Utah's Data Center Pipeline and Watershed Water Risk

The problem. Water scarcity is one of the defining resource challenges of the American West. Utah, an arid state that depends heavily on snowpack-fed rivers and reservoirs, is particularly vulnerable. At the same time, Utah has emerged as a rapidly growing technology and energy hub: tax incentives, available land, and access to power have attracted a pipeline of over 23 gigawatts of proposed data center capacity, with a single project (Stratos in Box Elder County) targeting 9 GW. Nuclear small modular reactors are also being revisited as a low-carbon baseload option, with proposed sites in the region. Both data centers and nuclear power plants are significant industrial water consumers: nuclear plants use water for steam generation and cooling, while data centers rely on evaporative cooling that can consume millions of gallons per day at scale.

The question this project asks: to what extent does thermoelectric water consumption from nuclear power plants and large industrial facilities such as data centers reduce effective water availability in Utah watersheds, and can we predict the impact of future growth in that consumption?

Why it matters. Utah's pipeline of proposed data centers is roughly 25 times the state's current operational data center capacity. The largest proposed sites — Sugarville (10,050 MW across three facilities), Headwaters Hansel Valley Wash (9,000 MW Stratos), and Whiskey Creek (4,000 MW HPC Campus) — all sit in endorheic Great Basin watersheds with no downstream outlet. Every unit of consumptive water loss in these basins is permanent. Yet only 8 of the 17 catalogued facilities have publicly reported water consumption figures, leaving water managers, state regulators, and downstream communities without a defensible basis for sustainable siting decisions. A quantitative model that links thermoelectric demand to watershed-level water availability fills exactly this gap.

EDA. The watershed panel covers 25,580 observations (2,558 watersheds × 10 years). The distribution of net availability is sharply skewed: most watershed-years are slightly water-positive, but a long left tail extends well below zero, reflecting the structural water deficit in arid central and western Utah. 8.9% of all observations are in deficit even before the proposed data center pipeline is operational.

The statewide hydrology time series shows that consumption is essentially flat across the decade (red squares hover near 0.5–0.6 mm/mo) while natural availability varies by a factor of five between wet years (2011 peak ≈ 8.5 mm/mo) and dry years (2018 trough ≈ 1.7 mm/mo). Net availability therefore tracks natural availability almost perfectly at the statewide aggregate scale. The thermoelectric effect — small in absolute terms — only becomes detectable when the watershed-level fixed-effects estimator isolates within-watershed year-to-year variation.

A scatter of natural availability against net availability makes the consumptive depression visible at the observation level. The non-thermoelectric majority of watershed-years cluster tightly along the y = x reference, indicating that net availability simply equals natural availability where human consumption is small. The thermoelectric-active watershed-years — roughly 76 observations corresponding to the three operating coal-plant watersheds in Emery County — deviate downward from this baseline. This is the empirical signal the OLS model extracts as a coefficient.

Data preparation. Utah's electricity generation mix has changed substantially over the 2000–2025 horizon. Coal generation has fallen from over 80% of statewide net generation to under 60%; natural gas has roughly tripled; and renewable generation (solar, wind, geothermal) has grown from negligible to a measurable share. The model's fossil_fraction feature captures this composition shift, and the cluster-robust OLS coefficient on fossil_fraction is −0.44 mm/mo (95% CI [−0.59, −0.28], p < 0.001) — statewide shifts toward fossil generation depress net availability.

The base panel was constructed by merging the USGS National Water Availability Assessment outputs (one row per HUC12 × year, fields including availab_mm/mo and consum_mm/mo) with the USGS thermoelectric water-use model (fields tewdftot_mgd for withdrawal and tecuftot_mgd for consumption). Thermoelectric records were available only for watersheds containing active facilities; missing rows were filled with zero. Watershed area in km² was joined from the HUC12 boundary GeoPackage downloaded via the pygeohydro WBD client.

