The business case for rightsizing AI is also an environmental one
Deployment must match the scale of the model with the task
ARTIFICIAL intelligence has become a boardroom priority, but chief executive officers are not considering its environmental footprint highly in the race to deploy it.
By now, many of us have heard the headline-grabbing statistics about AI’s energy and water use.
One widely cited estimate suggests that generating a single image using a powerful AI model can take as much energy as fully charging a smartphone.
A typical text-interaction with a large language model (LLM) may use 10 times more energy than an ordinary Google Search, Alphabet’s chairman John Hennessy reportedly told Reuters.
Water raises similar concerns. A medium-sized data centre in the US consumes an average of 416.4 million litres of water annually, enough to supply roughly 1,000 households for a year, with large and hyperscale centres consuming orders of magnitude more.
A rightsizing, or right-scaling, approach to AI deployment is needed: one that designs and selects AI systems that match the scale of the model to the complexity of the task, while treating energy and environmental costs as design constraints from the outset rather than externalities to be ignored.
In practice, this approach means asking before deployment: Does this actually require a frontier model, or can a smaller, purpose-built one accomplish the same goal?
As AI becomes embedded in business operations, companies can aim to avoid using more AI than the task requires.
Optimising for efficiency
For organisations and companies designing or procuring AI solutions, there are several principles that can be used to guide selection of the rightsized tool.
For example, in customising LLMs for specific tasks, there are simple design principles that can be implemented to lower the energy and water consumption.
A well-designed system can automatically route simple, routine queries to faster, lightweight methods and escalate only to heavier AI reasoning models when the task genuinely demands it.
A 2025 study found that applying model selection to AI tasks could reduce energy consumption by nearly 28 per cent, saving 31.9 terawatt-hours worldwide, or the equivalent of the annual output of five nuclear power reactors.
In the design of domain-specific LLMs, there are best practices that inform rightsizing approaches.
Caching to store and reuse proven reasoning approaches for common queries can help to deliver faster, more consistent responses while using a fraction of the energy compared to a system that treats every task as novel.
Since most organisations already have valuable, trusted information in the form of reports, databases and other expert knowledge, a well-designed system will rely on a lean, accurate knowledge base to ensure that AI systems pull from a credible knowledge store rather than large, generic models where hallucination and error rates can be high.
This also enables the domain-specific LLM to be more helpful for the use case, while keeping compute requirements minimal.
Lastly, designing outputs optimised for efficiency can also help to rightsize solutions. Longer, more verbose outputs require more energy and water, while output optimisation can help to reduce these impacts.
“Asking solution providers to disclose the energy and water consumption of their models on a per query, per deployment and at-scale basis turns an invisible cost into one that can be managed.”
Better procurement
Since companies generally cannot control the location of computing infrastructure, procurement is perhaps the biggest lever they have to shape the energy and water consumption of their AI adoption.
Just as companies now routinely require suppliers to disclose their carbon footprint or labour practices, the same approach can be extended to AI vendors through procurement practices.
Asking solution providers to disclose the energy and water consumption of their models on a per query, per deployment and at-scale basis turns an invisible cost into one that can be managed.
Yet, IBM found that while 63 per cent of businesses they served were planning to apply generative AI to “sustainable IT initiatives”, only 23 per cent were considering sustainability assessments during the design and planning stages of IT projects.
This disconnect highlights an opportunity for companies to treat sustainable AI usage not as a downstream compliance issue, but as a core design principle from the outset.
Policy levers
The GovAI Coalition, formed by the City of San Jose in the US, is one example from the public sector where agencies are being encouraged to ask AI vendors to disclose the environmental impacts of their products, including the energy and water used for training or per model use.
Governments have a larger lever than procurement. They can mandate that the build-out of data centres be combined with requirements to develop renewable energy for their power consumption, alongside deploying more efficient cooling technologies that reduce water use.
At the same time, these measures will be insufficient. The Jevons paradox reminds us that efficiency gains can also drive up the use of a technology and increase demand overall.
In the case of AI, faster or more efficient chips, cleaner electricity and better cooling may reduce the footprint of each AI task, but they will not stop total energy costs from soaring if deployment goes unchecked.
This brings us to a reality many boardrooms are avoiding. Rightsizing the technology is pointless if companies do not ask the fundamental question at the outset: whether AI is even the right solution.
The writer is an associate professor of public policy and environment, energy and ecology at the University of North Carolina at Chapel Hill. She founded and directs the Data-Driven EnviroLab.
