{"id":"eb548ae2-5795-4b38-ae73-81ecd71ae2b9","arxiv_id":"2506.22773","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SCARF is a water-stress-weighted impact metric that multiplies raw computing water use by a basin-level, time-discounted water stress factor, revealing large location and seasonal differences in the true water cost of AI and datacenters.","lead":"A new framework, SCARF, adjusts computing water-consumption estimates by multiplying them by local water stress, so the same workload in a dry region counts as worse than in a wet one. It adds a time dimension by discounting future water stress, and shows large location and season effects for LLM serving, datacenters, and chip fabs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Off-site water is assigned the facility's water-stress factor even though generation may occur in a different basin; this weakens the spatial accuracy claim of Eq. (6).","rationale":"Good-faith reading: the paper's contribution is a metric that multiplies consumption by basin-level water stress aggregated over time; this is a reasonable and useful idea, and the case studies provide illustrative evidence. I agree with the reader's weakest_assumption: the off-site attribution is the least secure load-bearing step. It is not a stylistic objection but a mismatch between what Eq. (2) measures (water consumed at generation sites) and what Eq. (6) multiplies (facility basin WSF). The fix is either to map generation basins or to state the scope limitation and show sensitivity. The framework can be made correct, so the verdict should remain conditional rather than accept or reject. The 'first general framework' novelty claim also needs softening relative to the water-aware scheduling literature, but that is secondary to the attribution issue and does not change the final recommendation.","tokens_in":10993,"tokens_out":6891,"duration_ms":79353,"concrete_test":"Use the Water Impact Tool's underlying county-level footprints to identify the generation basin (or a generation-weighted average of basins) for each facility's electricity mix, and recompute Eq. (6) with WSF of that generation basin in place of WSF of the facility basin. Then re-rank the locations in Figures 3, 7(b), and 9(b). If any ordering flips, or if any AWI changes by more than 20%, the paper must either correct the attribution or explicitly limit SCARF's spatial claim to on-site water.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Equation (6) defines AWI = (Won + Woff) × WSF_b, where b is the basin of the computing facility. Equation (2) models Woff as water consumed during electricity generation, and the case studies source location-specific WUEoff from the Water Impact Tool [41]. Nothing in §2 states that the generation sites lie in the same basin as the facility; in an interconnected grid they generally do not. Multiplying off-site consumption by the facility's WSF therefore attributes water impacts of generation to the wrong watershed. This is not a small parameter error: it can over- or under-state AWI by the ratio of WSF between basins, which in the case studies is large (Figure 4). Since the abstract's central claim is that SCARF accounts for 'where' water stress occurs, the framework as written does not yet provide a spatially correct treatment of Scope 2 water. All three case studies (Figures 3, 7(b), 9(b)) inherit the assumption, so their absolute AWI values and possibly their relative rankings are affected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SCARF, a four-step framework for evaluating the water impact of computing systems in a way that accounts for spatial and temporal variation in water stress. The framework models on-site (Scope 1) and off-site (Scope 2) water consumption, maps facilities to hydrological basins, constructs a Water Stress Factor (WSF) either from current conditions or from discount-rate-weighted future projections, and computes an Adjusted Water Impact (AWI) by multiplying total raw consumption by WSF. The authors demonstrate the framework on three case studies: LLM serving in Microsoft datacenters, Google's U.S. datacenters, and Intel's U.S. semiconductor fabs. The main claims are that deployment location and time can change adjusted water impact by orders of magnitude, that high consumption in medium-stress regions can outweigh moderate consumption in high-stress regions, and that discount-rate choice can alter long-term site rankings.","tokens_in":11212,"tokens_out":7938,"duration_ms":79621,"significance":"If the framework's spatial accounting is corrected, SCARF is a useful and timely contribution to sustainable-computing evaluation. Its strengths include a transparent, parameter-explicit metric; the use of publicly available basin-level water-stress data (Aqueduct 4.0); publicly released code; and a sensitivity analysis for the discount rate. The three case studies span the computing stack and illustrate that raw water volume alone is insufficient. However, the paper's central claim to be 'the first general framework' that accounts for where water stress occurs is weakened by the treatment of off-site water consumption, and the datacenter case study rests on an ad hoc power-capacity proxy. These issues are fixable, but they affect the numerical results and the interpretation of the metric as spatially