You have likely heard on campus about the “immense environmental cost” of AI. Student activists attribute these costs to three main characteristics of data centers. The first is that data centers use too much water. The Harvard Political Review made this argument, for instance, in a December article about a planned data center near Reno, Nevada. The second is that data centers use too much energy. When The Crimson Editorial Board told Harvard in May to “jump off the AI bandwagon,” they cited this reason as one of their primary motivations. The third is that data centers supposedly harm poor communities that live nearby. Two editors from Memphis wrote in The Crimson last September that Harvard’s use of AI harms communities like theirs due to the building of data centers. The Harvard Law Review similarly argued that fast-track permitting leaves nearby communities to “bear the burden” of data center construction.
Granted, data centers do have some environmental impacts. Yet every one of these articles places the harm in a town far from Cambridge and the remedy in the reader’s own habits. The Crimson’s September article asked Harvard to lead in “using AI sparingly, thoughtfully, and only when it truly advances academic inquiry.” The Political Review told its readers to demand transparency from Big Tech. The Crimson Editorial Board told the University to get off the “AI bandwagon” altogether. What is notable is how little any of them say about the data center Harvard already runs or about the emissions research Harvard has already published on the topic.
Once one looks at either, the case these papers are trying to make falls apart. The water and power figures these writers rely on are inflated. Harvard’s own scientists have found that a data center’s emissions depend on where it is built, meaning that giving companies greater autonomy on where to build data centers will only help the environment. Thus, the bans and moratoria that anti-AI activists demand push new data centers toward dirtier grids and worsen the problem these activists are supposedly protesting. As proof that AI data centers can often be clean, Harvard has run a data center in Holyoke, Massachusetts, on clean hydropower for more than a decade in a poor city that has gained from the presence of that data center. That facility is the best answer to the critics, and it is a model for how the rest of the country’s data centers should be run.
Let’s start by discussing water. A Crimson Magazine feature on a data center in Lowell, Massachusetts told readers that a single AI facility can use up to five million gallons of water a day. While it is technically true that there exist AI facilities that could use this much water, this is an extremely misleading figure. No actual AI facilities work that way. The Lowell data center, for instance, reports using between 60,000 and 120,000 gallons of water a day. Even more misleading, the Memphis op-ed said a specific data center, xAI’s Colossus, would use up to a million gallons a day. That is simply false. The Daily Memphian pulled the 2025 utility bills and found a peak of 381,000 gallons in one day, much less than the quoted figure. Harvard’s own data center in Holyoke, meanwhile, uses about 25,000 gallons of water a day, roughly what a small hotel uses.
Next, energy. The Crimson Editorial Board seems to think the energy used when prompting ChatGPT is a significant enough reason for the school to demand that students never learn how to use AI. One would think, based on the Editorial Board’s apoplexy, that AI usage would lead to something catastrophic like a massive depletion of the international fossil fuel supply. But that just isn’t true. Google reports that a typical text prompt to Gemini uses just 0.24 watt-hours, roughly what an LED bulb uses in two minutes. One student’s chatbot use will likely barely make up one percent of their personal energy consumption, let alone making a dent in national energy consumption.
The real energy consumption, instead, goes into the machines that train and run these models. While The Crimson Editorial Board chides students for using AI, Harvard has procured a large set of gas-guzzling AI trainers, the things that really use up energy. In March, Harvard’s Kempner Institute announced that its AI cluster, a bank of linked computers built to train models, had grown to 1,144 GPUs. GPUs are the specialized chips that do the actual computing, and they draw most of a data center’s power. In 2023, the cluster had 144 of them, so it has grown nearly eightfold in just three years.
Harvard has not published how much energy the cluster draws, but the building it is housed in is permitted for 12 megawatts. Say the cluster uses just one. An hour of that equals about four million prompts. If every undergraduate at the College sent fifty prompts a day, the whole College would use in a month what the cluster uses before lunch.1 You could quit AI for all four years, the school could ban student use of AI for anything whatsoever, and the University’s footprint would not change at all. The critics are focusing on the wrong thing. The individual energy use involved in AI is tiny, while Harvard’s institutional use is large. If the concern were really about Harvard’s impact on the environment, the cluster would be the first thing discussed. Instead, The Crimson Editorial Board did not even mention Harvard’s AI training cluster.
But The Crimson did not just make the basic-level mistake of not knowing what AI is or how it consumes energy. It also made the bigger mistake of not understanding how little energy AI uses in the first place. In November 2024, a team at the T.H. Chan School of Public Health under biostatistician Francesca Dominici published the first facility-level study of American data center emissions, covering 2,132 sites. According to their findings, data centers use just over four percent of the nation’s electricity.
The study’s more useful finding, however, concerns carbon intensity, meaning how much carbon is produced per unit of power used. Data centers came in at 48 percent worse than the national average, since 95 percent of centers sit on grids “dirtier” than the country as a whole, in places where coal and gas still dominate. In other words, the upshot of the study was that the environmental effects of data centers depend largely on where the servers are, not on how often people use them. Even if Harvard shut down Kempner and severely limited AI usage on campus, the Virginia or Kentucky grids would remain as fossil fuel-emitting as they were yesterday, and the total energy use of data centers would barely change. The same research shows that moving a data center to a clean grid would roughly halve its emissions.
