"It looks easy enough," Bloomberg's reporters wrote in one of the most widely cited explainers on artificial intelligence's hidden costs. "Ask ChatGPT something, and it responds." But, they continued, pull back the curtain and you find that every prompt consumes vast resources: millions of human beings engineering and correcting models, enough electricity to power entire countries, sprawling data center megacampuses, power lines and internet cables stretching across continents, and water, land, metals and minerals consumed at a scale most users never consider. Bloomberg's framing was deliberately stark and has only grown more urgent in the time since it was first published: artificial intelligence needs it all, and it will need more.

That warning is no longer an abstract prediction. It describes a resource competition that is actively playing out across the United States and around the world right now, straining electrical grids, draining water supplies in drought-prone regions, and creating genuine bottlenecks in the supply of metals most people have never heard of. Understanding what each of these resource categories actually means in practical terms, and what experts say about whether this trajectory can be made sustainable, matters considerably more than the optimistic talking points that typically accompany AI product launches.

What "More Data, More Chips, More Real Estate" Actually Means

Each item in Bloomberg's litany corresponds to a concrete, physical reality rather than an abstraction. The latest generation of AI products relies on large language models trained on enormous quantities of text, and tech companies are increasingly turning to synthetic data, content generated by AI itself, to feed that bottomless appetite while sidestepping some of the legal and ethical concerns tied to scraping content from the open web. Researchers have warned this approach carries its own risk, with studies on so-called model collapse showing that AI-generated outputs can become increasingly distorted when models are repeatedly retrained on their own previous creations rather than genuine human-produced content.

The chip and real estate dimensions of this resource appetite are visible in the sheer physical scale of recent buildout. Server racks that once held eight graphics processing units now commonly hold 72, requiring roughly 150 kilowatts of power per rack, and Nvidia's newest Rubin GPU and rack systems are expected to require around 300 kilowatts when they arrive later this year. Industry insiders are already bracing for future chip generations that could push individual racks toward a full megawatt of power consumption, enough on its own to power roughly 750 average American homes. This is precisely why AI companies have been racing to acquire enormous tracts of land for data center megacampuses, since the physical footprint required to house, power and cool this density of computing hardware has grown dramatically with each new chip generation.

The Water Connection: From Chip Fabrication to Data Center Cooling

Water enters this story at two distinct points in AI's supply chain, and both have drawn increasing scrutiny from researchers and journalists. The first is semiconductor fabrication itself, where producing the advanced chips powering AI systems requires enormous volumes of ultrapure water for the repeated cleaning steps involved in chipmaking, a process that has made major fabrication hubs like Taiwan and Arizona genuinely vulnerable to drought-driven supply disruptions in recent years. The second, and the one drawing the most direct public attention, is data center cooling itself, where the dense, high-powered server racks described above generate enormous heat that must be continuously removed to keep the hardware functioning.

The scale of this cooling-related water consumption is substantial and well documented. According to an April 2025 report from the International Energy Agency, an average 100-megawatt data center, which draws more power than 75,000 homes combined, also consumes approximately 2 million liters of water per day, equivalent to the water use of roughly 6,500 households. Many data centers rely on evaporative cooling methods, sometimes called swamp cooling, in which warm air is drawn through wet pads, with researcher Shaolei Ren of the University of California, Riverside noting that data centers typically evaporate about 80 percent of the water they draw, discharging only the remaining 20 percent back to wastewater treatment. Globally, the IEA estimates data centers already consume approximately 560 billion liters of water annually, a figure the agency projects could roughly double by 2030 as tech companies push for bigger facilities stocked with increasingly powerful, hot-running AI chips.

What makes this water consumption particularly fraught is where it is happening. Bloomberg's own analysis found that roughly two-thirds of new data centers built or in development since 2022 are located in places already experiencing high levels of water stress, with five states alone accounting for 72 percent of new centers built in these high-stress regions. Hydrologist Amy Bush, working near Abilene, Texas, where OpenAI is building a massive 1.2-gigawatt data center campus as part of its Stargate infrastructure venture, described the situation bluntly: every part of the state is facing what she calls a water-energy nexus crisis. This pattern extends well beyond the United States, with Bloomberg noting that Saudi Arabia, the United Arab Emirates, India and China are all building significant new data center capacity in dry regions where climate change is already pushing local water supplies toward their limits.

The Electricity Crisis Already Straining Global Grids

Power consumption represents perhaps the most acute and immediate resource constraint AI now faces. Researchers have estimated that a single ChatGPT query requires nearly ten times as much electricity to process as a traditional Google search, a gap that reflects the fundamentally different computational approach large language models require compared with conventional search indexing. Multiply that gap across billions of daily AI interactions worldwide, and the resulting electricity demand becomes difficult to overstate.

Elon Musk captured the severity of the looming bottleneck directly, warning earlier this year that the industry would very soon be producing more AI chips than it could actually power on, a statement that reflects a growing consensus among industry leaders that electricity generation, not chip manufacturing, has become the binding constraint on how fast AI capacity can actually expand. Currently, roughly 30 percent of the power flowing into data centers is not used for AI computation at all, according to Nvidia, with much of that share consumed by cooling systems and the electricity lost as power travels long distances across sprawling data center campuses. This inefficiency has prompted major hardware investments, including the shift toward direct-to-chip liquid cooling, which Nvidia and power equipment maker Vertiv have found can improve a data center's energy efficiency by roughly 15 percent and cut emissions tied to fossil-fuel-generated electricity by about 10 percent.

