New ZERO framework puts AI data-centre water and carbon decisions before consumption

Most assessments of artificial intelligence sustainability arrive too late. By the time electricity, water and materials have been measured, a workload has already been accepted, a facility has already been chosen and the resources have already been committed. A new Perspective published in the journal Carbon Neutral Systems argues that this ordering is the core problem, and it proposes a governance framework called ZERO that moves the decisive questions upstream, before consumption begins. The paper, authored by Maria Kollia and published on 2 September 2026, also includes a retrospective evaluation built entirely from public 2024 environmental data covering six Google data-centre locations in the United States.

ZERO is the name of the proposed framework, not an acronym and, the author stresses, not a claim of literal zero physical consumption. Its governing premise is that avoidable demand and preventable resource loss should be challenged before they are treated as inevitable. The framework integrates workload necessity, resource state, local constraints, recovery pathways, network coordination, evidence freshness and accountable intervention into a single decision architecture. Earlier conceptual papers by the same author introduced separate elements of zero-waste infrastructure, circular recovery and adaptive resource governance; the new paper consolidates those elements into an explicit operational model and adds the empirical replay.

The scientific contribution, the paper argues, is not a new efficiency indicator but a decision architecture with five testable properties. First, hard local constraints are applied before optimisation. Second, workload and resource states are represented separately. Third, only admissible options enter network comparison. Fourth, approval remains conditional on evidence validity and changing conditions. Fifth, the resulting decision and physical outcome are recorded against a credible baseline. Familiar facility metrics such as power usage effectiveness, water usage effectiveness and carbon usage effectiveness describe the performance of an operating system, the paper notes, but they do not by themselves determine whether a particular computation should start, whether its scale is justified, whether it should be delayed, or whether a technically available facility is locally admissible.

The framework is organised as an ordered resource hierarchy. The first duty is to avoid computation that has no defensible purpose or duplicates an adequate result. Necessary computation should then be reduced to an appropriate scale through model choice, precision, batching, reuse or workload redesign. Flexible work should be aligned with suitable temporal and geographical conditions. Residual energy, heat, water and material value should be recovered where technically and environmentally justified, and productive reuse must be distinguished from mere capture. Regeneration is the final and most demanding claim, because it requires evidence that the wider system has been strengthened rather than merely compensated. The ordering is deliberately non-compensatory: heat reuse cannot justify unnecessary computation, renewable electricity does not remove the need for efficient software and hardware, and water replenishment in one location cannot conceal pressure in another.

Formally, a workload request is represented as a tuple containing purpose and necessity, expected IT energy demand, duration, urgency and deadline, location and jurisdiction restrictions, service and resilience requirements, and flexibility attributes such as interruption, division, model substitution and migration. For each facility and decision time, a resource state captures electricity carbon intensity, water condition, cooling reserve, productive heat-use opportunity, grid or local infrastructure condition, material and equipment condition, and evidence validity, each with a measured or forecast value, source, confidence and time-to-live. Hard admissibility is a binary gate: a facility enters the admissible set only when every binding legal, security, safety, data-sovereignty, resilience and local environmental constraint is satisfied and the evidence used to evaluate those constraints remains valid. A failed binding constraint produces an excluded status, while an unresolved but resolvable value can be placed on hold, which is explicitly not a pass.

This rule is intentionally stricter than weighted optimisation. A sufficiently low carbon score cannot compensate for a critical cooling reserve, a prohibited data location or a binding water restriction. Weighted scoring, which combines normalised carbon, water, cooling, recovery, transition cost, service risk and evidence confidence, operates only within the already-admissible set and therefore cannot override a failed hard constraint. The weights themselves require governance: the paper permits stakeholder-derived vectors elicited through the analytic hierarchy process, objective vectors estimated by an entropy-weight method, or a declared hybrid of the two, with the method and weights fixed before inspecting the preferred outcome. Admission is also not a permanent licence. Long-running workloads continue only while their material conditions remain valid, and reassessment is triggered when a resource state crosses a declared threshold, evidence expires, confidence falls, a local restriction changes, or the predicted consequence differs materially from the observed one.

