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…Figure 2 · Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States · Alex C. Newkirk et al. · CC BY 4.0
This study builds a computer model to compare four ways to power large AI data centers, including grid power, natural gas, and solar panels with batteries. Each option must provide the same electricity reliably.
The model counts both energy costs and the value of GPU time lost while construction is underway. Across modeled continental U.S. locations, solar with direct-current wiring most often came out lowest-cost under the study’s baseline schedules. This is a modeled result, not a record of completed facilities.
The research
A power choice can trade a cheaper kilowatt-hour for months or years when GPUs sit idle before a facility opens. Counting that delay changes what “lowest cost” means: in this model, getting a data center running sooner can outweigh paying more for its electricity. The result depends on the value assigned to that delayed computing.
An open-source cost model compares four ways to power AI centers and counts compute lost during construction, showing why faster buildout can outweigh lower energy prices.
Across four power options, the model finds designs that meet the same facility load and reliability, then adds the rental value of GPUs unavailable during construction to energy costs. For 100,000-GPU facilities across modeled continental U. S. locations, DC solar with storage had the lowest combined cost in 84.9% under baseline timelines. This supports the conclusion that earlier operation can outweigh cheaper energy.
The ranking depends on the assumed value of delayed GPU time and modeled construction schedules. The grid option also omits potentially substantial upgrade costs for large loads.
What the evidence shows
Compare equal power service
The open-source technoeconomic model compares four ways to supply power to AI data centers: grid connection, natural gas with diesel backup, and solar-plus-storage using alternating-current (AC) or direct-current (DC) distribution.
It asks whether energy price alone ranks options well when construction delays keep GPUs from operating. Every modeled option must serve the same facility load at the same reliability.
Count the cost of waiting
The model adds the estimated value of GPU time unavailable during construction to energy costs. It uses GPU rental rates as a proxy for that value, making a system that starts operating sooner comparable with one that may provide cheaper electricity after a longer wait.
The resulting speed-adjusted cost combines energy costs with the opportunity cost of delayed computing.
DC systems cost less on average
Because both designs provide equivalent energy service, their modeled cost difference reflects the distribution architecture and its power-conversion path. The mean advantage does not mean DC systems are cheaper at every location.
Speed changes the ranking
After the model counts the value of delayed GPU time, DC solar-plus-storage has the lowest speed-adjusted cost in 84.9% of modeled locations for 100,000-GPU facilities under baseline timelines. This comparison includes grid connection, natural gas with diesel backup, and both solar designs.
It supports the conclusion that quicker operation can outweigh differences in energy costs under the model’s assumptions.
The gas timeline matters
The authors also model natural-gas systems becoming operational in three years instead of four. In that scenario, DC solar remains lowest-cost in 55.6% of locations, while natural gas is lowest-cost in 35.2%.
The shift shows that the baseline ranking depends in part on construction schedules; the accelerated timeline is a scenario, not a typical deployment time.