The political consequences of this electricity demand are already visible. Rampant power consumption from what industry insiders now call AI factories threatens to push up electricity prices for ordinary households and businesses sharing the same grids, and political backlash against new data center construction has already begun creating friction in communities asked to absorb the infrastructure without seeing proportional local benefit. Arizona State University professor and former Mesa, Arizona water director Kathryn Sorensen posed the question many of these communities are now asking directly: whether the increase in tax revenue and the relatively modest number of jobs these facilities actually create is worth the water and power they consume.

The Hidden Bottlenecks: Metals, Minerals and Microscopic Components

Beyond water and electricity, Bloomberg's reporting has highlighted a less visible but increasingly consequential category of resource constraint: the raw materials and microscopic components that make modern AI hardware function at all. Two obscure metals, gallium and germanium, have emerged as potential bottlenecks, and the stakes became starkly clear when China announced export restrictions on both materials as part of an escalating technology dispute with the United States. Copper, used throughout chips, data centers, electrical equipment and cooling systems, sits at the center of a looming three-way competition between the demands of AI infrastructure, renewable energy buildout and electric vehicle production, all of which depend on the same finite global copper supply. Steel, critical to data center construction and the broader infrastructure connecting these facilities, faces similar pressure as construction accelerates simultaneously across multiple continents.

Even at the smallest physical scale, bottlenecks are multiplying. Server racks that once required dozens of fiber-optic connections now require tens of thousands, and Bloomberg Businessweek's reporting on the components powering AI hardware describes an industry where microscopic motors spin tiny fans and threads of liquid thinner than a strand of hair run through copper plates to manage the heat of computation, with bottlenecks appearing across everything from memory chips to battery systems as each successive generation of AI chip demands ever more of these tiny, precisely manufactured parts.

What Experts Say About Whether This Can Be Made Sustainable

The honest answer experts offer is neither uniformly optimistic nor uniformly bleak, and it depends heavily on the pace of efficiency gains relative to the pace of overall AI adoption. Gary Dickerson, chief executive of chip equipment maker Applied Materials, has described efficiency as an emerging unifying force across the entire industry, telling investors that some AI companies are targeting hundredfold improvements in computing efficiency within five years, with others pursuing gains as large as ten thousandfold within fifteen years. Nvidia's own hardware has demonstrated real progress along these lines, with its Blackwell chip increasing processing capacity while consuming the same amount of energy as its predecessor, representing a genuine leap in energy efficiency per unit of computation delivered.

At the same time, researchers studying the water and energy dimensions of this buildout are notably less reassured. Newsha Ajami of Lawrence Berkeley National Laboratory has described the resource strain as a rapidly growing issue spreading everywhere, not a contained or isolated problem. The fundamental tension experts keep returning to is that efficiency gains per unit of computation do not automatically translate into reduced total resource consumption when the volume of AI usage is expanding even faster than those efficiency improvements can offset. Bloomberg's own reporting captures this tension precisely: AI's resource demands are simultaneously motivating companies to pour billions into genuinely promising alternative energy solutions, including long-stalled nuclear fusion research now receiving renewed investment, while also adding fresh pressure to keep burning fossil fuels to meet near-term grid demand, even as the world is already on track to miss critical climate targets.

What Realistic Global Coordination Would Actually Require

Addressing AI's resource appetite at the scale this buildout now demands would require coordination across several distinct fronts simultaneously, and experts studying the issue are clear that no single intervention will be sufficient on its own. Water-stressed regions need data center siting decisions that genuinely account for local water availability rather than treating water as, in the words of water consulting firm managing member Sharlene Leurig, the last consideration in facility planning because it is cheap relative to real estate and power costs. Some companies have already begun responding to this pressure directly, with Microsoft notably rethinking the pace of its own data center expansion and pulling back on certain projects amid growing scrutiny of resource impacts.

On the energy side, genuine coordination would require utilities, regulators and AI companies to plan grid capacity expansion years ahead of demand rather than reactively, alongside continued investment in the kind of efficiency technology, liquid cooling, optimized power conversion, more efficient chip architectures, that can meaningfully reduce the energy required per unit of AI output delivered. On the materials side, addressing bottlenecks in gallium, germanium, copper and steel will likely require diversifying supply chains away from single points of geopolitical failure, expanding recycling and materials recovery programs, and continued research into alternative materials that can reduce dependence on the most constrained inputs. None of this is simple, and none of it can be accomplished by any single company or country acting alone, given how thoroughly AI's supply chain now spans chip fabrication in Taiwan, rare metal processing concentrated heavily in China, data center construction across the United States, Gulf states and Asia, and power generation decisions made by utilities and regulators operating under entirely different national frameworks.

The Bottom Line

Bloomberg's blunt framing, AI wants more data, more chips, more real estate, more power, more water, more everything, has proven less like hyperbole and more like an accurate preview of where this technology's resource demands were actually headed. The water stress documented in two-thirds of new data center locations, the electricity strain serious enough that industry leaders now warn power generation, not chip supply, is the binding constraint on AI's growth, and the emerging bottlenecks in obscure but essential metals all point toward the same underlying reality: artificial intelligence's environmental and resource footprint has moved from a theoretical concern into a present, measurable strain on the planet's physical systems. Whether efficiency gains can outpace adoption growth fast enough to bend this trajectory toward something genuinely sustainable remains a live, unresolved question, one that depends less on any single technological breakthrough and more on whether the companies, governments and communities involved can coordinate fast enough to keep pace with a technology that, by nearly every measure available, continues wanting more.

*This article is for informational and educational purposes only. Data and quotes are sourced from Bloomberg News and Bloomberg Businessweek, the International Energy Agency, TechRepublic, and Lawrence Berkeley National Laboratory.*