Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States
Authors: Alex C. Newkirk, Daniel L. Gerber, Erica R. H. Fuchs, Alex Hubbard, M. Granger Morgan, Arman Shehabi, Sarah J. Smith, Constantine Samaras
Organizations: Department of Engineering and Public Policy, Carnegie Mellon University, Pittsburgh, PA · University of California, Berkeley, Berkeley, CA · Department of Civil and Environmental Engineering, Carnegie Mellon University, Pittsburgh, PA · Wilton E. Scott Institute for Energy Innovation, Carnegie Mellon University, Pittsburgh, PA
The rapid growth of artificial intelligence is driving demand for new data centers at large scale and on short deployment timelines. To power them, firms connect to the commercial grid or contract generation directly, but interconnection and gas-turbine delays now incentivize on-site generation. We compare grid-connected and off-grid data center energy infrastructure (AC- and DC-coupled solar-plus-storage, islanded natural gas, and grid interconnection) across the continental U.S. Using an open-source technoeconomic framework built on empirically calibrated AI training loads, climate-driven cooling, stage-by-stage power conversion, and physics-based battery degradation, we model designs providing equivalent energy services. Our central contribution is a speed-adjusted cost of energy that values the compute unlocked by faster construction, making options with different deployment timelines directly comparable on one physically consistent basis. We find DC coupling reduces lifetime energy costs by 18% ($0.09/ kWh) on average versus AC counterparts by eliminating inverters and improving efficiency. Though rarely cheapest on unadjusted energy cost, DC-coupled microgrids' faster deployment makes them the lowest-cost modeled option in 85% of U.S. locations once foregone compute is valued, or 55% under accelerated natural gas timelines. For AI infrastructure, deployment speed rather than unit energy cost determines the economically preferred path to power.
Figures & tables
Figure 1: Analytical framework. For a given quantity of chips, workload (in this research empirically calibrated to AI training) and location (climate, solar resource, gas price, grid rate, and interconnection queue), we build an hourly facility load combining IT demand with climate-driven cooling (PUE). We then size and dispatch three islanded architectures, AC solar+storage, DC solar+storage, and islanded natural gas with diesel backup, and compare them against grid interconnection. Each deliver equivalent energy service at equivalent reliability through an explicitly modeled power-conversion path. Each option is assessed on two axes: its unadjusted levelized cost of energy (/kWh)andthevalueofcomputeforgoneduringconstruction(setbydeploymentspeedandchip−hourvalue).Theirsum,aspeed−adjustedcost(\mathrm{LCOE_{adj}}$ ), determines the preferred option, which varies with location, scale, technology, and the value of time.
Figure 2: Candidate MV power flow models within the facility fence-line with modeled component efficiencies. Power flows are idealized, without batching or correspondence to any particular facility scale. Grid power and islanded natural gas are assumed to have equivalent power systems architecture inside the facility fenceline. Note that while not shown on the plot, the diesel backup power for the grid comparison and islanded natural gas cases is modeled as entering the facility through the same main power distribution path. Fault protection components, monitoring components, or supercapacitor banks are not shown. BESS and UPS components are shown abstracted from the facility layout and composition, and may in practice be centralized for the facility, distributed at the sidecar, rack, or node level, or some combination. Microgrids that use BESS for backup substitute these assets for the UPS via a small share of capacity embedded in the LV data hall (see Supplementary Information). AC power systems distribute power at 480 V bus voltage, with embedded PSU or motor drive power electronics converting to the voltages required by IT and HVAC loads. Bidirectional grid-forming inverters convert the DC outputs to AC at the specified bus voltage and frequency. DC power systems operate at a ±400 V DC bus voltage, with DC-DC converters stepping up to bus voltage and PSUs and ESCs stepping down to deliver the required load voltages.