accurate.","major_comments":[{"comment":"The AWI definition multiplies total raw water consumption, including Woff from Eq. (2), by the WSF of the facility's basin b. Because Woff represents water consumed at electricity generation sites, which generally lie in different watersheds from the computing facility, this assigns the wrong water-stress factor to Scope 2 water. All three case studies (Figures 3, 7(b), and 9(b)) inherit this assumption. Please revise Eq. (6) to use generation-basin WSF values, e.g., AWI = Won × WSF_b + Σ_g Woff_g × WSF_{b_g}, or clearly justify the approximation and quantify its effect using the spatially explicit WUEoff data from [41].","section":"§2.4, Eq. (6)"},{"comment":"Estimating each Google site's power capacity by taking the maximum reported capacity of any datacenter within a 100-mile radius is an ad hoc approximation. It can over- or under-attribute energy and water consumption to a specific Google site, and the paper provides no sensitivity analysis for the radius choice or for the choice of the maximum rather than another aggregation. Since Figure 7 and Takeaway 3 depend on the resulting consumption volumes, the quantitative datacenter results need to be re-examined or accompanied by a robustness check.","section":"§4.1, power capacity proxy"},{"comment":"The long-term AWI multiplies a current annual water-consumption value (Won+Woff) by a normalized weighted average of future water-stress projections. As written, this is neither a standard discounted lifetime impact (which would sum discounted annual impacts over the facility lifetime) nor a purely current annual-impact metric. The discount-rate sensitivity in Figure 8(b) therefore reflects the weighting of stress projections rather than the timing of the consumption stream. Please clarify whether the long-term WSF is intended as a forward-looking siting indicator or as a discounted lifetime impact, and adjust the aggregation and notation accordingly.","section":"§2.3, Eqs. (4)–(6)"}],"minor_comments":[{"comment":"The text says 'maximum reported power capacity P (kWh)', but power should be expressed in kilowatts (kW), not kilowatt-hours; the energy calculation E = P × 24 × 365 × 0.7 then uses power correctly, so the unit label needs to be fixed.","section":"§4.1"},{"comment":"The source and calculation of the monthly water-stress values plotted in Figure 4 are not described; §2.2 only mentions retrieving current and projected (2030/2050/2080) stress from Aqueduct 4.0, while Eq. (5) uses annual horizons. Please state where the monthly values come from and whether they are part of Aqueduct 4.0 or another dataset.","section":"§3.1, Figure 4"},{"comment":"The caption states that the star symbol 'refers to the location with lowest value' but does not specify which series (on-site WUE, off-site WUE, total WUE, or water stress); please clarify.","section":"Figure 2 caption"},{"comment":"The phrase 'first general framework' should be qualified, since prior work [17,18,21,28,40] already integrates water stress into scheduling and siting decisions; the novelty claim could be narrowed to the specific combination of basin-level mapping, temporal discounting, and a unified AWI metric.","section":"§1 and Abstract"},{"comment":"The site labels in Figure 7(a) run together (e.g., 'VA2OH3'), making them difficult to read; please add separators or a legend, and define the site abbreviations (NV2, OH2, OH3, VA2, VA3) in the caption or text.","section":"§4.2.1, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a workshop-style contribution with a transparent metric and useful case studies, but the off-site basin-mapping issue and the power-capacity proxy need to be addressed before the framework can support the paper's spatial-accuracy claims. I would not reject the paper, but the revisions are substantial enough to require a second round of review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SCARF is a clear, useful step for sustainable computing: it turns water volume into stress-weighted impact with a transparent formula, folds in future water stress via a discount rate, and ships code plus three case studies. The genuinely new component is the temporal aggregation; spatial weighting by water stress already exists in the water-aware scheduling work they cite (e.g., Islam et al., WaterWise), so 'first general framework' overstates the novelty. Still, the combined spatial-temporal packaging across LLM serving, datacenters, and fabs is worth having.\n\nThe metric and code are solid enough to use with eyes open. The LLM serving measurements are real measured power data; the Google and Intel case studies use public reporting and show how rankings shift with discount rate. That sensitivity analysis is good practice.\n\nThe load-bearing caveat is the off-site water attribution. Eq (6) multiplies W_on + W_off by the facility's basin WSF, but W_off is consumed at the power plant, which can sit in a different watershed. The paper never states or justifies that assumption, and all three case studies inherit it. In an interconnected grid this is not a small parameter error; it can over- or under-state AWI by the WSF ratio between basins, and Figure 4 shows those ratios are large. The framework survives because the fix is straightforward: use the generation basin's WSF for W_off, or disclose the approximation and bound its effect. But as written, the 'where' claim in the abstract is not fully delivered for Scope 2 water.