Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States

Alex C. Newkirk, Daniel L. Gerber, Erica R. H. Fuchs, Alex Hubbard, M. Granger Morgan, Arman Shehabi, Sarah J. Smith, Constantine SamarasarXiv 2026 · arXiv

TL;DR

In the race to deploy AI, the "speed-to-power" metric is more critical than the unit cost of electricity. This paper introduces a "speed-adjusted" levelized cost of energy (LCOE) that accounts for the massive opportunity cost of foregone compute during construction delays. The authors find that DC-coupled solar-plus-storage microgrids, despite higher unadjusted costs, are the most economical choice for 85% of U.S. locations because they can be deployed years faster than waiting for a grid connection.

What Problem Does This Paper Address?

The "decade of equilibrium"—where data center efficiency gains offset rising compute demand—has been shattered by the AI boom. Training frontier models requires gigawatt-scale power on aggressive timelines. However, the legacy U.S. electricity grid is struggling to keep up; interconnection queues now exceed five years in many regions, and even natural gas turbines face multi-year backlogs.

AI firms are increasingly looking at off-grid solutions to bypass these bottlenecks. But how do you compare a cheap, slow grid connection with an expensive, fast solar microgrid? This paper addresses the lack of a consistent framework to value deployment speed alongside traditional energy costs, providing a technoeconomic analysis of four primary architectures: AC solar+storage, DC solar+storage, islanded natural gas, and traditional grid interconnection.

Methods

The authors developed an open-source technoeconomic framework that models the entire energy lifecycle of an AI data center. Key components include:

  • Empirical Workload Modeling: Using power traces from NVIDIA H100 clusters to simulate realistic, near-baseload training demands.
  • Climate-Driven Cooling: Calculating hourly Power Usage Effectiveness (PUE) based on local weather data across the continental U.S.
  • Power Flow Analysis: A stage-by-stage model of conversion losses for both AC and DC architectures.
  • Physics-Based Degradation: Integrating the BLAST-Lite framework to model battery aging based on actual charge/discharge cycles and temperature.
  • The "Speed Lens": The central innovation is LCOE_adj, which adds the value of "lost compute" (idle GPU hours multiplied by spot market rates) to the capital and operating costs.
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) and the value of compute forgone during construction (set by deployment speed and chip-hour value). Their sum, a speed-adjusted cost (LCOEadj), determines the preferred option, which varies with location, scale, technology, and the value of time.

Datasets

The study synthesizes several high-fidelity public datasets:

  • Solar Resource: NSRDB PSM4 for hourly irradiance and temperature.
  • Costs: NREL Annual Technology Baseline (ATB) for solar and storage; PNNL for O&M; EIA for natural gas and industrial electricity rates.
  • Grid Timelines: LBNL data on median interconnection queue durations (2018–2023).
  • Compute Value: GPU spot market rental rates (e.g., $2.40/hr for an H100).

Key Insights and Ideas

  • The DC Advantage: Moving from AC to DC-coupled architectures eliminates the need for multiple inverters and improves path efficiency. This reduces lifetime energy costs by an average of 18% ($0.09/kWh).
  • Speed Over Efficiency: For a 100,000-GPU facility, a three-year delay in power isn't just a logistical headache—it represents billions in lost potential revenue. This "speed premium" dominates the economic decision-making process, often making solar microgrids preferable even in regions with mediocre solar resources.
  • Reliability is Expensive: Solar microgrids require massive overbuilding to reach "three nines" (99.9%) of availability because they must be sized for the worst-case multi-day cloudy stretch. However, the value of compute is so high that moving from 90% to 99% uptime is almost always economically justified.
  • Geography of Speed: While natural gas is cost-competitive for large scales in the Southwest, its deployment is hampered by turbine delivery delays. Consequently, DC solar becomes the "default" choice across most of the U.S. map when speed is valued.
Comparison of AC and DC power paths, highlighting the reduction in conversion steps for DC systems.

Key Results

  • DC Solar Dominance: Once foregone compute is valued, DC-coupled microgrids are the lowest-cost option in 85% of U.S. locations under baseline assumptions.
  • Cost Savings: DC coupling provides a mean advantage of $0.09/kWh over AC solar counterparts.
  • Threshold for Indifference: The median "indifference utility" for a GPU-hour is **$1.61**. Since current market rates are significantly higher ($2.40+), the economic case for fast-deploying microgrids is robust.
  • Natural Gas Sensitivity: If natural gas deployment could be accelerated to a 3-year timeline, it would become the preferred option in ~35% of locations, primarily in solar-poor regions like the Midwest and Pacific Northwest.
Maps showing the preferred architecture across the U.S. based on deployment speed assumptions.

Limitations and Future Work

The authors acknowledge several simplifying assumptions:

  • Grid Costs: The model uses prevailing industrial rates, which may understate the true cost of the massive infrastructure upgrades (substations, transmission) required for gigawatt-scale loads.
  • Supply Chain: The analysis assumes a mature supply chain for high-power DC equipment, some of which is currently pre-commercial or lacks a specialized workforce.
  • Land Use: Solar at this scale requires thousands of acres; the model assumes "frictionless" land acquisition, which may not hold in practice.
  • Future Tech: Future iterations could include "clean-firm" technologies like small modular reactors (SMRs) or advanced geothermal, though these currently lack the 24-month deployment timeline of solar.

Summary updated

Abstract

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.
BibTeX
@misc{newkirk2026speed,
  author = {Alex C. Newkirk and Daniel L. Gerber and Erica R. H. Fuchs and Alex Hubbard and M. Granger Morgan and Arman Shehabi and Sarah J. Smith and Constantine Samaras},
  title = {{Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States}},
  year = {2026},
  eprint = {2611.17098},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17098},
}
Figure