Figure 3: Cost advantage of DC-coupled systems over AC counterparts for 100,000 GPUs at 99% uptime across all modeled continental U.S. locations. Skew in the absolute distribution reflects variation in sensitivity to solar-to-load and battery-to-load path efficiency by location. Scale of DC advantage generally increases in locations with lower quality solar resources as marginal path efficiencies provide returns for highly overbuilt systems. In regions with substantial and consistent solar resources, both AC and all-DC microgrids can scope more aggressively, shrinking but not eliminating the DC cost advantage. (A) Percentage cost advantage distribution. Mean DC advantage is 18% of total cost of ownership for a given location. (B) Absolute cost advantage distribution. Mean DC advantage is $0.09 / kWh.
Figure 4: LCOE of the minimum cost natural gas configurations against different facility design loads at 99% uptime in eastern Colorado; natural gas price of 7.70/MMBTU.PointscorrespondtotheLCOEforagivendesign,withcolorindicatingturbineclass(coralforaeroderivative,tealforF−class)andshapeindicatingsinglecycle(SC)orcombinedcycle(CC)configuration.Step−sizesbetweenpointsareincrementsof2MWofdesignloadupfrom1−20MW,4MWofdesignloadfrom20−64MW,and8MWofdesignloadfrom64MWto320MW.DashedlinesindicatetheLCOEofmedium−voltageDC(green)andAC(yellow)PV+BESSsystemsatthesamelocation.Naturalgasplantsexhibitsignificantscaleeconomies,droppingby$0.15/kWhovertherangeofsystemscales.PV+BESSsystemsshowedminimalunitcostvariationwithnameplatecapacity(\pm$0.002$ / kWh), we plot these costs as idealized average values without this noise for visual clarity. Jaggedness on the natural gas curve demonstrates transitions between turbine models and classes, and plant configurations as data center facility scale changes.
Figure 5: Annual utilization requirements determine optimal power architecture across diverse geographic and market conditions. Each panel shows levelized cost of electricity versus annual share of energy-service-hours provided by the islanded microgrid for a 100,000 GPU datacenter, comparing MV DC-coupled solar (blue), natural gas with diesel backup (orange), and the grid comparison case (gray dashed). The horizontal axis reads two ways: the annual uptime of a fully islanded facility, with the remaining hours unserved; or the fraction of hours an otherwise grid-tied facility runs on its microgrid. Solar architectures at low utilization are constrained by power capacity rather than energy, meaning costs initially decline as utilization increases. Natural gas costs decline monotonically as fixed capital is amortized over more operating hours. Facility design loads vary by location due to location-specific cooling system performance. ( A ) Southeastern California: Exceptional solar resources and high electricity prices (0.185/kWh)favorDCsolaracrossallutilizationlevels.(B)Houston,Texas:NaturalgasbecomeslessexpensivethanDCsolarabove770.055/kWh). ( C ) Northern Virginia: Data center alley. ( D ) Eastern Washington State: Solar resource inconsistency makes high reliability with PV+BESS systems expensive. The hitch in natural gas at higher utilization reflects transition to an alternative plant configuration at higher uptimes.
Figure 6: Minimum speed-premium LCOE for 100,000 GPU facilities across the continental United States. Color indicates LCOE magnitude; hatching shows lowest- LCOEadj technology. Maps vary by natural gas deployment speed assumptions, with the top panel (A) showing the baseline case of 4 year combined turbine delivery and plant construction time, and the bottom panel (B) showing the accelerated 3 year case. (A) Baseline construction timelines: DC solar is the lowest LCOEadj architecture in 85% of locations due to rapid deployment despite higher unadjusted energy costs. Grid interconnection remains preferred in regions (15% of locations) combining poor solar resources, low industrial electricity rates, and abbreviated approval timelines. (B) Accelerated 36-month natural gas timelines: DC solar (solid) is preferred in ∼ 55% of locations. Natural gas (dots) is preferred in ∼ 35% of locations, forming coherent regional deployments in solar-poor areas (upper Midwest, Appalachian Basin, Pacific Northwest). Grid interconnection (diagonal hatching) remains optimal in ∼ 9% of locations. Even under this aggressive scenario, DC solar microgrids remain the most widely preferred option.
