Adjusted Water Impact (AWI) Metric

• Adjusted Water Impact (AWI) is a comprehensive metric that quantifies net water burden by incorporating on-site, off-site, and supply-chain water use with local scarcity factors.
• It employs normalized, scarcity-weighted formulations and models (e.g., SCARF, MEIO) to differentiate water impacts across spatial and temporal scales in energy-intensive operations.
• Its applications span workload scheduling, infrastructure siting, and supply chain optimization, though challenges include data gaps and modeling uncertainties.

Adjusted Water Impact (AWI) quantifies the net water burden attributable to a specific technological activity, economic entity, or node within an operational system, after accounting for direct (on-site), indirect (off-site), supply-chain, and spatiotemporally-varying water scarcity. It is a metric designed to allocate water consumption in a manner that reflects not only volumetric withdrawals but also hydrological context and responsibility chains, thus providing a more meaningful comparison of water footprint across geographies, workloads, and time horizons. AWI is central to sustainability accounting in advanced data center operations, large-scale AI, and multi-entity industrial systems.

1. Mathematical Definitions and Core Formulations

Several parallel formalisms for AWI exist, each targeting different system boundaries and analytic purposes. All build upon the foundational separation of water use into on-site (Scope 1), off-site (Scope 2), and supply-chain or indirect (Scope 3).

1.1. AWI for AI Workloads and Data Centers

AWI can be defined as normalized, scarcity-weighted total water use, with the canonical formula: $$\mathrm{AWI}_\mathrm{per\;unit} = \frac{S_r}{N_{\mathrm{base}}} \left( \sum_{t=1}^{T} e_t [\rho_{s1,t} + \theta_t \rho_{s2,t}] + \frac{T}{T_0} W \right)$$ where:

• $e_t$: energy used in time-slot $t$
• $\rho_{s1,t}$: on-site water usage effectiveness (WUE) (L/kWh)
• $\rho_{s2,t}$: off-site (grid) WUE (L/kWh)
• $\theta_t$: power usage effectiveness (PUE)
• $W$: total embodied water (L, Scope 3)
• $T_0$: asset operational lifetime
• $S_r$: regional water scarcity factor
• $N_{\mathrm{base}}$: normalization (e.g., total energy, tokens, inferences)

This structure enables separation of operational and embodied water while systematically capturing variability in local and temporal conditions (Li et al., 2023).

1.2. AWI with Explicit Water-Stress Weighting

The SCARF framework generalizes AWI as

$e_t$0

with

$e_t$1

and $e_t$2 being either a short-term or long-term basin-level water stress factor: $e_t$3 where $e_t$4 are temporal discount weights (Wu et al., 28 Jun 2025).

1.3. Marginal AWI in Electricity-Computation-Water Nexus

For dispatch-aware allocation: $e_t$5 with $e_t$6 as generator withdrawal intensities and $e_t$7 from water-aware optimal power flow (OPF); this formulation supports marginal sensitivity analysis across nodes (You et al., 25 May 2026).

1.4. Multi-Entity Input-Output (MEIO) AWI

MEIO-based AWI for an entity $e_t$8: $e_t$9 where $t$0 is the normalized supplier adjacency and $t$1 are immediate water footprints. This closed-form allocation supports chain-of-responsibility at sector/actor scale (Moghaddam et al., 2014).

2. Spatiotemporal and Scarcity Adjustments

AWI fundamentally encodes both spatial and temporal heterogeneity in water impact:

• Spatial heterogeneity: $t$2, $t$3, and the scarcity factor $t$4 or basin-level $t$5 incorporate local meteorological conditions, grid fuel mix, regional water yields, and withdrawal rates (Li et al., 2023, Wu et al., 28 Jun 2025).
• Temporal variability: Hourly to seasonal variation appears in $t$6 (cooling efficiency), $t$7 (grid dispatch), and PUE; in SCARF, time-projected stress series $t$8 allow discounting for future scarcity.
• Basins and grids: AWI values are basin-indexed or region-indexed, not globalized, so the same absolute water withdrawal yields different effective impacts if executed in, e.g., Arizona versus Sweden.

This explicit embedding enables AWI to reflect actual ecological vulnerability, as in demonstration of a $t$9 difference in AWI for identical AI inference workloads run in low- versus high-stress regions (Wu et al., 28 Jun 2025).

3. Normalization Protocols and Units

AWI requires normalization to support fair cross-comparison:

• Per-energy: Liters (adjusted)/kWh, e.g., for data center or model-level analysis.
• Per-model-token or per-inference: For AI/ML workloads, enabling statistical benchmarking across models and architectures.
• Per economic activity: In MEIO, per output unit or per sector (Moghaddam et al., 2014).

