A real cost that is hard to measure
AI systems run on computers in data centers, and those computers need electricity, cooling and physical hardware. That much is not in dispute. What is hard is putting a reliable number on it for any particular tool or task. Published estimates vary widely, depend on assumptions about hardware, location and usage, and sometimes come from sources with an interest in a particular conclusion.
This article does not offer figures, because any single number would give a false sense of precision. Instead it describes where the resource use comes from, how to think about it, and what organizations can reasonably do. It is general information rather than an environmental assessment.
Where the resources go
Building the model. Creating a large model typically involves running many specialized chips for an extended period, often weeks or longer, on large collections of training data. This is a significant one-time (or periodic) cost for the developer, and it is mostly invisible to the people who later use the product.
Running the model. Every time someone submits a request, servers do computation to produce a response. A single request is small, but widely used services handle enormous volumes, so the total can be large. Longer inputs and outputs, and larger models, generally use more computation than short ones.
Cooling and facilities. Data centers use energy to keep equipment cool and, in many designs, water for cooling. How much depends heavily on climate, design and the efficiency of the facility.
Hardware manufacturing and disposal. Chips, servers and networking equipment require mining, manufacturing and transport, and they eventually become electronic waste. These upstream and downstream effects are sometimes overlooked when only electricity is counted.
Why numbers differ
Several factors make comparisons tricky.
- The energy source. The same computation has a different climate impact on a grid dominated by renewables than on one dominated by fossil fuels. Location and timing matter.
- Efficiency changes quickly. Hardware and software improve, so figures that were reasonable for older systems may not describe current ones.
- System boundaries. Some estimates count only the electricity for the model, while others include cooling, networking, manufacturing and the devices used to access the service.
- Task differences. Generating a short answer, an image or a long video does not require the same effort, and tasks vary widely.
- Efficiency versus total demand. Even if each task becomes cheaper, more use can offset the savings. Whether total consumption rises or falls depends on how adoption grows.
Be cautious about claims that AI is either negligible or catastrophic. The accurate position is usually that it is a growing but uncertain share of a larger digital footprint, and that it varies by use.
The other side of the ledger
AI can also reduce resource use in some settings. Examples people point to include energy management in buildings, optimizing delivery routes, spotting equipment problems early through predictive maintenance, and reducing waste in production. These benefits are real in some cases but are not automatic, and results depend on how well the tool is implemented. It is reasonable to ask for evidence in your own context rather than assume a net gain.
It is also fair to compare AI with alternatives. A task done by AI may replace a trip, a manual process or a less efficient system, or it may simply add computation on top of what was already happening.
What organizations can reasonably do
You do not control how a vendor builds its infrastructure, but you have more influence than it may seem.
- Match the tool to the task. A simple rules-based automation or a basic search may do a job with far less computation than a large generative system. Choosing the right tool is also part of avoiding poor-fit tasks.
- Avoid wasteful use. Repeatedly regenerating content, running automated jobs nobody reads, or leaving unused experiments running adds up.
- Ask vendors questions. Many providers publish information about energy sourcing, efficiency and data center practices. Asking is a legitimate part of vendor evaluation, and the answers can feed into the total cost picture, which can include more than the subscription price.
- Prefer smaller models when they suffice. For narrow, routine tasks, a smaller system may perform well enough.
- Retire unused tools. Forgotten pilots and duplicate subscriptions can quietly keep consuming resources.
- Include the footprint in your risk review. A line in your assessment process is usually enough to prompt a short discussion.
Treat vendor sustainability claims with the same care as any marketing: look for specifics, not slogans, and be aware that offsets and credits are not the same as lower consumption.
Keeping a sense of proportion
For most small organizations, the footprint of their AI use is likely to be small compared with other parts of their operations, such as travel, heating and equipment. For large-scale deployments it can be more significant. Either way, the question is worth asking, not worth panicking over. Reasonable steps are to be deliberate about where AI is genuinely useful, to avoid waste, and to revisit the issue as better information becomes available.