The UofU Energy Landscape dataset provided statewide-annual features that vary in time but are constant across watersheds within a year: renewable_fraction, fossil_fraction, industrial_gwh, and industrial_yoy_growth. These were merged onto the watershed panel by year. The PacifiCorp IRP Table 7.4 supplied design-specification water intensities (Gal/MWh) used to (a) validate the UofU × intensity estimate of statewide thermoelectric consumption against USGS actuals and (b) construct projection scenarios for future technology mixes (nuclear SMR at 720 Gal/MWh, natural gas CCCT at 210, etc.).

Unit normalization. The target variable is net_avail_mm_mo = availab_mm/mo − consum_mm/mo, expressed in mm/month depth. The original USGS thermoelectric variables tecuftot_mgd and tewdftot_mgd are flow rates in million gallons per day — incommensurable with the target depth unit without watershed area. This unit mismatch is one of the most consequential structural decisions in the analysis. The transformation is:

\(thermo\_consum\_mm\_mo = tecuftot\_mgd × (365.25 / 12) × 3.785 × 10^{-3} × 1000 / areasqkm ≈ tecuftot\_mgd × 115.22 / areasqkm\)

The 10⁶ gal/MG and 10⁶ m²/km² factors cancel; the result is a depth-per-month rate directly comparable to the target. For Utah's active thermoelectric watersheds (median area 85.8 km²), the resulting depths range from 0.54 to 3.20 mm/mo — small relative to natural availability fluctuations, which motivates the within-watershed fixed-effects estimator used in model building.

A correlation heatmap of the demeaned features confirms no severe multicollinearity remains after the fixed-effects transformation. The high correlation between availab_mm_mo and net_avail_mm_mo (r ≈ 1.0) is expected and benign: it reflects the dominant physical relationship that net availability tracks natural availability when consumption is held constant. The strongest inter-feature correlation is between year and fossil_fraction (r = −0.73) — the secular decline in coal that defines the 2010s — well below the |r| > 0.9 multicollinearity threshold. The model's job is to extract the small thermoelectric departure from the availability baseline.

Temporal split and demeaning. Because the panel includes a temporal dimension, random k-fold splitting would leak future information into training (a model could see 2019 in training and 2017 in test, predicting the past from the future). The split was therefore made on time: 2010–2017 to training (20,464 rows) and 2018–2019 to test (5,116 rows). The 2018 test year is the driest in the dataset and 2019 is a wet rebound, so the test set spans contrasting hydrological conditions. Continuous features were standardized using a StandardScaler fitted on the training set only — transform-only on test.

The panel of 2,558 watersheds × 10 years contains substantial unobserved time-invariant heterogeneity. Geology, elevation, slope, land cover, and existing infrastructure differ across watersheds in ways the available features cannot fully describe; some of these time-invariant attributes are also correlated with thermoelectric siting decisions (coal plants were sited near coal seams and rivers; data centers cluster near power infrastructure). Within-watershed demeaning — subtracting each watershed's training-set mean from every observation of that watershed — removes time-invariant confounders by construction. The remaining identification is from year-to-year variation within each watershed. After demeaning, the coefficient on thermo_consum_mm_mo measures the within-watershed effect of additional thermoelectric consumption: precisely the estimand of interest for projecting impacts of new facilities.

Models trained. Three models on the watershed fixed-effects demeaned panel:

• Linear Regression (OLS with watershed fixed effects). Interpretable; the thermo_consum_mm_mo coefficient is a direct within-watershed estimate.
• Ridge Regression with leave-one-out CV alpha selection (α = 0.1). L2-regularized stability check on the sparse thermoelectric signal.
• Random Forest (300 trees, max_depth 10, min_samples_leaf 100). Captures any non-linear threshold effects.

Cross-validation strategy. GroupKFold with 5 folds and huc12_id as the group key was used for model selection. Each validation fold contains watersheds entirely absent from the training fold. This directly tests the deployment scenario: predicting net availability impacts in proposed-facility watersheds that have no thermoelectric history. Standard random k-fold would yield optimistic CV metrics because the model would see watersheds during training. GroupKFold CV RMSE was 0.221 ± 0.035 mm/mo for both OLS and Ridge.