Figure 7: Value of a chip-hour at which firms would be indifferent between DC-coupled microgrids and grid interconnection. Holding interconnection timelines at baseline (state-specific, 2.5-5 years). Color indicates indifference utility in /GPU−hour;whitecontourmarksthebaselineassumption(2.40/hr). Blue regions show where DC solar dominates even at low GPU values due to favorable solar resources and/or slow grid interconnection. The median indifference utility is 1.61/hr.Orange/redregions(concentratedinNewEnglandandpartsofTexas)showwheregridremainscompetitiveevenathighGPUopportunitycostsduetocheapelectricityandfastinterconnection.MostofcontinentalU.S.liesbelow2.40/hr (blue spectrum), indicating DC microgrids are preferable to grid interconnection when valuing deployment speed even at lower GPU value.
Figure S1 : Hourly power dispatch for the all-DC microgrid over a representative week (MV-coupled, Albuquerque, 100,000 GPUs). (A) PV generation and its disposition: power delivered directly to load (amber), used to charge the battery (blue), and curtailed (vermilion), with the facility load drawn for scale. (B) How the IT-plus-cooling load is met each hour, by direct PV (amber) and battery discharge (green); the solid line is total load and the dotted line the IT component, the difference being cooling. The eight-fold gap between peak PV and facility load, the heavy midday curtailment, and the full overnight reliance on storage are characteristic of an energy- rather than power-constrained islanded design, and underlie the low-stress battery duty cycle documented in the degradation subsection.
Figure S2 : Distribution of BESS state of energy over the hours in simulation year 0 for a candidate DC coupled design. The modeled design is in southern California (34.47 N, -116.16 W), would supply 50,000 GPUs (facility design load of 65.3 MW) on a MVDC architecture though 447 MW solar PV and 264 MW of BESS capacity. Note the strong skew towards high states of energy, as storage systems are scoped for outlier events. In practice, alongside the practically modest discharge rate for steady overnight operations, configurations capable of high uptime exhibit comparatively low-stress cycling patterns.
Figure S3 : Cumulative fade from cycling and calendar effects over modelled lifespan of LFP battery system. The modeled design is in southern California (34.47 N, -116.16 W), would supply 50,000 GPUs (facility design load of 65.3 MW) on a MVDC architecture though 447 MW solar PV and 264 MW of BESS capacity. Cumulative fade is 12% at replacement, with calendar fade contributing 51% and cycle fade contributing 49%. Note that cumulative fade and fade composition are both design specific so will differ system-to-system.
Figure S4 : Out-of-sample validation of the gray-box battery fade surrogate model. The parity plot compares predicted annual fade rates (vertical axis) against ground-truth BLAST-Lite simulations (horizontal axis) for 10 geographic locations not included in the training data. Each point represents one operational year (0-12, color-coded) for one of the validation designs. The tight clustering along the diagonal (dashed line) demonstrates that the surrogate maintains predictive accuracy when extrapolating to new locations.
Figure S5 : LCOE for all-DC microgrids across the continental United States. Baseline LCOE ranges from 0.15/kWhinthemostconsistentlysunnyregionsto1.10/kWh in the Olympic peninsula of Washington. System cost is more sensitive to worst-case cloud cover than latitude driven expected irradiance.
Figure S6 : Comparative LCOE of natural gas and DC Solar microgrid supplying a 24,000 GPU data center across the United States. DC solar achieves or nears cost-parity in high-insolation regions spanning southern California, Arizona, Nevada, New Mexico, and Colorado. In solar poor regions of the pacific northwest and upper midwest, natural gas significantly outperforms DC solar systems. Note the skew in performance geographically: the worst performing locations for solar are disproportionately expensive, while the spread of gas costs between the best and worst tiles is more modest.
