Artificial intelligence is often described as immaterial—lines of code, floating in the cloud, conjuring predictions, poems, or portraits with a few keystrokes. Yet, the truth is far from weightless. Every large AI model is rooted in an immense material infrastructure: sprawling data centers, server racks, and energy-hungry GPUs that crunch and train on oceans of data. This invisible scaffolding comes at a significant ecological cost, one that is only beginning to be reckoned with as AI becomes central to our daily lives.
The largest AI models today are trained on billions, sometimes trillions, of parameters, a process that requires weeks or months of continuous computation. Each training run can consume as much energy as five or six cars do across their entire lifetimes, including fuel. That comparison is not metaphorical—it’s a stark reminder that “virtual intelligence” is powered by a very physical system. The carbon emissions generated by this training phase alone are staggering, and unlike renewable energy projects that offset their load, AI training often draws from the existing fossil-fuel-powered grid.
But training is only one part of the story. Once a model is trained, every query—whether it’s a chatbot reply, an image request, or a search suggestion—requires further computation. Multiply that by the millions of people using AI daily, and the cumulative energy draw becomes immense. While individual queries might seem negligible, at scale they add up to an enormous and ongoing environmental cost. AI, then, is not a one-time expense but a perpetually resource-intensive technology.
Water use is another often-overlooked dimension of AI’s footprint. Data centers generate immense heat, and cooling them requires extraordinary amounts of water. Studies have shown that training a large AI model can consume as much fresh water as a small town uses in a year. In drought-prone regions where many data centers are located, this raises serious ethical questions about resource allocation—should water cool GPUs, or should it irrigate crops and sustain communities?
The environmental burden is compounded by the geographic unevenness of these infrastructures. Many data centers are built in areas with cheaper electricity, land, and water, often in regions with vulnerable ecosystems or less stringent regulations. The communities near these centers may shoulder the environmental degradation while reaping few of the economic benefits. AI’s ecological toll, then, is not evenly distributed—it is entangled with global inequalities and patterns of exploitation.
There is also the issue of hardware itself. GPUs, TPUs, and other specialized chips are made from rare earth minerals and metals whose extraction is environmentally destructive. Mining cobalt, lithium, and nickel leaves deep scars on landscapes, contaminates waterways, and often involves exploitative labor practices. The glamour of artificial intelligence is thus underwritten by extractive industries that echo the darker side of technological progress.
Some argue that AI might eventually pay back its ecological debt by helping us fight climate change. Optimists point to AI-assisted energy grids, climate modeling, and agricultural efficiencies as potential benefits. Yet, these applications coexist with the reality that the industry’s current trajectory is toward bigger, more resource-intensive models, each racing to outsize the last. It is not clear that efficiency gains from “AI for climate” will outweigh the costs of AI itself.
The companies building these models are beginning to acknowledge the problem. Some have committed to carbon neutrality, or to powering data centers with renewable energy. Yet transparency remains a serious issue—few companies disclose the true carbon and water costs of their models. Without accountability, the conversation risks being more about green branding than genuine sustainability.
What’s at stake is not whether AI will continue—it will—but how it evolves. Will efficiency and sustainability become central design principles, or will the industry continue to prioritize scale and speed above all else? Regulation, public pressure, and consumer awareness may ultimately push the sector toward greener practices. But unless sustainability becomes embedded in the very culture of AI development, the environmental costs will continue to mount.
In the end, AI forces us to confront an uncomfortable truth: the digital world is not detached from the natural one. The “cloud” is not a celestial abstraction but a network of very earthly resources, extracting energy, water, and minerals at enormous scales. As we rush toward an AI-driven future, the challenge will be to align technological ambition with ecological responsibility—before intelligence, artificial or otherwise, becomes too costly for the planet to bear.