91ºÚÁÏ

What does it cost to ask AI a question? Not in subscription fees, but in electricity, water, and rare earth minerals. These costs are real, they are growing, and they remain largely invisible.

AI technology has existed for decades, but generative AI (systems that create text, images, video, and code on demand) represents a fundamental . Since ChatGPT's launch in late 2022, these tools have moved from research labs into , and their environmental costs have scaled with them.

Yet conversations about AI tend to focus on capabilities and ethics while environmental dimensions receive far less attention. Part of this silence stems from difficulty in measurement: companies rarely disclose energy consumption, researchers lack standardized metrics, and the infrastructure supporting AI is largely invisible to end users. This resource provides a starting point for understanding AI's environmental footprint across its full lifecycle — from the mining of rare earth minerals to the disposal of hardware.

How to use this resource:

This resource follows AI's lifecycle journey from raw material extraction through manufacturing, training, and daily use. Questions about — who bears the costs, who makes the decisions, and whose communities absorb the harm — run throughout every stage of this journey. 

You don't need to read everything or follow a linear path. Choose the sections most relevant to your interests or questions; the resource is designed for self-directed exploration. Each section includes two tiers of sources: jumping off provides accessible overviews, while deeper dive is more technical and data-driven. This is a living document, and updated as new research emerges.

Before we unpack these impacts in detail, here are a few starting points for understanding the scale of AI’s physical footprint:

• Article: “” MIT Technology Review, James O’Donnel and Casey Crownhart, May 20, 2025.
• Article: “” Engadget, Daniel Cooper and Cheyenne MacDonald, October 3, 2025. 
• Video: "." The New York Times Opinion Video, Video by Kate Crawford, Ryan S. Jeffery and Adam Westbrook, September 26, 2025. 
• Podcast: “.” Uncanny Valley:WIRED, Lauren Goode, Michael Calore, and Zoe Schiffer, October 23, 2025.

Unlike a car’s fuel efficiency or refrigerator energy label, AI systems carry no standardized environmental reporting. Companies treat energy consumption and emissions data as proprietary information. Researchers work with incomplete data and estimates. The infrastructure spans global supply chains, making it nearly impossible to track the full picture.

The opacity isn't just a technical problem, is it also a political and ethical one. What we choose to measure shapes what we can address. Without agreed-upon metrics, we can't compare models, hold companies accountable, or decide which AI applications justify their environmental cost. This section introduces key frameworks and terminology, and explores why measurement remains so contested. The stakes of that contest are not abstract: without reliable data, the communities most affected by AI infrastructure have the least power to demand accountability.

Every AI model runs on physical hardware (processors, memory, storage systems) and that hardware carries substantial environmental costs long before it reaches a data center. The rare earth minerals essential for modern chips come from mining operations that devastate ecosystems and exploit labor. Manufacturing semiconductors requires extraordinary amounts of pure water, toxic chemicals, and energy. And unliking a building or a car, AI hardware has remarkable short lifespan. The chips are replaced after just one or three years, generating mounting streams of electronic waste. 

This material footprint is often invisible in conversations about AI's environmental impact, which tend to focus on energy use alone. But the environmental and human costs of extraction, fabrication, and disposal are often externalized to communities near mining sites, factories, and informal e-waste processing operations. This makes AI's hardware lifecycle a matter of global environmental justice. 

Behind every AI interaction sits massive physical infrastructure: data centers filled with servers that require enormous amounts of electricity to run and water to keep cool. The energy required for a single prompt may seem trivial, until it is multiplied across billions of daily queries. New facilities are increasingly being built in drought-stressed regions, where their demand competes directly with the drinking water and agricultural needs of local communities. These communtieis rarely have a seat at the table when decisions are made. 

Companies frequently cite renewable energy commitments, but these claims deserve scrutiny. Purchasing renewable energy credits is an accounting mechanism, not a direct connection to clean power in real time. Meanwhile, efficiency improvements at the individual facilities are often overwhelmed by the explosive growth in AI deployment (a data center might become more efficient but total resource consumption still climbs). Sustainable AI will require not just better technology, but transparent accountability for local impacts.

Training a large AI model is enormously resource-intensive. Thousands of specialized processors run continuously for weeks or months, consuming massive amounts of electricity and generating substantial carbon emissions. And training is rarley a one-time cost because models are frequently retrained updated, and fine-tuned for specific applications. Each iteration carries its own environmental cost.

The race toward ever-larger models reveals a tension. The efficiency improvements that make training feasible also enable the construction of even more resource-hungry systems. Without mandatory disclosure and environmental accounting, this cycle will likely continue. Understanding what training actually costs and who bears those costs is the starting point for any meaningful accountability. 

While training a model happens once or periodically, inference (the process of actually using a trained model to generate responses) happens billions of times per day. Individual querires may seem negligible, but at the scale of widespread deployment, inference can rival or exceed training as a source of environmental impact. And unlike training, inference never stops.

This is where Jevons' Paradox becomes relevant: efficiency improvements that reduce the cost of each query can paradoxically increase total consumption, as cheaper and faster AI gets used more, in more contexts. Technological efficiency alone can't solve the problem. The harder questions are about scale and purpose: which applications actually justify AI's resource costs, and where is it simply convenient?

AI's environmental footprint is not distributed equally. The minerals that make AI hardware possible are extracted primarily in the Global South, where mining operations contaminate water supplies and displace communities. Semiconductor manufacturing concentrates in areas with weaker labor and environmental protections. E-waste travels from wealthy consumers back to low-income countries where it is processed unsafely. Meanwhile, the benefits (the profits, the capabilities, the decision-making power) concentrate in wealthy nations far removed from these harms.

These aren't accidental outcomes. They follow historical patterns of how industrial development distributes benefits and burdens along lines of race, class, and geography. The communities most burdened by AI's material demands are frequently those least responsible for driving AI adoption and most vulnerable to the climate impacts that AI's energy consumption accelerates. Understanding this lifecycle geographically involved asking not just what AI costs, but who pays.