Figure S7 : Required grid interconnection timeline for grid and DC-coupled solar microgrids to achieve equal speed-premium LCOE. For a 100,000 GPU datacenter, holding GPU opportunity cost constant at $2.40/hr. Note this plot displays only the absolute interconnection time required for indifference and is not relative to regional interconnection estimates.
Figure S8 : Required natural-gas construction time to match DC-coupled solar on speed-premium LCOE. For each hexagon we solve for the gas-plant construction duration at which the gas system’s speed-premium-adjusted LCOE equals that of the optimized DC-coupled solar design (100,000 GPUs, 99% uptime; GPU opportunity cost 2.40/hr;DCsolarbuiltin2.0years).Whitemarksthe3.0−yearaccelerated−gastarget.Blueregions—2,801of4,317hexes—requiregastobebuiltfasterthanthisacceleratedpacemerelytomatchsolar,andconcentrateinthehigh−insolationSouthwestwheresolarischeapest;orangeregionscanbebuiltatthestandard\sim$ 4-year pace and still win, reflecting solar’s higher cost in the cloudy North. For a facility of this scale NG can match DC solar at some build speed in every hexagon; the map shows how fast it must be built to do so.
Figure S9 : Optimization stability across representative climate zones. DC-coupled, 100,000 GPUs at 99% uptime, medium-voltage spine, 24 independent optimizations (one per seed) per panel. Each point is one optimization, colored by LCOE (/kWh).Interiorsolutionsindicateoutlierrunswithsuboptimalconvergence(>2%$ LCOE penalty). Component sizing variation (solar CV 3.7–10%, battery CV 2.9–10% across these four sites) substantially exceeds LCOE variation (0.3–2.3%), indicating near-flat objective function regions where many configurations achieve nearly identical economic performance. Note the regional variation in optimal design cluster between comparatively battery favoring and compartively PV favoring. Interior solutions represent convergence failure by the optimizer to identify the performant design basin.
Figure S10 : Committed fuel-price elasticity of natural-gas LCOE, by facility scale. Each point is one (city, scale) plant, sized once at its state’s baseline gas price; the plant is then held fixed and its LCOE recomputed at a set of alternative delivered fuel-price levels. Axes show deviations from each plant’s baseline point ( Δ delivered fuel price, Δ LCOE). The common near-linear trend is the simple-cycle heat-rate regime that prevails for the smaller facilities; combined-cycle plants, selected only in high-gas price locations, sit off-trend at large fuel-price deviations, reflecting their lower heat rate and flatter elasticity.
Figure S11 : Distribution of single-year LCOE deviations from the TMY design basis. Pooled across 28 locations (two per DOE climate zone) and the fourteen historical years 2010–2023, for the DC-coupled architecture at 40,000 GPUs and 99% uptime (392 location-years). Each observation is one historical year’s LCOE minus the TMY-based LCOE at the same site. The distribution is right-skewed—most years cluster near the TMY value while a thin tail of adverse years exceeds +\0.2/kWh—pullingthemeanto+$0.0121$ /kWh (dashed line), the basis-wide optimism of TMY.
Figure S12 : Modular Low Voltage PV+BESS Power Flow Architectures. Candidate low voltage power topologies. Power flows are idealized, without batching or correspondence to any particular facility scale. Note that in the AC case, grid forming and power quality services are provided by the inverter. For cases with fully conditioned power, there would be a double conversion storage system embedded between the switchgear and the loads, analogous to the UPS placement in maintext figure 6.
Parameter
Low
High
Δ% Low
Δ% High
($/kWh)
($/kWh)
(%)
(%)
AC-Coupled Solar+Storage
Baseline LCOE: $0.3743/kWh
Discount Rate (3–12%)
0.2685
0.5339
-28.26
+42.61
Solar BOS (×0.8–1.2)
0.3448
0.4039
-7.89
+7.89
Battery Cost Y0 (×0.8–1.2)
0.3581
0.3906
-4.33
+4.33
Table S1 : Base LCOE sensitivity analysis across power system architectures (excluding GPU idling costs). Percent changes are relative to each architecture’s baseline LCOE (shown above result columns). Solar+storage and natural gas results are shown for the 100,000-GPU facility (solar+storage MV-coupled); the per-row percentage sensitivities of solar+storage are nearly scale- and topology-invariant, differing by less than one percentage point at the 24,000-GPU LV-coupled scale. Baseline assumptions: Solar construction = 2.0 yr; NG construction = 4.0 yr; discount rate = 7%; NG price = $7.70/MMBtu.
