Ringside: The Energy and Water Footprint of AI

Whether artificial intelligence should be regulated, and if so, how, has almost overnight become one of the biggest political issues of our time.  And with AI comes a surge in construction of massive data centers, for which we should minimize impact on the surrounding communities. But to the question of how we will find enough energy and water to supply data centers, concerns may be overstated.

When it comes to water, there is good news. A recent report published by Richard Evans, an economist with the Abundance Institute, describes recent recycling innovations that promise to reduce ongoing water consumption by data centers to almost nothing. Using a liter per kilowatt-hour metric, modern designs are already down from as high as 2.5 L/kWh with traditional evaporative cooling to 0.12 L/kWh with closed-loop cooling technologies.

It may get even better than that. Evans elaborates, “And those fleet averages actually understate what is occurring in new construction. Microsoft says its newest liquid-cooled AI data centers use direct-to-chip closed loops with zero water evaporation. Oracle says the same about AI facilities it is building in New Mexico, Michigan, Texas and Wisconsin. The cooling system is initially filled and subsequently has essentially no evaporation, blowdown, or continuous makeup-water requirement.”

Even at 0.12 L/kWh, with worldwide data center electricity consumption estimated at 485 terawatt-hours in 2025, that only equals 47,202 acre feet. To put this into perspective, total worldwide water withdrawals per year are estimated to be 2.6 billion acre feet for agriculture, 746 million acre feet (MAF) for industry, and 373 MAF for municipal use. Data centers with state-of-the-art designs already have the potential to use less than one-eighth of one percent (0.013%) of municipal water consumption, with next generation, already proven designs capable of bringing that down to zero.

What about energy?

According to the International Energy Agency (IEA), as noted, data centers worldwide consumed 485 terawatt-hours (TWh) in 2025. That figure is roughly corroborated by Gartner’s estimate of 448 TWh. And both of these sources project consumption to double by 2030 to 950 TWh and 980 TWh, respectively. But this doesn’t tell the whole story.

The Statistical Review of Global Energy, using data gathered by S&P, reported data power demand at 788 TWh in 2025. The difference can be explained, but reveals the complexity inherent in making these estimates. S&P’s numbers are derived from their inventory of individual data-center facilities around the world. The IEA relies on industry projections. Both methods are credible, and both require huge sets of assumptions.

It’s important to recognize that “data centers” have already been around for quite a while. They are defined as “servers, storage, networking equipment, and the infrastructure needed to support them.” Those components didn’t just show up when AI came along. And they are part of a larger information and communications technology (ICT) ecosystem that also includes telecommunications networks, end-user devices, and cryptocurrency mining. For example, in 2023, the IEA estimated global ICT electricity demand as follows: data centers 360 TWh, networks 280 TWh, end-user devices 440 TWh, and cryptocurrencies 125 TWh. Total: 1,205 TWh.

Putting data center electricity demand into perspective must first acknowledge how difficult it is to make an accurate estimate, but let’s go with the Statistical Review’s estimate, compiled from S&P’s data. The worldwide total in 2025 was 788 TWh, and of that, U.S. demand was 313 TWh. China’s data centers used 206 TWh, and the rest of the world used 269 TWh. How does that compare to total electricity production?

Worldwide electricity production in 2025 was 32,202 TWh, meaning data centers consumed 2.4 percent. In the US, we produced 4,772 TWh, and our data centers used 6.6 percent of it. China’s total electricity production was 10,575 TWh, and their data centers used 1.9 percent. Of course, overall, China’s population of 1.4 billion only utilized 7.5 megawatt-hours per person in 2025, compared to 13.8 in the U.S. On a per capita basis, we have far more electricity to spare than the Chinese do, making our 6.6 percent share for data centers less onerous than it would otherwise be.

When we speak of the dawning electric era, the electricity hogs are not just data centers. Converting homes to heat pumps, cars and trucks to EVs, investing in large scale desalination, and deploying millions of robots will require tens of thousands of additional TWh. And all of those innovations will require an expanded information and communications infrastructure – i.e., the more we electrify and automate, the more data centers we’re going to need.

However, there is also good news here. Without diving too deep, just imagine how much electricity a primitive computer using vacuum tubes would require just to perform basic calculations. We’ve come a long, long way. Computational power is often measured using “FLOPs,” or floating point operations per second. There are other measuring units, depending on the task, but for describing the relentless progress still being made in processor efficiency, FLOPs will do.

In the previously referenced analysis of water and energy use by data centers, author Richard Evans cited extensive research that projects the ratio of electricity to FLOPs to improve by 34 percent per year until at least 2030. Concurrently, most estimates project electricity use by data centers to double by 2030. But the point here is that the computational power of data centers will improve far more than mere doubling by that time.

Again, without going too deep, the current benchmark efficiency for data center processors, writ large, is 12.5 octillion FLOPs per terawatt-hour. By 2030, that number is projected to increase to 54.2 octillion FLOPs per TWh, which is 4.3 times greater than 2025. If the electrical input doubles, that means the capacity of data centers in 2030 will be 8.6 times greater than it is today. We may hope that is enough.

Ultimately, data centers are at the center of a broad energy transition that will require far more electricity than the projected requirements of the data centers themselves. Our need for data centers should be beyond serious debate. Fraught with potential perils that could rival their stupendous upside, they are nonetheless coming, whether we build them here or let others take the lead.

Here in California, data centers present an opportunity analogous to the regulatory situation facing our in-state oil and gas industry. We can enact smart regulations that empower the industry to thrive and set a positive example to the world, or we can drive it away to more welcoming states and nations.