To examine two of these properties with observed data, the study assembled a retrospective secondary-data panel rather than a deployment trial. Six U.S. locations were retained where a public FY2024 Google water entry, an unambiguous county, an explicit Google Cloud location-to-region mapping and a 2024 regional grid-carbon value could be joined without interpolation: The Dalles in Oregon, Midlothian in Texas, Ashburn in Virginia, Henderson in Nevada, Council Bluffs in Iowa and Berkeley County in South Carolina. Three data layers were harmonised: Google’s 2025 Environmental Report water withdrawal, discharge and consumption figures by site; Google Cloud’s 2024 annual regional grid carbon intensity and carbon-free-energy percentages; and 52 weekly county-level observations from the U.S. Drought Monitor of the percentage of county area in moderate-to-exceptional drought, joined by county FIPS code. The resulting panel contains 312 location-week records.

The empirical question was deliberately narrow: given the observed 2024 regional carbon ranking and the observed weekly county drought state, how often would a water-related hard gate remove the carbon-only spatial choice, and how often would changing conditions reopen admissibility? A carbon-only comparator selected, in each week, the candidate with the lowest reported 2024 grid carbon intensity. Because that carbon source is an annual regional average, its within-year ranking was fixed, and all weekly variation came from the observed drought series. Under the ZERO replay, a site remained admissible when county drought coverage fell below a declared threshold and was placed on hold when coverage met or exceeded it. Three thresholds of county-area drought coverage were predeclared for sensitivity analysis: 50, 75 and 90 percent. The paper is explicit that these are not proposed as universal ZERO thresholds, are not attributed to Google or the U.S. Drought Monitor, and represent progressively less precautionary spatial-coverage rules applied to the same observed series.

The results show a non-zero, threshold-sensitive decision effect. The carbon-only comparator selected The Dalles in all 52 weeks, because Oregon had the lowest 2024 regional grid carbon intensity in the six-location set at 79 grams of CO2-equivalent per kilowatt-hour, followed by Texas at 303, Northern Virginia at 323, Nevada at 357, Iowa at 413 and South Carolina at 576. Once the drought gate was applied, that choice was no longer invariant. At the 50 percent threshold, The Dalles was held in 13 weeks and the ZERO choice differed from the carbon-only choice in 13 of 52 weeks, or 25.0 percent; Midlothian was the lowest-carbon admissible substitute in 11 of those weeks, Council Bluffs in one, and in one week no location in the six-site sample remained admissible. At the 75 percent threshold the carbon leader was held in six weeks, or 11.5 percent, and at 90 percent in three weeks, or 5.8 percent, with Midlothian selected in every case. Reported FY2024 site water consumption in the sample ranged from 56.0 million gallons at Ashburn to 1,010.2 million gallons at Council Bluffs, values the paper treats as descriptive context rather than workload-normalised efficiency.

The weekly series also exercised the revalidation property. All six counties entered moderate-to-exceptional drought during at least part of 2024, and at the 50 percent threshold the six sites underwent 17 admissibility-state transitions in total, with every site changing state at least once at all three thresholds. In Wasco County, home to The Dalles, early-year drought coverage exceeded all three thresholds, while later coverage returned to roughly 50 percent, so a 50 percent rule reopened the placement decision while the stricter rules did not. An independent sensitivity check using U.S. EPA eGRID2023 state-level emission rates confirmed that Oregon remained the lowest-carbon state among the six locations, though the ranking of several substitutes differed. The author and the paper’s own conclusions are careful about scope: the replay does not show that Google should have moved workloads, that county drought caused facility-level water scarcity, that the thresholds are optimal, or that ZERO reduced emissions or water use in operation. Those claims, the paper states, require site telemetry, workload traces, service constraints, locally authorised thresholds and observed post-decision outcomes, and the proposed validation path runs from time-matched retrospective data through shadow-mode prototyping to controlled pilots and, eventually, cross-federation trials.

Subject of Research: A governance framework that applies pre-emptive environmental admissibility constraints to AI data-centre workload placement, evaluated retrospectively with 2024 public water, carbon and drought data.

Article Title: Governing AI data-centre resources before consumption: the zero framework

Article References: Kollia, M. (2026). Governing AI data-centre resources before consumption: the zero framework. Carbon Neutral Systems, 2(1), Article 15. https://doi.org/10.1007/s44438-026-00037-z

Image Credits: AI Generated

DOI: 10.1007/s44438-026-00037-z

Keywords: AI data centres, ZERO framework, resource governance, water use, carbon intensity, drought monitoring, admissibility constraints, revalidation, Google data centres, carbon-aware computing, sustainability metrics, retrospective replay