility built without this screen becomes a stranded-asset risk decades before its financing is repaid.

8.7 Step Seven — Engage Local Consent Early, Not After Design Is Fixed

The Netherlands and Uruguay cases in Part V both involved opposition that hardened after community members felt a decision had already effectively been made before they were consulted. Genuine early engagement — sharing the water and power footprint numbers publicly before finalising design, not after — costs schedule time upfront but measurably reduces the risk of the kind of escalation that stalls or cancels a project outright later in construction, when far more capital is already committed.

8.8 Step Eight — Confirm Regulatory and Water-Rights Stability

Verify that water allocation rights for the site are legally secure and not subject to a senior competing claim (agricultural or municipal) that could be prioritised over the facility's allocation during a future drought — the precise vulnerability exposed in Uruguay and Chile. Favour jurisdictions with an explicit protected-minimum allocation for municipal and agricultural use already in law, per Pillar Eight, over jurisdictions where industrial water rights are informally negotiated case by case.

Part IX — Anticipating Industry Pushback

8.1 "Efficiency Mandates Will Slow AI Investment and Innovation"

Singapore's experience is the direct rebuttal available: the 2026 DC-CFA2 round, gated on some of the strictest efficiency standards in Asia, attracted sufficient investor interest to justify expanding the programme toward a 700 MW dedicated park. Resource constraints have historically driven innovation in cooling and siting technology faster than abundance has — the closed-loop cooling and modular design advances now being deployed at scale first emerged in exactly the water- and power-constrained markets (Singapore, parts of the Middle East) that had no choice but to innovate.

8.2 "Cost Allocation Reform Will Just Raise Data Centre Prices, Which Gets Passed to AI Users Anyway"

That is, to a degree, the intended and correct outcome: the current allocation understates the true infrastructure cost of AI compute by shifting a share of it onto residential ratepayers who receive no benefit from the facility. Correcting that allocation makes AI compute pricing more accurately reflect its actual resource cost — a standard economic argument for internalising an externality, not a novel one specific to this paper.

8.3 "Siting Restrictions in Water-Stressed Regions Will Just Push Facilities to Countries With Weaker Regulation"

This is a genuine risk and the strongest objection this paper faces. It is also, however, an argument for coordinated international disclosure standards (Pillar One) rather than an argument against siting policy altogether — the same "race to the bottom" concern applies to environmental and labour regulation generally, and the conventional answer (baseline international disclosure norms, even where enforcement varies) applies here with equal force.

Part X — An Original Metric: The Grid-Water Stress Score

Every efficiency metric in current use — PUE, WUE, carbon intensity — measures one resource in isolation. This paper proposes the Grid-Water Stress Score (GWSS), a composite index intended to let regulators, investors, and communities compare a proposed data centre site on both binding constraints simultaneously, rather than approving a facility that scores well on power while sitting in an "extremely high" water-stress zone, or vice versa.

The weighting (40/40/20) reflects this paper's own judgement that siting fundamentals matter more than facility-level engineering, consistent with Cornell's finding that siting alone can move the water outcome by up to 86% — a scale of impact facility efficiency improvements alone have not been shown to match. Regulators adopting a GWSS-style tool would set their own weights and thresholds; the framework's purpose is to make the trade-off explicit and comparable across proposed sites, not to prescribe a single correct weighting.

Part XI — Policy Implications

• For national energy regulators: adopt Ireland's bidirectional-capability requirement and Singapore's efficiency-gated allocation as the two most immediately transferable, already-implemented models, rather than starting policy design from first principles.
• For utility rate regulators: revisit large-load cost-allocation rules using Virginia's experience as the cautionary case study — a five-fold demand-forecast revision inside five years is evidence that existing forecasting and cost-recovery frameworks were not built for this scale of single-sector growth.
• For water regulators in stressed basins: adopt an explicit protected-minimum allocation for municipal and agricultural use ahead of any new industrial water permit, directly responding to the Bengaluru and Uruguay patterns in Part V.
• For technology companies: treat Cornell's siting finding as the highest-leverage lever available — a facility sited in a low-water-stress region with strong renewable availability starts from an 86%-better position than the same facility built in a contested basin, before any cooling-technology investment at all.
• For investors and capital allocators: a GWSS-style composite screen, applied before capital commitment, would have flagged the resource risk in several of the specific projects in Part V ahead of the public opposition that ultimately materialised.

Part XII — Conclusion

The AI build-out is, at its physical core, an energy and water infrastructure build-out wearing a software story's clothing. Roughly 945 TWh of projected 2030 electricity demand, a water footprint that could reach over a trillion litres annually in the United States alone, and six geographies — from Virginia to Bengaluru, from Dublin to Montevideo — already living the consequences of that scale are not future risks to be modelled. They are present-tense facts, documented in utility filings, peer-reviewed water research, and, in the Netherlands' case, in an act of physical sabotage that made international news.

