Read a sample: 1. The Physical Side of an AI Answer
You open an AI assistant and type a simple request: “Help me plan a vegetable garden for a small balcony.”
A moment later, suggestions appear on your screen. The answer might mention containers, sunlight, herbs, and watering. You see words arrive in a familiar interface, and the exchange can feel almost weightless. Nothing has been delivered to your door. No machine is visibly moving. From your point of view, the activity consists of typing, waiting, and reading.
Yet the answer depends on a physical system. Your phone or computer uses electricity to display the interface and send your request. Networking equipment carries information between your device and the computers providing the service. Those computers perform calculations using electronic components housed somewhere in the world. Power systems supply electricity, and cooling equipment helps keep the machines within operating conditions. Long before you asked about tomatoes or basil, factories, mines, construction sites, and transportation networks contributed materials and equipment to that system.
None of this means that asking one question is necessarily a major environmental event. It means that an AI answer is not separate from the material world. The visible interaction is a small window into a much larger arrangement of shared infrastructure.
Following that arrangement is the starting point for understanding AI’s environmental footprint. Along the way, we will encounter several kinds of resource use: electricity used while equipment operates, water used in some cooling and electricity-generation processes, materials used to make hardware, and the energy and resources involved in building and replacing it. We will also encounter a harder question: how much of that activity should be attributed to one request?
The journey begins with your device. Whether you use a phone, tablet, laptop, or desktop computer, the interface runs on physical electronics. The screen displays your words, a processor handles local tasks, and a network connection allows the device to communicate with the service.
Your device would use some electricity even if you were doing something else. That matters when thinking about responsibility for an individual request. It would be misleading to assign all the electricity your laptop uses during an afternoon to one AI question. On the other hand, treating the device as if it had no physical presence would also leave part of the picture out. What belongs in the calculation depends on what the calculation is trying to show.
After you submit your gardening question, the text is converted into digital information and sent across a network. It may travel first through a Wi-Fi router or a cellular connection, then through equipment operated by internet and telecommunications providers. Along the way, it passes through devices that direct information toward its destination.
The network is not a single wire running from your screen to an AI computer. It is a collection of connected systems: cables, routers, switches, radio equipment, and other infrastructure. Much of that equipment serves many purposes at once. Your gardening question shares a world of infrastructure with video calls, online shopping, work messages, and countless other activities.
Those shared systems use electricity and require equipment. But they do not necessarily spring into action solely for your question. A router may already be powered on; a network link may be carrying many users’ data. This makes it difficult to draw a neat line between the power a system uses to be available and the additional activity associated with a particular request.
The physical distance your information travels can matter too, but the path is not always obvious to the user. The service may direct a request to a facility based on available capacity, system design, or other operational considerations. The location of a company’s headquarters does not tell you where a particular answer was produced.
At some point, your request reaches computers that can run the AI model. These machines are commonly housed in data centers: facilities designed to contain computing equipment and provide the power, connections, and environmental conditions it needs.
A data center is more than a room full of computers. It includes equipment to distribute electricity, networking hardware to move information, and systems to remove heat. Depending on the facility, it may also include backup power equipment, water infrastructure, and monitoring systems. Some data centers support many kinds of digital services, not just AI. Even where AI is a prominent workload, the facility’s total resource use cannot automatically be treated as the footprint of a single model or application.
Inside the computing equipment, processors perform the calculations needed to produce an answer. AI services may use specialized chips designed to handle the kinds of mathematical operations common in these models. Memory holds information the processors need, and connections move that information among components. The processors do not work alone: the surrounding hardware and software help turn a question into a usable response.
For a text-generating model, the request is processed and the response is produced in pieces. The model uses patterns represented in its learned parameters to predict what text should come next, given the information it has received and generated so far. Those repeated calculations continue until the answer is complete or the system stops generating.
Some services do more than run a single model once. They may search for information, examine an image, retrieve stored data, check an output, or call other software. A seemingly simple question can therefore involve different amounts and kinds of computing depending on how…