\n\nTwo smaller soft spots. The Google power capacity proxy (max reported capacity within 100 miles) is crude, and the paper gives no error bars or sensitivity around it; the ranking takeaways probably hold but the absolute AWI numbers are rough. And the novelty claim should be narrowed: refs [17,18,21,28] already weight water use by location stress; what is new is the time-discounted aggregation, not the spatial weighting itself.\n\nWho benefits: system operators and sustainability teams wanting a single number for siting or scheduling trade-offs, and researchers comparing water impact across layers. It is a workshop-grade but useful contribution; with the off-site fix and a revised novelty claim it becomes a respectable full paper. I would send it to peer review rather than desk reject: the code ships, the metric is transparent, and the flaw is correctable. My own verdict would be 'conditional accept' pending the Scope 2 fix.","headline":"Temporal water-stress discounting is a real increment, but the off-site attribution and the 'first framework' claim need fixing before the numbers mean what the abstract says.","tokens_in":11710,"tokens_out":2267,"would_cite":true,"duration_ms":24437,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Computing's water impact is not its consumption volume: a new metric multiplies consumption by local, time-varying water stress, and case studies show order-of-magnitude swings by location and season.","keywords":["water stress","adjusted water impact","SCARF","sustainable computing","datacenter water consumption","LLM serving","semiconductor fabrication","water footprint"],"falsifier":"Take a datacenter that buys electricity from a grid whose power plants lie in a low-stress basin, and compute AWI two ways: with the datacenter's basin stress, as SCARF does, and with the plants' actual basin stress; if the rankings of two candidate sites flip, the single-basin attribution is the point of failure.","tokens_in":10764,"feed_emoji":"💧","tokens_out":7931,"duration_ms":80871,"temperature":0.7,"pith_summary":"This paper sets out to change how the computing industry measures its water footprint. Rather than counting liters consumed, it argues that the same liter is much more damaging in a stressed watershed than in an abundant one, and that water stress changes over seasons and decades. To capture this, it introduces SCARF, which computes an Adjusted Water Impact (AWI) by multiplying raw on-site and off-site water consumption by a basin-level Water Stress Factor, optionally discounted over future years. Three case studies—LLM serving, datacenters, and semiconductor fabs—show that AWI can differ by orders of magnitude across locations and months, and that rankings by raw consumption do not match rankings by adjusted impact. If the metric becomes standard, siting, scheduling, and procurement decisions could be made on real water burden rather than volume alone.","feed_headline":"Where a workload runs can change its water impact 1,000-fold","feed_subtitle":"SCARF multiplies water use by basin-level stress and season, exposing hidden choices in siting and scheduling.","key_machinery":"The central object is the Adjusted Water Impact identity $AWI = (W_{\\mathrm{on}} + W_{\\mathrm{off}}) \\times WSF_b$, where water stress is the ratio of local water demand to supply. On-site consumption is $W_{\\mathrm{on}} = P \\cdot t \\cdot WUE_{\\mathrm{on}}$, and off-site consumption is $W_{\\mathrm{off}} = P \\cdot t \\cdot PUE \\cdot WUE_{\\mathrm{off}}$, so the framework hinges on two efficiency ratios—water per kWh on site and per kWh of purchased electricity—combined with a basin-level Water Stress Factor. $WSF_b$ is the mechanism that carries the argument: for immediate impact it is simply the basin's current water stress, and for long-term facilities it is a discount-rate-weighted average of projected stress, making the policy choice about the future explicit. Multiplying the two turns a volume metric into a burden metric.","core_discovery":"The paper's central claim is that water impact assessments for computing should weight consumption by where and when it occurs, and that this can be done with a single unified metric. SCARF maps each facility to its hydrological basin, obtains current and projected water stress from a global risk dataset, and forms the Water Stress Factor: current stress for short-term analyses, or a discounted sum of stress in 2030, 2050, and 2080 for long-lived infrastructure. The Adjusted Water Impact is $AWI = (W_{\\mathrm{on}} + W_{\\mathrm{off}}) \\times WSF_b$, where on-site water comes from cooling and operations and off-site water comes from electricity generation. In the case studies, the same LLM served in a high-stress, inefficient location can have over 1,000 times the adjusted impact per request as a low-stress location; Arizona fabs outrank far-larger Oregon consumers in AWI; and changing the discount rate can reverse which datacenter looks sustainable. The paper reads these results as evidence that a stress-weighted metric reveals hidden opportunities for water-sustainable computing that volume-only accounting misses.","pith_inferences":["Beyond the paper: the same AWI construction