Datacenters over America: The Power Dance with AI (Part II)

In Part I of this two-part examination of datacenters and AI, we explored the frantic boom in AI-driven datacenter construction and the furiously negative public reaction to it. Now let’s take a closer look at the actual impacts of datacenters on communities and the environment, and the complex interplay between how people feel about AI and how they feel about datacenters.

Power Demand

AI requires microprocessors to run flat out, and the power draws can be substantial, such that datacenters are projected to increase total electrical power use in the United States by 10-12 percent by 2030, up from about 4 percent today. Some in the industry have suggested that datacenters could be using 50 percent of today’s total electrical energy generation capacity by 2040.

If this appetite for electricity is met by fossil fuel generation, datacenters will be pumping more CO₂ and other pollutants into the atmosphere on a huge scale. To the extent that they use cleanly generated electricity, this problem is reduced, but they will still compete for green energy with other users, which can increase carboniferous generation elsewhere. The pace of datacenter construction is extremely rapid, and in many cases the operators turn to dedicated off-grid gas turbine generation since they frequently can’t make timely deals for green power via the grid.

Microsoft, for example, gets power however it can: it uses clean hydroelectric power in Quincy, Washington; it paid to have a Three Mile Island nuclear power plant turned back on in Pennsylvania (also green power as far as atmospheric carbon is concerned); and it did a deal with Chevron in Texas for gas turbine generation, which is not remotely green. However, the manufacturer of the turbines Microsoft will use in Texas, GE Vernova, touts itself as a leader in carbon capture technologies capable of extracting 95-98 percent of CO2 from exhaust, so perhaps Microsoft will plan its project there to take advantage of this technology at some point in the future.

Water Use

It’s true that early datacenters located themselves in places where they could buy lots of fresh water to use for evaporative cooling—out of the ground or river, through the datacenter, and off into the atmosphere. Using local fresh water to cool datacenters is the easiest solution, but in some locations such use can create serious problems.

In Arizona, for example, agriculture is already draining groundwater and aquifers at an unsustainable rate, and rapid population growth, replete with numerous golf courses, is further depleting the available water. The idea of adding large-scale datacenter water usage to that already-unsustainable situation is simply terrible. However, datacenter designs that draw no water are in development, and jurisdictions in Arizona could require such designs for datacenters that would otherwise compete for the already over-stressed local water.

Under pressure to reduce water use, datacenters have turned to air cooling, frequently using systems analogous to a car’s cooling system: a recirculating fluid which carries heat from the servers to a radiator which dissipates the heat into the air. Other mitigations are in development: more efficient chips that do more computing while generating less heat, immersion of servers in a flow of non-conductive liquid to carry heat away more efficiently, and systems which cool only the hot chips, not the entire datacenter, to give three examples.

There might also be opportunities to transfer heat to large bodies of cold water (looking at you, Puget Sound) in ways that have little or no negative impact. Cornell University has been cooling its entire campus (8 million square feet inside over 100 buildings) since 2000 using the coldness of nearby Cayuga Lake with almost immeasurably small impacts on the temperature of the lake. Cayuga Lake is about three times the volume of our own Lake Washington, and about 1/18^th the volume of Puget Sound, not to mention Puget Sound’s partner, the Pacific Ocean. Western Washington could conceivably offer locations where datacenters could access vast amounts of cooling while neither using fresh water nor heating up the air.

The scale of datacenter water use needs to be understood in context. It’s tiny compared to agriculture, which accounts for nearly 80 percent of all water use in the US. It’s even tiny compared to watering golf courses: all of Google’s datacenters combined, for example, use about as much water as 50 golf courses and Google claims to replenish nearly 80% of the water it uses. There are 16,000 golf courses in the US, with about 15 new ones and 60-70 restored and reopened ones added per year.

The US has about 30 million golfers, whose combined water use dwarfs the water used by the much larger number of datacenter end users. There is some controversy around water use by golf courses (not enough, in my view), but activists and political leaders are hardly crawling over hot coals to oppose golf the way they are to oppose datacenters. There are, as far as I know, no moratoria on playing golf currently in the United States, where about half of the world’s golfers and golf courses are found.

Loudoun County makes more than enough money from datacenters to fully mitigate the impact of water use, if it wishes to. Out of datacenter revenues, the county could afford to take water from Chesapeake Bay, desalinate it, and pump it nearly two hundred miles to Loudoun County, with money to spare, and no harm done to the Bay. Interestingly, citizens appear to want to spend that money in other ways. Even so, Loudoun County may never have to consider pulling water from the Chesapeake, as datacenter designers find ways to use less (or no) water if required to.

Land use is also a flash point, since individual hyperscale datacenters are really big—one proposed for Utah would be twice the size of Manhattan, and six times the size of the aforementioned Hypergrid in Texas. To New Yorkers that sounds huge; to residents of the American West, not so much.

Datacenters are generally built flat on the ground, since it would be more expensive to build multi-story structures that would bear the weight of all those server racks. If you have a vacation home on a ridgetop overlooking a verdant valley and somebody puts a datacenter on the valley floor, it’s going to be ugly, no question. Still, there’s a lot of land in this country that isn’t in anybody’s view, and the total land used by datacenters is simply quite small, even if each one is large. By one estimate, we devote about half as much land to datacenters as we do to Christmas tree farms. It seems that, with permitting power and public support, governments could prevent poor land use choices by datacenter operators without preventing datacenter development in general.