The choice of normalization ($\rho_{s1,t}$0) materially affects interpretive precision. It is critical to maintain congruent baselines when benchmarking sustainability across sites, models, or production chains.

4. Data Sources, Assumptions, and Computational Techniques

AWI metrics are built from heterogeneously sourced data and assumption-driven surrogates:

Data Type	Examples/Sources	Notes
On-site WUE	Literature/monitoring data; cooling-vendor curves	Highly site/weather-specific
Off-site WUE	EWIF database, eGRID subregions (fuel mix dependent)	Grid-dispatch variable
Scarcity factors	UN Water Stress Index, Aqueduct 4.0	Hydrological model errors
Embodied water	Semiconductor LCA, fab benchmarks	High Scope-3 uncertainty
Dispatch patterns	OPF models, SCARF time-series data	For dynamic/marginal AWI

Typical assumptions:

• Amortization of embodied water over asset life is linear.
• Scarcity factors and water stress indices represent current or modeled future hydrological states.
• OPF-based methods assume differentiability and exact grid topology (Li et al., 2023, Wu et al., 28 Jun 2025, You et al., 25 May 2026, Moghaddam et al., 2014).

Computational implementations employ fixed-point iteration (for virtual water content), automatic differentiation (for OPF sensitivity), and matrix inversion (in MEIO).

5. Case Studies and Quantitative Illustrations

AWI enables quantitative, context-sensitive assessment and comparison, as demonstrated across scenarios:

• GPT-3 Training: Microsoft datacenter, 1,287 MWh training, mean PUE 1.17, $\rho_{s1,t}$1 L/kWh, $\rho_{s1,t}$2 L/kWh. Raw withdrawal ≈5.44×10⁶ L. With $\rho_{s1,t}$3 (moderate stress), $\rho_{s1,t}$4 L(adjusted)/kWh (Li et al., 2023).
• LLM Inference: Per-request AWI varies from 0.008 L (Illinois, low stress) to 0.28 L (WA, high stress) for Qwen2.5, solely due to site water-stress differences (Wu et al., 28 Jun 2025).
• Data Center Dispatch: IEEE 30-bus, incremental 1 MW load shift from high- to lower-AWI node can reduce water withdrawal by ≈0.46 m³/h (You et al., 25 May 2026).
• MEIO Example: For three-entity chain (A→B→C), total AWI for entity C aggregates direct water of all upstream suppliers, exceeding nominal direct withdrawal (Moghaddam et al., 2014).

6. Applications, Insights, and Limitations

Applications:

• Location-aware workload scheduling: Shifting AI computation to low-stress regions based on marginal or total AWI.
• Infrastructure siting/planning: Capital expansion decisions aided by long-term, discount-rate-weighted AWI.
• Supply chain optimization: For manufacturing (e.g., fab plant siting), AWI enables multi-basin, multi-facility comparison.
• Virtual water accounting: In power grid management, informs optimization by associating real water withdrawal sensitivities to nodal demand.

Insights:

• AWI exposes "hidden" environmental costs invisible to volumetric metrics and reveals order-of-magnitude differences between nominally similar operations.
• Incorporating both direct and indirect flows, AWI delivers more comprehensive responsibility statements for actors/entities (Moghaddam et al., 2014).

Limitations:

• Scarcity-weighted AWI relies on spatially- and temporally-resolved stress indices, subject to modeling error and subjective parameterization (discount rates, scarcity indices).
• Data gaps (confidential PUE/WUE, undisclosed grid dispatch, or proprietary fab water use) can necessitate using proxies.
• Current MEIO and SCARF AWI do not model water quality or ecological threshold effects; scaling linearly with demand may miss non-recoverable ecohydrological impacts.
• Scope-3 (upstream supply chain) attribution is challenging except in full LCA or MEIO frameworks and often omitted in SCARF/OPF models (Li et al., 2023, Wu et al., 28 Jun 2025, Moghaddam et al., 2014, You et al., 25 May 2026).

7. Relationships to Broader Methodological Landscape

AWI occupies a spectrum bridging engineering, sustainability science, and operational research:

• LCA/footprinting: Extends traditional Life Cycle Assessment with spatiotemporal contextualization and explicit marginal sensitivity.
• Statistical accounting vs. operational optimization: OPF-based AWI supports direct embedding of water attribution into dispatch and dispatch-planning, enabling dynamic response rather than passive reporting (You et al., 25 May 2026).
• Inter-entity reallocation: MEIO AWI uniquely quantifies direct and indirect handprints within sectoral or multi-actor networks, facilitating nuanced policy or supplier decision support (Moghaddam et al., 2014).

A plausible implication is that adoption of AWI will increasingly drive both disclosure standards and operational practices in AI, data center, and manufacturing sectors, reflecting the explicit coupling of water, energy, location, and climate factors in advanced sustainability management.