Parameter
Low
High
Δ% Low
Δ% High
($/kWh)
($/kWh)
(%)
(%)
AC-Coupled Solar+Storage
Baseline LCOEadj : $0.8188/kWh
Discount Rate (3–12%)
0.5548
1.2256
-32.24
+49.69
GPU Idling ($1–4/chip-hr)
0.5595
1.1151
-31.66
+36.19
Construction Time (1.5–3.0 yr)
0.7128
1.0943
-12.95
+33.65
Table S2 : LCOEadj sensitivity analysis. Percent changes are relative to each architecture’s baseline (shown above result columns). All results are shown for the 100,000-GPU facility (solar+storage MV-coupled); solar+storage percentage sensitivities are nearly scale- and topology-invariant. Baseline assumptions: Solar construction = 2.0 yr; NG construction = 4.0 yr; discount rate = 7%; NG price = 7.70/MMBtu;GPUspotprice=2.40/chip-hr.
Parameter (Range)
AC Solar
DC Solar
Natural Gas
Baseline: $0.3743/kWh
Baseline: $0.2880/kWh
Baseline: $0.1729/kWh
Battery Performance
Battery RTE
-\0.0093to+$0.0119$
-\0.0075to+$0.0073$
—
(0.85 – 0.95)
−2.48% to +3.19%
−2.61% to +2.54%
Replacement Year
-\0.0023to+$0.0191$
-\0.0033to+$0.0100$
—
(Year 16 – Year 10)
−0.61% to +5.10%
−1.14% to +3.48%
Table S3 : Technical parameter sensitivity across power system architectures for the 100,000-GPU facility (solar+storage MV-coupled; natural gas F-class) in southeastern Colorado. Values show the LCOE impact range (absolute $/kWh and percentage change) for each parameter variation, computed by re-optimization under the perturbed parameter using common RNG seed. Sensitivity at 24,000 GPUs (aeroderivative) and 500,000 GPUs (H-class) similar in magnitude. The replacement-year row re-optimizes with the degradation schedule re-anchored to the replacement year: a later replacement (year 16) requires modestly more initial storage to hold reliability at the more-degraded pre-replacement point, and battery sizing moves by a few percent across the range. Parameters that do not bind the energy-limited solar design—inverter load ratio, design contingency, and the double-conversion UPS stage return the identical re-optimized design and hence no first-order LCOE effect. Inverter load ratio matters only in the rare hours when clipped DC output coincides with excess bus demand, which the large power overbuild renders negligible.
Table S4 : Cost levers in PV+BESS topology, ordered by magnitude. Penalties relative to base LCOE modeled for an illustrative high-resource site (SE Colorado, 100k GPUs, 99% uptime).
Parameter
Low
High
Δ% Low
Δ% High
($/kWh)
($/kWh)
(%)
(%)
24,000 GPUs
Baseline LCOE : $0.2082/kWh
Natural Gas Price ($3.01–12.75/MMBtu)
0.1549
0.2657
-25.60
+27.57
Discount Rate (3–12%)
0.1790
0.2557
-14.03
+22.78
NG CapEx (×0.8–1.2)
0.1967
0.2198
-5.56
+5.56
Table S5 : Natural Gas baseline LCOE parameter sensitivity for the 24,000 and 500,000 GPU cases Modeled facilities are in southeastern Colorado. Percent changes are relative to each scale’s baseline. Baseline assumptions: Discount rate = 7%; NG price = $7.70/MMBtu.
