Water, Noise,Power: The Real Costs of Data Centers
I expect AI data centers to become a major issue in the 2028 presidential primaries, particularly on the Democratic side. I recently argued that the case for building more data centers and compute is getting lost amid the backlash, and I have also written about the financial bubble risks…
I expect AI data centers to become a major issue in the 2028 presidential primaries, particularly on the Democratic side. I recently argued that the case for building more data centers and compute is getting lost amid the backlash, and I have also written about the financial bubble risks surrounding the buildout. But the political debate is increasingly about what happens outside the data center: what these facilities consume and what nearby communities are asked to absorb. So I want to look more closely at three concerns that come up again and again: water, electricity, and noise. Subscribe to our weekly newsletter Water Water use is not a national crisis, but it can become a real local constraint depending on where a data center is built and how it is cooled. At a national level, data centers do not look like a water crisis. U.S. facilities consumed about 66 billion liters directly in 2023. Even in Virginia, the country’s largest data center market, they accounted for less than 0.5% of statewide water withdrawals. But those averages hide a lot. At six Virginia water utilities, data centers ranged from a negligible share of consumption to roughly one-fifth. Most facilities used no more water than a large office building, while one consumed 243 million gallons in a year, about as much as 36 large offices. So I think the useful question is not how much water data centers use nationally. It is how large a particular facility is relative to the water supply available where it is built. There is another wrinkle. The water used at the data center is only part of the footprint. In 2023, generating electricity for U.S. data centers consumed close to 800 billion liters of water, roughly 12 times the amount used directly by the facilities. Location and design also change the numbers dramatically. Researchers find more than a 10,000-fold difference in water use for the same amount of computing depending on the servers, their utilization, the local power supply, cooling design, and climate. Evaporative cooling can save electricity but use more water, while dry cooling does the reverse. This is why I would be skeptical of simple claims about how many liters an AI query consumes. Water is less a property of the query than of the system and the place where it runs. Noise Noise is a real problem at some sites, especially near homes, but the evidence points more to siting and measurement than to data centers being inherently loud. Noise appears to be a real problem, but at a minority of data centers. Virginia identified problematic noise around roughly 15 facilities, about 10% of operating sites. Siting seems to be an important part of the story. Statewide, 29% of data-center properties were within 200 feet of residentially zoned land, roughly half the length of a football field. In Fairfax County the figure was 55%, compared with 24% in Loudoun and 21% in Prince William. Those differences do not prove that proximity causes the complaints, but they suggest that where a data center is built matters a lot. The issue is also less about extreme volume than persistence. Cooling fans, HVAC systems, and transformers operate continuously, so a low-frequency hum that would be easy to ignore for a few minutes can become much harder to live with every night. There is also a measurement problem. One Virginia field study recorded a continuous 55 Hz tone at about 51 dB outdoors and estimated roughly 39 dB inside a typical house. Run that through A-weighting, the filter behind most dBA noise measurements, and the reading drops to only 14 dBA. In other words, the standard metric can make low-frequency sound look much less significant than it feels to someone hearing it. There is an important caveat: the researchers did not establish that the data center produced that particular tone. They estimated the facility itself added only about 2 to 4 dB above the traffic background in some low-frequency bands, and it remained within the proposed limits. My read is that noise is primarily a siting, design, and measurement problem. The evidence for annoyance is credible. The evidence that data-center noise causes specific long-term health problems is still extremely thin. Electricity Electricity is the bigger concern: data-center demand is rising fast enough to strain parts of the grid, while the impact on household bills depends largely on who pays for the new infrastructure. Electricity is where I think the concerns about data centers are most justified. U.S. data centers consumed about 4.7% of U.S. electricity in 2024. The central projection puts them at 11.8% by 2030, more than tripling their electricity use in six years. AI is driving much of that growth. In the central projection, AI servers account for 55% of total data-center electricity by 2030. The harder problem is how quickly this new demand is arriving. Grid planners increased their ten-year forecast for summer peak-demand growth from 132 GW to 224 GW in a single year, with data centers responsible for most of the increase. Data centers can be built faster than new power plants and transmission lines, so generation, transmission, and grid connections all have to catch up at once. The impact on electricity prices is more complicated. Data centers have already raised average wholesale electricity prices by roughly 3% to 5%, with much larger effects in some regions. Yet the best recent study of household prices found that data-center growth reduced average residential electricity rates between 2015 and 2024. For the average growth experienced from 2019 to 2024, the estimated reduction was about 6% relative to the counterfactual. Both can be true. Data centers increase the cost of producing the next unit of electricity, but they can also spread the grid’s large fixed costs across more electricity sales. The question is what happens when utilities have to build substantial new infrastructure. That is why we are starting to see long contracts, minimum payments, collateral, and customer-specific infrastructure charges for large users. I think that is the question to watch: not whether data centers raise or lower electricity prices in general, but who pays for the new grid capacity they require. Related content The AI Data Center Backlash Has a Blind Spot Why Data Centers Became the Face of the AI Backlash The “Data Center Rebellion” is here The Bear Case for AI Data Centers The post Water, Noise,Power: The Real Costs of Data Centers appeared first on Gradient Flow.Source: Gradient Flow — Published — Category: Models