On November 12, 2026, researchers from Carnegie Mellon University and the University of California, Berkeley / Lawrence Berkeley National Laboratory published a technoeconomic analysis demonstrating that deployment speed, rather than unit energy generation costs, dictates the most economical power architecture for artificial intelligence data centers in the continental United States. The preprint, titled ‘Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States’ (arXiv:2611.17098), introduces a speed-adjusted levelized cost of energy (LCOE) metric valuing the compute unlocked by faster construction, finding that multi-year grid interconnection delays make off-grid direct-current (DC) solar-plus-storage microgrids the lowest-cost power option across roughly 85% of continental U.S. locations.
How does deployment speed alter data center energy economics?
The research team—Alex C. Newkirk, Daniel L. Gerber, Erica R. H. Fuchs, Alex Hubbard, M. Granger Morgan, Arman Shehabi, Sarah J. Smith, and Constantine Samaras—modeled a 100,000-GPU training facility (131 MW design load, based on eight-H100 nodes drawing 7.3 kW average and 8.5 kW peak with cooling) across four power architectures: alternating-current (AC) solar-plus-storage, direct-current (DC) solar-plus-storage, islanded natural gas, and commercial grid interconnection.
Under unadjusted metrics, commercial grid electricity or natural gas appears cheapest per kilowatt-hour. However, grid interconnection queues across U.S. transmission operators median 4 to 5+ years, while gas turbine backlogs impose 36 to 48-month lead times. In contrast, solar-plus-storage microgrids can be energized in 24 months. Incorporating the Value of Lost Compute (calibrated at .40 per GPU-hour) reveals that idling compute during interconnection delays destroys far more economic value than energy price differences. In speed-adjusted terms, DC solar microgrids are lowest-cost across 84.9% of continental U.S. locations under baseline 48-month gas timelines, and 55% under accelerated 36-month gas timelines. Grid power remains preferred in only 9% to 15% of sites where short queues, low industrial power rates, and poor solar insolation align.
| Power Architecture | Deployment Timeline | Unadjusted Cost Profile | Speed-Adjusted Preference (% of US Sites) | Key Operational Dynamics |
|---|---|---|---|---|
| DC-Coupled Solar + Storage | 24 months (2.0 years) | bash.21–bash.61/kWh across US; 18% cheaper than AC | 84.9% (baseline) / 55% (fast gas) | Eliminates DC-AC-DC conversion stages; avoids inverter hardware costs. |
| AC-Coupled Solar + Storage | 24 months (2.0 years) | Higher than DC; ~bash.09/kWh average penalty | Sub-optimal to DC | Faces round-trip efficiency losses through repeated power conversion. |
| Islanded Natural Gas | 36–48 months (3.0–4.0 years) | Low unit energy cost where pipelines exist | 0% (baseline) / 35% (fast gas) | Constrained by multi-year turbine delivery times and gas price exposure. |
| Grid Interconnection | 48–60+ months (4.0–5.0+ years) | Lowest unadjusted LCOE at industrial rates | 15.1% (baseline) / 9% (fast gas) | Severely delayed by regional utility queues and transmission upgrades. |
Architectural efficiency: Why DC-coupled microgrids cut energy costs
Beyond timeline advantages, power system topology yields substantial operational savings. Across modeled U.S. locations, DC-coupled systems cut lifetime energy costs by an average of 18% (~bash.09/kWh) versus AC counterparts. Eliminating redundant DC-to-AC and AC-to-DC conversion stages removes inverter hardware and improves round-trip efficiency along both solar-to-load and battery discharge pathways. In a 100,000-GPU Permian Basin installation, DC coupling lowers energy costs from bash.33/kWh to bash.27/kWh, generating a 70 million net present value reduction in energy expenditures.
The study also revealed that microgrid costs track the worst sustained five-day low-insolation window rather than annual averages, creating threefold cost differences between sites with identical annual insolation (bash.21/kWh in Colorado versus bash.61/kWh in Northern California). Furthermore, while moving from 90% to 99% uptime saves ~bash.20/kWh in avoided compute loss, pushing beyond 99.9% uptime is nowhere economically justified on compute value alone due to exponential battery oversizing costs.
Implications for academic HPC and national research infrastructure
These findings carry significant implications for university high-performance computing centers and research offices planning AI cluster expansions. While commercial cloud operators can finance multi-hundred-million-dollar off-grid microgrids to capture commercial compute revenues, academic institutions and national data infrastructure initiatives operate under fixed public grants and capital constraints, risking an exacerbated compute divide.
Moreover, institutional initiatives around sustainable HPC and mitigating the compute carbon footprint face shifting grid dynamics. When hyperscalers draw massive grid loads, required transmission upgrades are routinely socialized across utility rate bases, forcing general ratepayers to cross-subsidize commercial data centers. Conversely, if commercial operators build parallel off-grid microgrids to bypass queues, they sever ties with shared grids, leaving research institutions to grapple with congested utility queues and uncertain clean energy transitions.
Sources
- Newkirk, A. C., Gerber, D. L., Fuchs, E. R. H., Hubbard, A., Morgan, M. G., Shehabi, A., Smith, S. J., and Samaras, C. (2026). ‘Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States.’ arXiv preprint arXiv:2611.17098v1: https://arxiv.org/abs/2611.17098 (checked 2026-11-15)
- Lawrence Berkeley National Laboratory (2025). ‘Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection.’ LBNL Electricity Markets and Policy Group (checked 2026-11-15)