What distinguishes this paper's argument from a simple call for caution is the evidence, assembled in Part VII, that specific, already-implemented interventions — Singapore's efficiency-gated capacity allocation, Cornell's siting research, the nuclear power-purchase wave, Ireland's bidirectional-grid requirement — have already measurably worked, at real scale, without halting the underlying build-out. The choice facing regulators, companies and communities is therefore not between AI growth and resource protection. It is between adopting mechanisms that have already been shown to reconcile the two, and continuing to let the two collide first and negotiate the cost afterward, as Virginia's ratepayers and Bengaluru's dry borewells are already doing.

References

International Energy Agency (2026). "Energy and AI." IEA, Paris.

International Energy Agency (2026). "Electricity 2026: Analysis and Forecast to 2028." IEA, Paris.

International Energy Agency (2026). "Electricity Mid-Year Update 2026." IEA, Paris.

Ren, S. et al. (2023). "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models." University of California, Riverside / arXiv:2304.03271.

University of California, Riverside News (2023). "AI Programs Consume Large Volumes of Scarce Water."

You, F. et al. (2025). Cornell University research group, published via Nature Sustainability-linked analysis on U.S. data centre water withdrawal projections to 2030.

Knowable Magazine (2026). "How Much of a Problem Is AI's Water Use?"

CBS News (2026). "How Much Water Does AI Really Use?"

Genetic Literacy Project (2026). "AI Data Centers and Water Usage: It's a Genuine and Growing Problem."

Roosevelt Institute (2026). "Virginia's Data Center Alley Is a Cautionary Tale for Equity and Climate."

Virginia Business (2026). "Data Center Growth Could Outpace Power Generation." Dominion Energy grid challenges reporting.

Data Center Dynamics (2026). "Dominion Energy Admits It Can't Meet Data Center Power Demands in Virginia."

Amperon (2026). "Why Virginia Data Centers Matter to All of PJM."

Broadband Breakfast (2026). "Dateline Ashburn: Data Centers Drive New Energy Disputes in Northern Virginia."

William Fry (2026). "Ireland's Data Centre Connections: Back Online." Legal analysis of CRU final policy.

KPMG Ireland (2026). "Ireland's Data Centre Policy Reset."

EnergyConnects (2025). "Ireland Ends Moratorium on New Power Links to Data Centers." December 2025.

DC Byte (2026). "Ireland Reopens the Grid for Data Centres."

Pinsent Masons, Out-Law (2026). "New Irish Data Centres Must Generate Power Back Into National Grid, Rules Regulator."

Introl (2026). "Singapore's Green Data Center Gamble: DC-CFA2 2026."

KWM (2026). "Singapore Launches 200MW Data Centre Call for Application (DC-CFA2)."

Morgan Lewis (2026). "Singapore Announces Data Center Capacity Allocation Call."

Data Center Dynamics (2026). "Singapore Authorities Invite Applications for New Data Centers."

Down To Earth (2026). "India's AI Data Centre Boom Is Draining Water: Weak Regulation Threatens Long-Term Water Security."

Down To Earth (2026). "India's AI Data Centres Are Draining Water as El Niño Threatens a Weak 2026 Monsoon."

Observer Research Foundation (2026). "The Cloud Runs on Water."

The Better India (2026). "AI Needs More Than Data & Power: Is Water India's Next Hidden Cost?"

Reccessary (2026). "Data Center Water Use to Double by 2030: Whose Water Is AI Drinking in India?"

Countercurrents (2026). "AI Data Centres vs. Water Crisis: How Digital Infrastructure Is Draining India's Wells Dry."

Nearshore Americas (2026). "Water Scarcity Jeopardizes Data Center Projects in Chile and Uruguay."

Nearshore Americas (2026). "Water-Guzzling Data Centers Spark Outrage Across Latin America."

Waging Nonviolence (2026). "The Data Center Fight Is Going Global." July 2026.

The Register (2026). "Amsterdam Activists Throw Acid at Microsoft Datacenter Project." July 2026.

The Borgen Project (2026). "New Data Center Threatens Further Water Scarcity in Uruguay."

EnkiAI (2026). "Microsoft Nuclear 2026: 1,920 MW Amazon PPA" and "Data4 Nuclear 2026: 837 MW Microsoft Deal." Nuclear power-purchase-agreement tracking.

Introl (2025). "SMR Nuclear Power AI Data Centers 2025."

Ramboll (2023). "Meta: Surplus Heat to District Heating." Odense, Denmark case study — 215,000 MWh annual heat recovery, 45 MW capacity, 12,000+ homes served.

Microsoft Local (2019). "Surplus Datacenter Heat Will Be Repurposed to Heat Homes in Denmark."

Google DeepMind (2016, ongoing generalisation to present). "DeepMind AI Reduces Google Data Centre Cooling Bill by 40%."

Amazon Sustainability (2026). "How AWS Uses Recycled Water in Data Centers." Expansion to 120+ U.S. locations; 530 million gallons preserved annually; 53% progress toward 2030 water-positive target.

Gravel2Gavel, Pillsbury Law (2026). "Water Reuse Emerges as a Critical Strategy for Data Center Development." June 2026.

This white paper is offered as a contribution to the debate on the physical economics of the AI build-out. The authors welcome comments, critique, and correction.