Harvard solved this location problem in 2009, when it and four other universities went looking for cheap, clean power and found Holyoke. The city’s municipal utility owned a hydroelectric dam at Hadley Falls, built for paper mills that had long since closed. Its clean power had few buyers left. So these universities stepped in and spent about $95 million to put the Massachusetts Green High Performance Computing Center on the demolished mill site. Nowadays, about 90 percent of Holyoke’s electricity is carbon-free. The server room runs at 80 degrees to save on cooling, and for eight months of the year, outside air handles the temperature. Racks there cost Harvard about $7,000 a year against $28,000 in Boston. Harvard’s Holyoke facility shows that data centers need not be environmentally unfriendly. In fact, the only reason they tend to be energy-guzzling is that misguided environmentalists force them to only be built in areas where all electricity emits significant carbon.
Holyoke is also notable in that it is exactly the kind of low-income, majority-minority town the “environmental justice” literature criticizing data centers worries about. It is 52 percent Hispanic, and its poverty rate is 27 percent, double the state’s. What the data center brought was lower taxes, coding classes for local kids, and, as of this year, a $31 million state AI hub on a previously moribund dam. The city’s own view of the arrangement is telling. While Holyoke this year became the first city in Massachusetts to ban data centers, this ban had one conspicuous exception, written into the ordinance by name: the center Harvard built. The council president said most residents did not even know it was there. The exemption gives the game away. Holyoke did not decide data centers were bad. It decided it liked its data center so much it wanted to give it a monopoly.
Harvard needs data centers like Holyoke’s. Researchers at the facility sequence genomes, model the spread of disease, and run the climate simulations on which the environmentalists’ own work depends. The Kempner Institute is equally indispensable, training the foundational models that let academic labs do work that would otherwise only be possible for companies like Google and Meta. AI built on this hardware is already reading cancer pathology slides and finding new uses for existing drugs. Harvard needs these centers if it wants to remain a medical research center, if it cares about its basic mission of research and education. The question, if Harvard views itself as a university and not an activist workshop, is not whether to use data centers, but where and how to build them.
Holyoke provides an answer to this question beyond just Massachusetts. It shows how to create the data centers necessary to grow our economy and defend our nation without destroying our society. Spare power capacity is not unique to New England. In 2024, Texas curtailed more than eight terawatt-hours of wind and solar because the turbines produced more than the transmission lines could carry east; that power was lost to the wind. Grid operators in the Midwest and Mid-Atlantic report similar patterns. With a bit of planning, we could build data centers that could soak up that extra power, easily pay the Americans who live in these sparsely populated counties, and, for anyone worried about emissions, keep the load off a coal-heavy grid.
Luckily, Harvard has built the tools necessary to know where these data centers should be built. In June, two Harvard researchers won a Harvard Ascend award to build a tool that tells developers and towns where a data center belongs, where it can bring the most good without destroying local communities or raising gas prices. Harvard researchers have shown the way to the future of data center construction.
Harvard students, on the other hand, have ignored and ridiculed that future. Not knowing the nuances of data center construction or even how AI energy consumption works, students have put forth false claims that only distract from the true question — not whether to build data centers but how to build them. More centrally, these students’ failings show a fundamental problem in the progressive project — its focus on broad claims about social justice over particular claims about how to serve real people.
This same progressive mistake is playing out across Massachusetts. In 2024, the state rightfully passed a tax incentive to bring data centers onto one of the cleanest grids in the country. Since then, activists have pushed Lowell and Westfield to pass moratoria, and Governor Maura Healey has frozen the incentive. In September, she signed an order letting towns veto a data center project that had started anywhere nearby at any time for any reason. These sanctimonious policies ignore that servers are still going to be built. Meta plans to spend up to $145 billion on infrastructure this year; its Louisiana campus will run five gigawatts on ten new gas plants. Capital goes where it is wanted, and the places that want it tend to have “dirtier” grids, leading to substantially worse environmental effects than if the moratoria hadn’t been imposed in the first place.
Take away the misleading water figures and the prompt counts that do not matter, and the environmental case against students’ AI use has little left. Its strongest version, that data centers belong where the power is clean, is one Harvard already acted on in Holyoke. What remains is unreasoned anxiety: a radical environmentalism that risks leaving the environment worse off. Harvard should recognize the spurious nature of these complaints and disregard them completely.
Author’s calculation from the previously cited figures, assuming a one-megawatt draw.


Don’t know why my comment posted before I finished. The need for a duplicate backup continuously reliable system capable of providing peak power for weeks at a time makes wind and solar uncompetitive. Battery backup at grid scale is not close to affordable or practical.
Data centers developers are well aware of the energy and water issues and could not obtain financing if they do not have persuasive answers for them.
Good article but it does not mention that data centers are increasingly using air cooled or closed loop water cooling systems with waste water as the input. These systems may use more electricity than evaporative water cooling systems, but there is a great deal of innovation and discovery in industry’s efforts to minimize water and electricity impacts of data centers.
The point about moratoriums forcing data centers into dirty grids is trivial. While hydropower and nuclear are the cleanest electricity generators,, wind and solar are neither cleaner nor cheaper than fossil fuel plants when one considers the entire life cycle and