AI Elephants in the Room: Power and Water

The views expressed are those of the author and do not necessarily reflect the views of ASPA as an organization.

By Barbara L. Neuby
December 23, 2024

Use of artificial intelligence is exploding at all levels of government. Reports from academia and many agencies laud efficiencies and project wondrous things yet to come. A 2024 Ernst & Young survey found that 51 percent of public workers use AI daily, most frequently to summarize information (62 percent), conduct research (60 percent) or brainstorm topics (59 percent). Federal agency uses are noted in Figure 2 below. These efforts are made possible by data centers that consume enough power for millions of homes. While agencies like the Office of Management and Budget (OMB) and Government Accountability Office (GAO) issue lengthy memoranda promoting the ethics, innovation, risk management and efficiency, and President Biden’s 2023 Executive Order 14110 sets expectations and parameters for AI use, the two elephants in the room are power and water.  Only the Department of Energy (DOE) and a few states are addressing these problems head on.

A Mad Dash for Data Centers

A data center contains thousands of servers with millions of advanced, power-hungry microchips. Microsoft’s Copilot and Statista report that there are 5,381 data centers in the United States, a number that far exceeds that of any other country and represents 49 percent of the global supply of 10,900 centers. However, the International Energy Agency (IAE) claims the U.S. only has a 33 percent share of global data centers that consumed 200 TWh and this figure is expected to rise to 260 TWh in 2026. The U.S. share of power consumed by data centers would then be 66 TWh. According to the American Power Association, U.S. daily electrical capacity is 1.3 trillion watts (TWh). If we multiply the daily capacity of 1.3 TWh x 365 days, we find that our annual generation capacity is 474.5 trillion watt hours. Dividing our data center power consumption by our total capacity we learn data centers take up 13.9 percent of generating capacity. A watt is a unit of electricity flow. Table 1 is provided for scale. To keep this in perspective, one bitcoin mining search requires 800 KWh, or enough to power 25 homes for a day. A single natural language processor requires 10 GWh and produces a carbon footprint equal to the life cycle of 5 gas powered vehicles.

Chip advancements are fueling the power demand. Two of Nvidia’s new “Blackwell” chips consume 2,700 watts of electricity—a 300 percent increase in power consumption over the prior generation. Producers are scaling up their size and capacity since Big Tech (Microsoft, Google, Amazon and Meta) have increased orders by as much as 75 percent over the last two years. All are actively involved in constructing data centers.

Private Sector Response to Power Demands

Power created by the present electric grid barely keeps up with demand and the grid’s age is a de facto limit on future production. Natural gas and coal together supply roughly 64 percent of our electricity but they come with a carbon footprint. Renewables supply about 21 percent according to the EIA but cannot meet the growth curve demanded by data centers at this time. The DoE reports an additional 15.6 GWh of solar power added to the grid in 2024 with nearly 100 more projects slated to come online in 2025. However, it is nuclear power that can produce the mega wattage needed to support AI without the carbon emissions.

Introducing the small modular reactor (SMR) that needs less fuel, is smaller in size, can be shipped in units and erected in much less time and more cheaply than a regular reactor. Each SMR can produce roughly 50 to 300 MWh and there are several new, more efficient designs under review at the Nuclear Regulatory Commission (NRC). Google, Amazon, Meta, Microsoft, Alphabet, Elon Musk and BlackRock all have committed billions to build new data centers and SMRs. Microsoft and BlackRock have promised $30 billion to try to restart the Three Mile Island reactor in Pennsylvania that almost melted down in 1979.  North and South Carolina, Alaska, Pennsylvania, Oregon and Washington have signaled their support of small reactors as well. Many local utilities are aware they cannot provide the extra power need to run the centers and have been generally supportive of reactors. This party is just getting started but is a lot of investment considering that a 2024 USA Facts survey reported that 81 percent of the public say they do not trust AI to return truthful information.

Water and Servers

The other elephant in the room is the water used by the data centers. Mytton wrote in his Energy Policy article that a medium sized data center uses as much water as three hospitals or two golf courses and most of this water is drawn from potable sources. Water that is used to cool the servers is lost to the atmosphere through evaporation and takes about a year to make its way back to the ground, especially in warmer climates. Most data centers are in Virginia, perhaps due to the number of federal agencies in the region, but siting data centers in arid western states becomes a questionable policy decision. Training one GPT system uses 700,000 liters of water. Google’s data centers used 2.5 billion liters in 2022 a 20 percent increase over 2021, and Microsoft’s water consumption grew 34 percent in those two years. Together these two companies withdrew 1.5 billion cubic meters of water from the potable supply. Already, conflicts have arisen in city councils and county commissions over siting and water rights.

Conclusion

AI offers clear benefits and is here to stay, even with the surge in power and water consumption. While the private sector is the main creator of power, states and communities are stepping up. Less has been done to address the water issue. In the last series of these articles, we will look at some of the newly recognized best practices in power generation and water conservation to enable continued AI success. 

---

Author: Barbara Neuby teaches at Kennesaw State University, north of Atlanta, where she researches advancing issues like the central bank digital currency, global financial problems, and the pros and cons of artificial intelligence. She believes these issues are connected and will play out in a new global financial system. Neuby can be reached at [email protected].

(2 votes, average: 5.00 out of 5)
Loading...