Test-set evaluation. OLS and Ridge produced essentially identical test-set performance (RMSE = 0.326 mm/mo, R² = 0.994). The slight scatter visible at the bottom-left of the predicted-vs-actual test panel corresponds to 2018 drought observations, where the model conservatively underestimates how dry the year was — a recognized limitation discussed below. Random Forest performed substantially worse (test RMSE = 2.46, R² = 0.67). With only 36 thermoelectric watershed-years in training (out of 20,456), the forest cannot construct splits on thermo_consum_mm_mo at the min_samples_leaf = 100 threshold required to avoid overfitting. The linear models are clearly preferred for this problem.

Statistical inference. OLS standard errors assume independent residuals, but observations within the same watershed share hydrology, geology, and infrastructure. Cluster-robust standard errors clustered by huc12_id correct for this. The cluster-robust 95% confidence interval for the key coefficient is [−0.356, −0.241]. An independent pairs-bootstrap with 500 resamples gave [−0.358, −0.260], confirming the cluster-robust interval.

Cluster-robust 95% CIs: fossil_fraction −0.436 [−0.594, −0.277], thermo_consum_mm_mo −0.299 [−0.356, −0.241], year −0.015 [−0.018, −0.012], industrial_gwh +0.00011 [9.4e-5, 1.3e-4], availab_mm_mo +1.006 [1.004, 1.008]. All five coefficients are unambiguously distinct from zero. The two positive coefficients are not plotted because of incompatible scale: availab_mm_mo ≈ 1.0 reflects the physical baseline that net availability tracks natural availability when consumption is constant, and industrial_gwh is on the order of 10⁻⁴. The thermoelectric effect is the small departure from the natural-availability baseline.

Quantile regression at τ = 0.10 (the dry-year tail) yields a thermo_consum_mm_mo coefficient of −0.328 versus −0.299 at the conditional mean — a 10% amplification of thermoelectric stress in dry years. The fossil_fraction coefficient amplifies 3.6× from −0.44 at the mean to −1.58 at τ = 0.10. Both findings confirm that thermoelectric and energy-mix impacts are disproportionately concentrated in the years when water is already scarcest.

Robustness check. A leave-driest-training-year-out check refits the model after removing 2012 (the lowest-availability training year, mean 2.68 mm/mo). The thermo_consum_mm_mo coefficient changes from −0.299 to −0.297 — a +0.4% change well within the 95% CI of the full estimate. The thermoelectric coefficient does not depend on a single drought year.

Residual diagnostics. Test residuals have mean 0.006 and standard deviation 0.326, indicating overall unbiased prediction. The Q-Q plot shows a heavy left tail (skew = −6.80), consistent with the model under-estimating the severity of the 2018 drought. This is a known limitation: no training year matches 2018's aridity, so the model treats it as an extreme extrapolation. For projection purposes, this means the model is conservative on drought risk — the actual impact of data center demand in a drought year is likely to be modestly larger than the projection implies.

Projection methodology. For projection scenarios, the unscaled OLS coefficient is applied directly. Each proposed facility's capacity (MW) is converted to cooling water demand (MGD) using the project's standard helper dc_water_mgd(mw, wue) with a default mid-range Water Usage Effectiveness of 1.5 L/kWh (low 0.5 and high 2.0 also computed). The MGD is then converted to mm/month depth using the watershed area, multiplied by the OLS coefficient (−0.299), and the cluster-robust 95% CI bounds [−0.356, −0.241] are propagated through the same multiplication to give an uncertainty interval on the projected impact.

For watersheds hosting multiple proposed facilities, demand is summed before conversion. This produces the cumulative per-watershed projection below.

