AI is becoming one of the most powerful technologies available to businesses.
It can write content, analyse documents, answer customer questions, automate workflows, generate software, summarise meetings and help employees make decisions.
But there is another side to the AI revolution that businesses need to understand.
AI consumes resources.
Behind every AI application are data centres, servers, GPUs, networking equipment, cooling systems, electricity, water, semiconductor manufacturing and eventually electronic waste.
The question is not whether we should stop using AI.
The better question is:
How do we get more business value from AI while using fewer resources?
That is the idea behind Sustainable AI.
AI Has an Environmental Footprint
It is tempting to think of AI as software.
After all, we interact with AI through a browser or mobile application.
But AI doesn't run in the cloud in some magical, weightless environment.
It runs on physical infrastructure.
A typical AI ecosystem involves:
Data → Servers → GPUs/AI accelerators → Storage → Networking → Cooling → Electricity → Water → Hardware manufacturing → Eventually e-waste
The United Nations Environment Programme recommends looking at AI across its entire lifecycle, including data preparation, model development, training, deployment, inference, hardware production, data-centre construction and disposal.
This is important because focusing only on electricity used by the GPU gives us an incomplete picture.
How Much Energy Does AI Use?
The answer is:
It depends.
A simple text query and a complex AI reasoning, image-generation or video-generation workload can have dramatically different resource requirements.
The IEA's latest analysis makes this distinction particularly important. Energy consumption per AI task has been falling rapidly because of improvements in hardware and software efficiency. At the same time, increasingly sophisticated applications such as video generation, reasoning and agentic AI can consume hundreds or even thousands of times more energy per query than simple text generation.
So the environmental problem isn't simply:
"AI uses too much energy."
It is more accurately:
"AI usage is growing rapidly, while some of the new AI workloads are becoming much more computationally intensive."
The scale matters.
The IEA estimates that global data-centre electricity consumption was around 415 TWh in 2024, about 1.5% of global electricity consumption. Its 2025 analysis projected data-centre electricity consumption could reach around 945 TWh by 2030 in its base case.
And the latest IEA analysis reports that global data-centre electricity demand increased 17% in 2025, while electricity consumption from AI-focused data centres increased even faster.
This is why efficiency matters.
But What About Water?
This is where the discussion becomes more complicated.
AI doesn't literally "drink" water.
Water is primarily associated with the infrastructure supporting computation.
Data centres generate enormous amounts of heat. Depending on the cooling system, water may be used directly for cooling.
There is also indirect water consumption associated with generating the electricity used by the data centre.
And there is another often-forgotten component:
Semiconductor manufacturing itself can require substantial amounts of water.
The OECD identifies these three areas as important parts of AI's water footprint: cooling, electricity generation and semiconductor production.
This means that saying:
"One AI question uses X litres of water"
can be misleading if presented as a universal number.
The actual amount depends on the model, workload, hardware, data centre, cooling system, electricity mix and location.
A Real-World Measurement Is More Useful
In 2025, Google published a detailed measurement of AI serving using production infrastructure for Gemini.
Its methodology included not just the active AI accelerator but also CPU/RAM, idle capacity and data-centre overhead.
Google estimated that the median Gemini Apps text prompt consumed:
- 0.24 Wh of energy
- 0.03 grams of CO₂e
- 0.26 mL of water
Google also reported that the energy consumption of its median text prompt had fallen 33x over the preceding year, while its carbon footprint had fallen 44x.
That is an important lesson.
AI efficiency can improve dramatically.
But efficiency improvements alone don't solve the problem if usage grows even faster.
This is the classic rebound effect:
If something becomes cheaper and easier to use, people may simply use much more of it.
The Biggest Environmental Mistake Businesses Can Make
The biggest mistake isn't using AI.
It is using more AI than the business problem requires.
Imagine a company wants to classify incoming customer emails.
There are several possible approaches:
Option 1
Send every email to the largest available reasoning model.
Option 2
Use simple rules for obvious cases and AI only when necessary.
Option 3
Use a small model for classification and escalate complicated cases to a larger model.
Option 3 could provide essentially the same business outcome with substantially less computation.
This leads to a simple principle:
Don't use the most powerful AI model. Use the smallest model that can reliably solve the problem.