could be applied to other spatially variable burdens—such as watershed nutrient loading or local air pollution—wherever a basin-level stress factor exists.","Beyond the paper: because off-site consumption is assigned the datacenter's basin stress, extending SCARF to use the power plant's actual watershed could change rankings in regions that import electricity across basins; this is a direct test of the spatial weighting.","Beyond the paper: with sub-monthly water-stress data, the temporal axis could be pushed from seasons to hours, allowing water-aware job shifting inside a single day for LLM serving.","Beyond the paper: AWI could be paired with carbon accounting to expose trade-offs, since a low-carbon site in a stressed watershed may score worse on water than a higher-carbon site elsewhere."],"forward_implications":["Datacenter and LLM-serving sustainability comparisons should report AWI, not raw water volume, because rankings by the two measures differ.","Workload schedulers can shift inference jobs across months to lower water impact without reducing consumption, since water stress varies seasonally.","For long-lived facilities, the discount rate is a policy parameter: a site can look better or worse depending on how much future stress is valued, so sustainability claims should state the discount rate.","Siting decisions for fabs and datacenters should weigh basin stress alongside efficiency, since high efficiency in a stressed basin can still carry a larger burden than moderate consumption in a wet basin."],"supporting_citations":[{"why":"provides the prior water-consumption methodology and Scope 1/Scope 2 accounting that SCARF builds on.","marker":"[27]"},{"why":"supplies the basin-level water stress values and future projections that define the Water Stress Factor.","marker":"[23]"},{"why":"provides the watershed mapping API used to locate each facility in its hydrological basin.","marker":"[14]"},{"why":"gives location-specific off-site water intensity values used in the LLM serving case study.","marker":"[41]"},{"why":"is the LLM serving platform used to collect power and latency data.","marker":"[24]"},{"why":"reports on-site WUE and PUE for Microsoft datacenters used in the LLM case study.","marker":"[31]"},{"why":"reports Google datacenter site-level water consumption and PUE.","marker":"[12]"},{"why":"reports Intel fab water consumption used in the semiconductor case study.","marker":"[16]"},{"why":"is the source of the 3% discount rate for long-term water stress aggregation.","marker":"[1]"}],"fun_headline_variants":["Water impact of computing depends on time and place","New metric weighs computing water use by regional stress","SCARF shows water impact varies 1000x by location","Stress-aware water metric for sustainable computing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that water consumed off-site to generate electricity feels the same water stress as the datacenter's own watershed, even if the power plant sits in a different basin; the case studies all rely on this.","fun_headline_variants_meta":{"raw":{"variants":["Water impact of computing depends on time and place","New metric weighs computing water use by regional stress","SCARF shows water impact varies 1000x by location","Stress-aware water metric for sustainable computing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000154,"raw_usage":{"total_tokens":1192,"prompt_tokens":907,"completion_tokens":285,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":224}},"tokens_in":523,"tokens_out":285,"duration_ms":3714,"temperature":1.0,"reasoning_tokens":224,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:58:19.820050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a datacenter that buys electricity from a grid whose power plants lie in a low-stress basin, and compute AWI two ways: with the datacenter's basin stress, as SCARF does, and with the plants' actual basin stress; if the rankings of two candidate sites flip, the single-basin attribution is the point of failure.","supporting_citations":[{"cited_title":"Islam, and Shaolei Ren","cited_arxiv_id":null,"evidence_quote":"provides the prior water-consumption methodology and Scope 1/Scope 2 accounting that SCARF builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the basin-level water stress values and future projections that define the Water Stress Factor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the watershed mapping API used to locate each facility in its hydrological basin."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives location-specific off-site water intensity values used in the LLM serving case study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the LLM serving platform used to collect power and latency data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"reports on-site WUE and PUE for Microsoft datacenters used in the LLM case study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"reports Google datacenter site-level water consumption and PUE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"reports Intel fab water consumption used in the semiconductor case study."},{"cited_title":"Water Resources Development Act of 1974","cited_arxiv_id":null,"evidence_quote":"is the source of the 3% discount rate for long-term water stress aggregation."}],"review_version":1}