The AI energy bill is finally being measured — and the supply answer is already being built
A study of US data centers finds simple persistence beats electricity-scaling for water estimates, a UAE study prices the cooling trade-off, and Redwood and Antares are building storage and nuclear supply. Measurement first, then megawatts.
- In a preprint, persistence models provide lower error in annual water-withdrawal estimates for data centers than electricity-scaling rules, which consistently overestimate actual usage. Self-reported m:…8279
- In a preprint, data-center expansion in the UAE could consume up to 11.5% of national electricity generation by 2040, dry cooling cuts water demand by 98.2% at an 11 to 13% electricity penalty, and system-wide greenhouse gas emissions rise by 2.27 to 19.9 million tonnes of CO2 equivalent annually when demand is met by natural-gas-dominated grids. Self-reported m:…7101
- Redwood Materials says it repurposes retired electric vehicle batteries for stationary energy storage, deployed a 12 MW / 63 MWh microgrid for Crusoe, and reports that the microgrid has maintained over 99% availability since launch. Self-reported m:…0275
- Antares says the US Department of Energy approved the Nuclear Safety Design Agreement for its Mark-1 reactor, which is scheduled to operate at full electrical power for several months in 2027 at Idaho National Laboratory and is commercially representative of the plant it will deploy under the Army Janus program in 2028. Self-reported m:…9991
Measuring the water bill
The debate about AI's resource footprint has often run on estimates built from other estimates, and a preprint on US data centers tests whether the most common shortcut works. Persistence models provide lower error in annual water-withdrawal estimates for data centers compared with electricity-scaling rules, and electricity-scaling rules for data center water withdrawal consistently overestimate actual usage. Persistence here means the unglamorous assumption that next year's withdrawal resembles last year's, and it beats scaling water use up and down with electricity consumption. For planners, that is a useful warning: a model that feels more sophisticated because it tracks electricity can still be systematically wrong about water. m:…8279
The same preprint offers a reason the scaling shortcut fails. The withdrawal-to-electricity ratio for the studied data centers declined annually between 2017 and 2023, so a fixed ratio carried forward from earlier years overstates later water use. Efficiency improved, in other words, and any estimation method that assumes a constant relationship between power and water will drift upward as that ratio falls. The implication is practical rather than academic: water plans built on electricity-scaling rules should be re-checked against observed withdrawals before they drive infrastructure decisions. m:…8279
Pricing the trade-off
A second preprint, on data-center expansion in the UAE, puts numbers on the stakes of getting the supply mix wrong. Data-center expansion in the UAE could consume up to 11.5 percent of national electricity generation by 2040, and system-wide greenhouse gas emissions increase by 2.27 to 19.9 million tonnes of CO2 equivalent annually when data-center demand is met by natural-gas-dominated grids. The wide emissions range reflects different build-out paths, but even the low end is a national-scale addition. The point is not that data centers must not grow; it is that their growth lands on a grid, and the grid's fuel mix decides the climate result. m:…7101
Cooling is where the water and energy bills collide. Dry-cooling configurations in UAE data centers reduce water demand by 98.2 percent but incur an 11 to 13 percent electricity penalty. That trade is close to a straight swap: nearly all the water saved, at the cost of roughly a tenth more electricity. In a water-scarce setting that trade may be worth making, but it only helps if the extra electricity is clean. Dry cooling powered by a gas-dominated grid trades a water problem for an emissions problem, which is why the preprint's two findings belong together rather than apart. m:…7101
Building the supply answer
The supply response is already taking physical form, starting with storage built from the waste stream of another industry. Redwood Materials says it repurposes retired electric vehicle batteries for stationary energy storage, developed its Pack Manager electronics to handle diverse battery chemistries and voltages, and manages the full energy stack including acquisition, integration, and site design. The company says its assessment tool evaluates battery capacity and health in a few minutes, which is the operational detail that makes second-life batteries usable at scale: each retired pack is different, and sorting them quickly is the bottleneck. m:…0275
Redwood says it deployed a 12 MW / 63 MWh microgrid for Crusoe, that the microgrid has maintained over 99 percent availability since launch, and that Crusoe increased compute power draw on the campus by 7 times since the microgrid launch. The company also says its Pack Manager configuration reduces transformers and inverters by 90 percent in solar-plus-storage sites, and that it is moving toward integrating solar and storage to provide clean firm power. Those are company claims, not independent verification, but they describe a coherent proposition: storage that lets a compute campus grow without waiting for grid upgrades, assembled from batteries that already exist. m:…0275
Antares offers the nuclear side of the same answer. The company says the US Department of Energy approved the Nuclear Safety Design Agreement for its Mark-1 reactor, that Mark-1 is scheduled to operate at full electrical power for several months in 2027 at Idaho National Laboratory, and that Mark-1 uses the same core, TRISO fuel, and test facility as Mark-0. Antares says Mark-1 will be the first full-scale electricity-producing microreactor test in the United States and is commercially representative of the plant it will deploy to Fort Bragg under the Army Janus program in 2028. Measurement is catching up with the footprint, and supply is being built in parallel: storage from retired batteries, and small reactors aimed at firm power. The energy bill is no longer an abstraction, and neither is the response. m:…9991
What is still unresolved
- Whether persistence models generalize beyond the studied US data centers to other climates and cooling systems.
- Which UAE build-out path materializes, and therefore where in the 2.27 to 19.9 million tonne emissions range the outcome lands.
- Whether the Mark-1 test proceeds on the stated 2027 schedule and the Janus deployment follows in 2028.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.