Build speed, not power price, decides how to power AI data centres

A computer model of 4,317 US sites finds that solar panels and batteries wired for direct current win in most places once lost computing time is priced in. The study is a preprint and has not been peer reviewed.

6 min readIIIOpen sourcePreprint, not peer reviewed
Save
A studio photograph of a brass balance scale tipping toward an hourglass with blue sand, outweighing stacked copper coins.AI illustration
Illustration: AI generated

What the study shows

The authors built an open-source cost model of four ways to power an AI training data centre with 100,000 GPUs, the chips that do the computing. The facility needs about 131 MW at peak. The four options were: solar plus batteries with conventional alternating-current (AC) wiring; solar plus batteries with direct-current (DC) wiring; an off-grid gas plant with diesel backup; and a normal grid connection. Each option was sized to deliver the same power, with 99% uptime, at each of 4,317 sites across the continental US. This is a simulation and a preprint that has not been peer reviewed.

Solar panels and batteries produce direct current. Most facilities convert it to AC with inverters and then back again inside the servers. Each conversion wastes some power. The authors report that keeping the whole system on DC cut lifetime energy cost by 18% on average, about 9 cents per kWh. The saving comes from fewer inverters and smaller losses along the way. For one West Texas site, they put the saving at about $570 million over the project's life.

On energy price alone, solar plus batteries was rarely the cheapest choice. A grid connection usually beat every off-grid option, although the authors say their grid cost leaves out upgrade costs and is a lower bound. Large gas plants also beat solar at big scale.

The authors then added a speed-adjusted cost. It charges each option for the computing lost while it is being built, priced at $2.40 per GPU-hour, the rental price for an H100 chip. Solar is assumed to take 2 years to build, gas about 4 years, and a grid connection 2.5 to 5 years depending on the state. On this basis, DC solar plus batteries was the cheapest option in about 85% of locations. The grid won in the rest, where power is cheap and connection is fast, such as New England and East Texas.

If gas plants could be built in 3 years, the picture shifts. DC solar stays cheapest in about 55% of locations, gas wins about 35%, mostly in cloudy regions such as the upper Midwest, and the grid about 9%.

Solar cost depended more on the worst run of cloudy days than on average sunshine. The system has to ride out the longest dull spell. Two sites with similar yearly sunlight, one in Colorado and one in foggy northern California, differed in cost nearly threefold.

Our reading

In our reading, the paper makes two claims of different strength. The DC wiring result comes from engineering arithmetic: fewer conversions, less lost power, less equipment. The authors' checks show it holds unless DC equipment costs reach roughly twice their central estimate. It is the more durable finding.

The speed result depends on a pricing choice. It treats every idle GPU-hour as worth its rental price and adds that to the energy bill. The authors show where the answer flips. The break-even value between DC solar and the grid has a median of about $1.61 per GPU-hour, and ranges from under $1 to over $4 by location. So the verdict depends on what a given firm thinks an hour of its chips is worth.

A concrete case: a firm that values its chip time well below the rental price would find the grid competitive in more places. A lab that values its training runs above $2.40 an hour would find the argument for a fast-built solar site stronger still. The model lets each firm put in its own number, which is its main practical use.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

III
Open source
LevelStatusEvidence
I  Lab onlyMetA simulation study with a published method, and no real-world deployment of the proposed DC design.
II  BenchmarkedMetThere is no standard public benchmark for this kind of study. It does include internal validation: out-of-sample checks of the battery surrogate, optimiser-stability runs, a historical-weather re-run at 28 sites, and sensitivity tables. Its inputs come from public datasets (NREL, EIA, LBNL). This is a loose fit for 'benchmarked'.
III  Open sourceMetThe full model (power flow, dispatch, degradation, optimisation) is released under an MIT licence at github.com/acnewkirk/AI-Datacenter-Microgrid-Analysis. It includes result tables, config files and a notebook that reproduces the main figures. Re-running the full sweep needs a free NLR API key.
IV  In productionNot yetNo named organisation is documented using this framework or running a 100 MW+ all-DC data centre. The cited DC pilots are small (for example, 200 kW) and predate this work.

What would move it up
To reach level 4: a named operator would need to run a data centre on DC-coupled solar plus batteries and report cost, uptime and build time. Better evidence for the main result: measured demand-side interconnection timelines, real build times for solar and gas, and validation of conversion losses and power-transient handling at 100 MW scale. Peer review would also help.

Checked automatically every week, last on 16 November.

The study

Alex C. Newkirk, Daniel L. Gerber, Erica R. H. Fuchs et al.. Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States. arXiv, 12 November 2026. Read the paper

Preprint, not peer reviewed · Licence CC BY 4.0

Notes

  1. Newkirk, A. C. et al. Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States. arXiv preprint, 12 November 2026, not peer reviewed. CC BY 4.0. Source
  2. Code, result tables and configuration files, MIT licence. A notebook reproduces the main figures; rerunning the full sweep needs a free solar-data API key. Source
  3. Exact figures: DC advantage 18.3% mean, $0.0879/kWh mean (Fig. 3). Permian Basin example $0.27 vs $0.33/kWh, $570 million NPV saving. Speed-adjusted shares: DC solar 84.9% baseline; accelerated gas case DC solar 55.6%, gas 35.2%, grid 9.2%. Break-even GPU value median $1.61/hr. Colorado vs northern California $0.21 vs $0.61/kWh. Near-tie sites 149 of 4,317 (3.5%) baseline, 348 (8.1%) accelerated gas.
  4. An earlier version was posted under the title 'Technoeconomic Analysis of Microgrids for AI Data Centers in the Continental United States'. Source

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Efficiency