You begin to see them when you drive out toward Prince William County, Virginia, beyond the strip malls, commuter subdivisions, and the area where the landscaping gives way to cleared flat ground. On a peaceful morning, the buzz of industrial cooling units operating in rows outside the walls of enormous low grey buildings set back from the road could be heard from a hundred yards away. The structures were ringed by security fence. Data centers are what these are.
As it happens, Northern Virginia has more of them than any other place in the world, and more are being constructed. These are not your typical data centers; some of the biggest and newest are owned by Meta. These facilities, which the business now refers to as “superintelligence hubs,” house tens of thousands of GPUs that are constantly operating and training the massive language models on which Meta is placing its bets for the next ten years. The cooling units must be noisy.
The hidden environmental cost of Meta’s superintelligence hubs starts with electricity, particularly its quantity and source when the renewable supply is insufficient. Meta has made large investments in solar and wind power and discusses them in sustainability reports with the careful attention of a business conscious of its audience. However, powering GPU clusters of this magnitude necessitates a degree of reliable, dispatchable electricity that is currently unattainable from renewable sources alone, regardless of the time of day or weather.
Natural gas and sometimes coal replace the void on the regional grid when the wind stops blowing and solar output decreases. The lack of renewable energy doesn’t cause the facilities to shut down. The carbon accounting subtly changes while they continue to operate. According to research, Meta’s AI operations use more than 100 megawatt-hours per million dollars of sales, which places it far higher than competitors like Apple in terms of energy intensity per dollar generated.
In certain respects, the water story is more readily apparent at the local level. The majority of these facilities control server temperatures by evaporative cooling, which removes massive amounts from nearby sources by pulling water through the system and enabling heat to escape as vapor. The addition of a Meta facility that draws hundreds of millions of gallons a year does not result in an abundance of slack capacity in areas of Texas and Arizona where baseline water stress is already severe and aquifer depletion is a known continuing issue. The nearby local wells have dried up.
As water districts handle the increased demand, local households’ utility costs have skyrocketed. These are not hypothetical predictions; rather, they are actual results that have actually happened in the vicinity of large-scale data center campuses, and the trend continues when additional facilities in comparable areas come online.
Due in part to the less direct causal chain, the air quality component typically receives less attention than power and water. PM2.5 and nitrogen oxide emissions build up through the supply chain for semiconductor manufacturing (the production of the chips used in Meta’s servers generates detectable particulate pollution) and indirect emissions from the non-renewable electricity used for model training.
Over the course of its training run, training a single large model like Llama can indirectly account for tons of PM2.5 and NOx, contributing to the decline in regional air quality and the associated healthcare expenditures. There is a conflict that is difficult to fully reconcile when one looks at the company’s sustainability communications in conjunction with these numbers: the language of green commitment coexists with data that characterizes one of the world’s most energy-intensive computing operations.
The least noticeable and arguably most permanent aspect of the toll is the clearing of land. Hundreds of acres are needed for these campuses, and the locations chosen for them are usually woodland or agricultural property on the outskirts of tiny towns. These locations were picked because they have more acreage available, are closer to electricity infrastructure, and require less permits than denser regions. There are no longer any woodlands.
The facilities bring with them light and auditory pollution, which changes the surrounding land’s environment in ways that are hard to quantify but regularly reported. It’s also uncertain if the industry will continue to grow under primarily voluntary guidelines or if federal regulatory frameworks will evolve quickly enough to demand true environmental accountability for AI infrastructure at this scale. In any case, the amenities continue to be constructed.
