The Real Cost of Running AI: Water, Power, and Community Impact

The Infrastructure Boom Nobody Asked For

Shaolei Ren grew up in a coal-mining province in northern China during the 1980s, where black carbon settled on streets and families rationed water from storage tanks. Today, as an associate professor of electrical and computer engineering at the University of California, Riverside, he studies a different kind of infrastructure challenge: the environmental footprint of GPU-powered data centers spreading across the American landscape.

His research cuts through the abstract metrics of carbon emissions to focus on what happens to the communities that host these facilities. Air quality deteriorates. Water supplies tighten. Utility bills climb. And unlike the semiconductor fabs and fast-fashion supply chains concentrated in Asia, this wave of infrastructure is landing in backyards from Virginia to Arizona, making the costs impossible to ignore.

At DailyTechWire, we have tracked the data center buildout across North America for the past eighteen months. What began as a straightforward capacity story - more GPUs for more AI workloads - has evolved into something harder to parse: a clash between the promise of transformative technology and the tangible burden it places on local power grids, watersheds, and air quality.

The GPU's Expanding Appetite

The numbers tell a blunt story. GPU-accelerated AI servers in US data centers drew roughly 2 terawatt-hours in 2017. By 2023, that figure had surged past 40 terawatt-hours, according to a 2024 Lawrence Berkeley National Laboratory study. Projections for 2028 range between 165 and 326 terawatt-hours annually - the lower bound equivalent to the yearly consumption of more than 8.7 million American homes.

A 2025 analysis by Alex de Vries-Gao, a PhD candidate at Vrije Universiteit Amsterdam Institute for Environmental Studies, estimates that AI likely overtook Bitcoin mining in power consumption during the year, accounting for nearly half of global data center electricity use. The resulting carbon emissions sit somewhere between 32.6 million and 79.7 million tons per year, roughly comparable to New York City's annual climate footprint.

Gaming, by comparison, consumed about 34 terawatt-hours annually in the US as of a 2019 study - equivalent to the emissions of roughly 5 million cars. Newer consoles have pushed that figure higher, but the trajectory of AI infrastructure dwarfs it. The GPU, once a niche component for rendering polygons and textures, now anchors an entirely different energy regime.

Air Quality and the Hidden Health Toll

Ren and colleagues at UCR and Caltech published a preprint in 2024 estimating that training a model the size of Meta's Llama 3.1 generates air pollution equivalent to 10,000 round trips between Los Angeles and New York City by car. Public health costs associated with AI adoption could exceed $20 billion by 2028, the study found, with up to 1,300 premature deaths annually by 2030 linked to air pollution from data center operations.

Those costs do not distribute evenly. The NAACP filed suit against xAI - now operating as SpaceXAI - over air pollution from gas generators installed to power the company's data centers. The civil rights organization has signaled it will support similar campaigns nationwide, framing the issue as environmental justice: low-income neighborhoods and communities of color disproportionately bear the burden of siting decisions driven by land cost and zoning latitude.

Catherine Flick, a professor of ethics and games technology at the University of Staffordshire, notes that GPUs have always sat at the intersection of technological progress and ethical scrutiny. From their origins in arcade hardware in the 1970s through Nvidia's 1999 release of what it called the world's first GPU, these chips have powered applications - games, virtual reality, now generative AI - that reshape how people interact with the world. The question has never been whether GPUs enable new capabilities, but whether those capabilities justify the costs.

Water Scarcity and Demand Spikes

Power consumption is only part of the equation. Running GPUs around the clock while preventing hardware from overheating requires enormous volumes of water for cooling systems and electricity generation. De Vries-Gao estimates that AI consumed between 312.5 billion and 764.6 billion liters of water in 2025, roughly in line with global annual bottled water consumption. Ren's 2023 study projected demand could reach 600 billion liters by 2027.

Yet annual totals obscure the real strain, Ren argues. Water use at data centers is "spikey," peaking during hot weather when cooling demands surge. A typical home might use 1.5 to 2.5 times its baseline water during peak periods. A data center can hit six to ten times its baseline, creating sudden stress on local water districts precisely when drought conditions are most likely.

In Southern California, where Ren's university is located, the landscape is already dominated by fulfillment centers and logistics hubs serving e-commerce. The arrival of hyperscale data centers adds a new layer of infrastructure demand in a region where water scarcity is chronic. Retirees watch quiet neighborhoods transform into industrial corridors humming with generator noise and cooling systems. The parallel to warehouse sprawl is hard to miss: national advocacy groups focus on carbon emissions while residents organize around local air quality and noise.

The Value Proposition Problem

Some venture capitalists have blamed slow consumer adoption of AI on negative press about its environmental impact. Ashley Striblet, who works in product strategy and consumer AI, pushed back on that argument in a recent TikTok video, calling it an attempt to blame consumers for investment decisions that misjudged demand.

Flick finds the framing absurd. "Isn't it nice to have the environment as a scapegoat?" she says. Consumers tolerate the environmental costs of fast fashion or cheap electronics because they perceive clear value: lower prices, convenience, entertainment. AI has yet to deliver a comparable tradeoff for most people. "It'd be nice if we all had personalized butlers or whatever it is," Flick notes. "But how do you get from point A to point B? It's not with what we've got."

Striblet makes a similar observation. People understand that ordering clothes from Shein or Amazon carries environmental costs, yet they continue because the value is immediate and tangible. AI, by contrast, remains largely promissory. The infrastructure is being built - and the environmental costs are being incurred - before the use cases have proven themselves outside narrow verticals.

The question is not whether GPUs in iPhones or gaming PCs deserve scrutiny. They do. But the scale and pace of data center expansion for AI represent a different order of magnitude, with infrastructure decisions being made in real time by communities that may see little direct benefit.

A Community-Integrated Vision

Ren envisions what he calls a "community-integrated data center approach" that minimizes harm to nearby residents. His whiteboard at UCR is covered with equations exploring how to adjust data center operations to reduce resource consumption and pollution. "It's definitely doable," he says. But it is not the default.

The challenge is not purely technical. It requires aligning the incentives of hyperscalers, utilities, local governments, and residents - a coordination problem that has proven difficult even in less contentious industries. At present, the buildout continues at pace, driven by capital flows and competitive pressure among cloud providers and AI labs.

Whether the benefits of AI - advances in weather forecasting, drug discovery, wildlife conservation - ultimately justify the environmental and social costs remains an open question. What is clear is that those costs are no longer abstract. They are showing up in utility bills, air quality indices, and water district reports across the United States. The GPU, once a component for rendering game worlds, now underpins an infrastructure regime whose real-world impacts are becoming impossible to ignore.