Parameter
Low
High
Δ% Low
Δ% High
($/kWh)
($/kWh)
(%)
(%)
24,000 GPUs
Baseline LCOEadj : $1.1909/kWh
GPU Idling ($1–4/chip-hr)
0.6177
1.8460
-48.13
+55.01
Discount Rate (3–12%)
0.8018
1.8361
-32.67
+54.18
NG Construction Time (×0.5–1.2)
0.6428
1.5626
-46.02
+31.21
Table S6 : Natural Gas LCOEadj parameter sensitivity for the 24,000 and 500,000 GPU cases. Modeled facilities are in southeastern Colorado. Percent changes are relative to each scale’s baseline. Baseline assumptions: Discount rate = 7%; NG price = $7.70/MMBtu.
Parameter
Value
Source
COMPUTE HARDWARE
GPUs per node
8
DGX H100 standard [ 30 ]
Node power (average)
7.3 kW
[ 115 ]
Node power (maximum)
8.5 kW
[ 115 ]
Design contingency
1.05
5% safety margin
ECONOMIC ASSUMPTIONS
Table S7 : Economic, Hardware, and Solar PV System Parameters
Parameter
Value
Source
BATTERY STORAGE SYSTEM
Capital Costs
Battery cells (Year 0)
$826/kW
[ 88 ]
Battery cells (Year 13)
$508.12/kW
[ 88 ] moderate case
Battery BOS (AC-coupled)
$385.31/kW
[ 89 ]
Battery BOS (DC-coupled)
$292.20/kW
Proportional reduction
Table S8 : Battery Storage System Parameters
Parameter
Value
Source
NATURAL GAS TURBINES
Simple Cycle Reference Plants
Aero SC (211 MW) CapEx
$1,510/kW
EIA [ 93 ]
Aero SC Fixed O&M
$35/kW-yr
[ 93 ]
Aero SC Variable O&M
$8/MWh
[ 93 ]
F-class SC (233 MW) CapEx
$1,319/kW
NLR ATB [ 88 ]
Table S9 : Natural Gas, Diesel, and Gas Turbine Performance Parameters
Turbine Model
Capacity (MW)
Efficiency
Availability
Aeroderivatives (Simple Cycle)
Siemens SGT-A05 KB7S
5.8
0.323
0.92
GE LM2500+G4
35.0
0.390
0.95
Siemens SGT-A45
44.0
0.404
0.94
GE LM6000PC
51.0
0.397
0.94
Siemens SGT-800
62.0
0.411
0.92
Table S10 : Gas Turbine Models Available for System Configuration. Aeroderivatives optimized for simple cycle (SC); F-class suitable for both SC and combined cycle (CC); H-class preferred for CC configurations.
State
ISO/RTO
Queue (yr)
Elec (¢/kWh)
Gas ($/MMBtu)
Region
Connecticut
ISO-NE
2.5
15.20
7.87
New England
Maine
ISO-NE
2.5
12.01
9.47
Massachusetts
ISO-NE
2.5
16.79
12.04
New Hampshire
ISO-NE
2.5
15.39
12.02
Rhode Island
ISO-NE
2.5
16.42
11.57
Vermont
ISO-NE
2.5
11.02
5.16
Table S11 : State-Level Grid and Fuel Parameters (2022 USD). Electricity and gas prices from EIA 2022 industrial rates [ 98 , 148 ] . Interconnection queue data from LBNL 2018–2023 median [ 13 ] .
Sep 8, 2026·Alexander Crosier, Kyle Onghai, Ronnie SircarPower Systems
Formerly Princeton University · ORFE Department, Princeton University · ORFE Department, Bendheim Center for Finance and Andlinger Center for Energy & the Environment, Princeton University
May 30, 2026·Dana Golden, Aruna Balasubramanian, Niranjan BalasubramanianPower Systems
Department of Economics, Stony Brook University, Stony Brook, 11794, New York, USA · Department of Computer Science, Stony Brook University, Stony Brook, 11794, New York, USA