Results: projecting future demand. Today, consumption is broadly distributed across Utah's populated and agricultural areas, with modest concentrations along the Wasatch Front and in Emery County (existing coal-plant watersheds). After the proposed pipeline is operational, three Great Basin watersheds — Sugarville, Headwaters Hansel Valley Wash, and Whiskey Creek — emerge as new extreme hotspots, each absorbing an order of magnitude more cooling water than any current site. The map below toggles between the two states. Green-bordered dots are existing power plants, blue are existing data centers, red are proposed power plants and data centers; marker radius scales with MW capacity. Click a watershed for facility detail; the fullscreen button opens a zoom/pan view.

Six watersheds host planned facilities in the Cleanview inventory. Aggregating facility-level cooling demand within each watershed and propagating through the OLS coefficient yields per-watershed projected reductions in net availability. The three large Great Basin sites dominate; the other three watersheds (Bingham Creek, Summit Creek, Big Dry Creek) host smaller expansions with minimal projected impact.

Cluster-robust 95% CIs from the OLS coefficient: Sugarville [−41.67, −28.22], Hansel Valley [−19.17, −12.98], Whiskey Creek [−10.49, −7.10], Bingham Creek [−0.573, −0.388], Summit Creek [−0.075, −0.051], Big Dry Creek [−0.027, −0.018].

Statewide aggregate. The total planned pipeline is 23,222 MW across the six watersheds, corresponding to a cumulative cooling-water demand of approximately 177 MGD at mid-WUE. This is roughly four times the entire current statewide thermoelectric consumption baseline of ~43 MGD. The mid-WUE worst-case watershed is Sugarville, projected to lose 34.9 mm/month of net availability; the historical average for that watershed is already −15.6 mm/month, so the combined post-buildout average would be approximately −50 mm/month. Headwaters Hansel Valley Wash (Stratos) is currently barely water-positive at +0.5 mm/month average, but the proposed 9,000 MW facility would drive it to approximately −16 mm/month — a sign change in the basin's water balance.

What the model tells us. The cluster-robust within-watershed coefficient on thermoelectric consumption is −0.299 mm/month per mm/month, with a 95% confidence interval that excludes zero by a comfortable margin ([−0.356, −0.241]). The pairs-bootstrap confirmation ([−0.358, −0.260]) provides an independent check via a different statistical mechanism. The robustness check (leave-driest-year-out) confirms the coefficient is not driven by extreme-drought training observations. Quantile regression at the drought tail shows the effect amplifies modestly (~10%) in dry years and the energy-mix effect amplifies dramatically (~3.6×). The model generalizes to never-thermoelectric watersheds (OOS test RMSE = 0.32 mm/month, R² = 0.994 across 5,098 observations) — the exact deployment scenario for projection.

When the coefficient is applied to the 23,222 MW Utah data center pipeline at mid-WUE, three Great Basin watersheds emerge as severe stress hotspots: Sugarville (−34.9 mm/mo), Headwaters Hansel Valley Wash / Stratos (−16.1 mm/mo), and Whiskey Creek (−8.8 mm/mo). All three are endorheic with no downstream outlet; consumptive loss in these basins is permanent. The remaining three planned-pipeline watersheds have minor impact, but the headline finding is unambiguous: the proposed pipeline produces large, statistically significant water stress concentrated in a small number of closed-basin watersheds that are not currently in deficit.

Is the model ready to be deployed? The model is ready for use as a scenario planning tool for water resource managers and as input to facility-siting reviews. Its strengths are clear: it is built on the authoritative USGS NWAA dataset, uses correct unit normalization and a watershed-fixed-effects estimator that identifies a causally interpretable within-watershed effect, reports uncertainty via two independent statistical methods, has been stress-tested against a leave-one-year-out robustness check, generalizes to unseen watersheds, and produces a directly usable mm/mo-per-mm/mo coefficient that can be applied to any new facility for which a cooling-water demand can be estimated.

The model should not, however, be the sole basis for facility approval or denial. The training set contains only 36 thermoelectric watershed-years, all at existing Utah coal plants. Extrapolating to data-center evaporative cooling assumes the same physical mechanism applies, which is reasonable but not directly validated. Sub-watershed dynamics (groundwater recharge, return flows, intra-month variability) are not resolved by the HUC12 × year panel. The model is therefore best understood as a defensible baseline for understanding the order of magnitude of impact and the relative ranking of risk across proposed sites — not as a substitute for site-specific hydrological studies.