Noise pollution

The noise they generate comes from massive cooling machinery (think fans, compressors and pumps) and electrical generation, often diesel backup generators or full-time gas turbines.  The noise is generated 24/7, and an unusual proportion of it is at very low frequencies, including below the hearing threshold. It can cause sympathetic vibrations within nearby homes and offices, and “nearby” can mean within a mile of a smallish datacenter, and several miles of a large one.

People living within the affected areas say it keeps them from sleeping, makes them sick, drives wildlife away, and trashes the value of their properties. First generation stand-alone datacenters are mostly significantly smaller than the latest hyperscale centers, but most are in urban and suburban areas, near homes, schools, offices, parks, churches, etc.

Planned new construction is much more likely to be in rural areas, which is a partial way to mitigate noise: on cheaper rural land, it would be easier to establish a wider buffer, and stock it with soft abatements such as plantings. Other mitigations include reducing the use of air cooling with fans by using liquids closer to the chips (more expensive, and potentially requiring greater water use), and physical barriers in and around the datacenter to absorb noise before it can radiate into neighboring properties.

Many jurisdictions have noise ordinances, but they are often mismatched to the new kind of noise produced by datacenters. The law creating the EPA stipulated noise pollution control as an EPA function, but the Reagan administration zeroed out the budget for that department in the 1980s, and it has never been restored. If datacenter builders decide that they need to stop making enemies so gratuitously, a commitment to noise abatement would be relatively low-hanging fruit for them.

Mitigation Deals?

Deals with datacenter operators are likely to be able to mitigate most impacts except electrical power use fairly straightforwardly, if the governments are determined to make the deals. Why wouldn’t they? Well, if the controlling jurisdiction is a red state and it has been generously lobbied by the AI industry, there’s a good chance concerns about impact will be downplayed.

There are existing datacenters which have produced problematic noise and air pollution, but governments could require mitigation in future datacenters, of which many more have yet to be built. Similarly, there are some datacenters today making unwise use of water, and some in the planning stages which should not be built for this reason, but negotiation and innovation can prevent more of this.

So far, datacenter power use hasn’t necessarily resulted in immediate increases in electrical rates for other users, but there are clearly examples where it has, and that will certainly become more common as they accelerate demand faster than supply can keep up. To some extent datacenters limit this problem by installing backup generation which can reduce their load on the grid at peak times (at a cost in air pollution and noise pollution), and the trend toward self-powered datacenters will also reduce their drain on the grid, but also result, at least in the near term, in more CO₂ emissions and more noise.

The deals governments make with datacenter operators about electrical power will probably be for longer-term mitigation of the electricity problem: grid upgrading, green generation, and so forth. There’s also an opportunity for time-shifting data center use, where they would install large battery banks, charge those batteries at times of day when usage by others is low (the middle of the night, for example), and then reduce their draw during peak hours by relying on the batteries, rather than noisy, dirty diesel generators.

If governments (including utilities) are actually able to negotiate good deals, such that most impacts are well-mitigated, will that quell the surge in opposition to datacenters? If the stated concerns about impact are the only concerns, the answer should be yes. If instead the opposition to datacenters is really a stand-in for opposition to AI, then turning the datacenters into model corporate citizens won’t satisfy the opposition.  Then what?

As mentioned above, there’s a case to be made that AI is overpromising and underdelivering, in which case there could be an AI bust in the near future, in which corporate bean counters rebel against the money being spent on tokens, Wall Street decides that the AI companies are overvalued, their stock prices crash, and demand for more datacenters quickly withers.

As with the boom-and-bust cycles associated with railroads and the Internet, this is likely to be a temporary reprieve. There is going to be a substantial market for AI, even if its impacts are not as apocalyptic as its promoters have predicted. After a shakeout, demand for datacenters will begin to climb again.

Datacenters’ demands for electricity and cooling have inspired furious innovation by chipmakers and the providers of underlying technologies, which could, over time, result in improved efficiency for the chips that provide processing and memory, reducing net demand for electricity per unit of compute, less water use, and less noise.

Long-term, contributions from AI firms to the development of renewable generation and a smarter grid could help Electricity 2.0 accelerate. AI firms claim that AI itself has the potential to improve the efficiency of energy use across the economy, offsetting the impacts of AI’s own use. That may be true but would require independent verification.

Ultimately, if citizens develop a more positive view of AI, and if datacenter developers develop a willingness to accommodate public concerns about energy, water, noise, and land use, datacenters could move back down the list of things to be mad about.

Citizen enthusiasm (or lack thereof) for AI will depend on how AI’s costs and benefits unfold, and that unfolding is going to take time. There’s some sketchy evidence that AI is not only NOT taking away jobs but may even be increasing hiring: at least some firms find that AI-supported humans create so much more value that they want more such humans. If this mini trend continues, perhaps concerns about a job apocalypse will subside. It also appears that the Edge Lords of AI have begun to temper their predictions of apocalyptic consequences. Perhaps they have concluded that the benefits of juicing the stock are outweighed by the consequences of enraging the public.

The wave of talk about the imminent arrival of AGI—Artificial General Intelligence—which would render humans unnecessary,has also subsided dramatically, as it becomes clearer that AGI is not going to be achievable simply by scaling LLMs—Large Language Models—which are what power all of today’s AI best-known applications. AGI may not be achievable, if at all, for a very long time, meaning that more narrowly-talented AI models, working as partners with humans, may be the winning model. In that case, again, fear will subside.

Even a big fat AI crash, which burns a lot of reckless investors, kills off some companies, tanks the stock markets for a while, and flushes some of the least-loved leaders into early retirement might help the public view of AI, simply by proving it mortal, and therefore less of an existential threat. If at the same time governments are able to tame the worst excesses of the datacenter building boom, tempers may cool. Or not. The next few years will tell us a lot.