Recommendations.

• For Utah water resource managers: prioritize watershed-level water-availability studies for Sugarville (HUC12 160300050805), Headwaters Hansel Valley Wash (160203090501), and Whiskey Creek (160300051502) before approving the proposed facilities in those basins. The combined impact of multiple facilities in Sugarville is the single largest projected risk in the pipeline.
• For state regulators: require facility-level water consumption measurement and public reporting for all proposed and operating data centers. The current 9-of-17 reporting gap leaves regulators without a basis for calibrating projection models or enforcing water-use conditions.
• For technology companies conducting water impact assessments: use the model's mm/mo-per-mm/mo coefficient with cluster-robust 95% CI as a defensible projection framework. Site-specific WUE values should replace the default 1.5 L/kWh assumption where measured cooling-system performance data is available.
• For environmental planners: recognize that the endorheic nature of the high-impact Great Basin watersheds means the consumptive losses are permanent and irreversible. Mitigation through return-flow recovery is not available in closed basins.

Limitations and additional data that would improve accuracy. The model produces statistically robust within-watershed estimates with a tightly bounded 95% CI, and projections track expected physical magnitudes. Several data gaps still constrain confidence in the projection numbers:

1. Facility-level measured water consumption. WRA (2025) reports actual water use for only 8 of 17 catalogued operating data centers. Verified consumption data for the remaining 9, plus mandatory reporting for proposed facilities, would replace the WUE-based intensity assumption (currently 0.5–2.0 L/kWh range, midpoint 1.5) with directly measured values.
2. Post-2019 USGS NWAA observations. The watershed model panel ends in 2019. Data center industrial load roughly doubled in Utah between 2019 and 2025; the most recent thermoelectric responses to that growth are not in the training set. Re-extracting NWAA 2020–2025 outputs when the USGS releases them would strengthen both the historical coefficient estimate and the calibration of future projections.
3. Groundwater and aquifer recharge data. The target net_avail_mm_mo captures surface water; the Great Basin watersheds (Sugarville, Hansel Valley, Whiskey Creek) where the largest proposed facilities will operate are endorheic and partly groundwater-dependent. USGS National Water Information System (NWIS) well data per HUC12 would allow extending the analysis to total water balance.
4. Climate projections for natural availability. Future scenarios in this notebook hold availab_mm_mo at its historical mean. Coupling to downscaled IPCC CMIP6 projections for the Utah/Great Basin region would quantify how climate-driven declines in availability compound with data-center demand growth.
5. Return-flow accounting. OLS implicitly assumes all thermo_consum_mm_mo is fully consumptive. Some cooling-tower configurations have measurable blowdown that returns to nearby surface water. Facility-level return-flow audits would refine the within-watershed coefficient by separating consumption from net withdrawal.

These gaps do not alter the qualitative conclusion (proposed pipeline produces large, statistically unambiguous reductions in net availability in three Great Basin watersheds). They affect the precision of the absolute numbers, not the direction or significance.

Data sources.

• United States Geological Survey (USGS): Thermoelectric Power Water-Use Model; National Water Availability Assessment Outputs (CONUS 2025)
• Western Resource Advocates (WRA): Data Center Impacts in the West: Policy Solutions for Water and Energy Use
• Salt Lake Tribune: Can Utah become a data center hub without draining its water supply?
• Stanford University: Nuclear Reactor Water Usage and the Implications of Limited Water Availability
• University of Utah: Utah's 2025 Energy Landscape and Outlook for 2026 and Beyond
• PacifiCorp: 2025 Integrated Resource Plan: Volume 1
• Utah Clean Energy: Estimated Emissions and Water Consumption from the Proposed Stratos Data Center
• Deseret News: Data centers on the way to Utah, but how fast and at what cost?
• ABC4 News: Tracking where AI data centers are being built in Utah: Interactive map