AI’s promise – and its hidden footprint
Artificial intelligence has captured imaginations with its potential to tackle everything from climate modelling to healthcare diagnostics. The United Nations Environment Programme points out that AI is already being used to map illegal dredging and chart methane emissions, giving policymakers powerful tools to protect our planet. A Boston Consulting Group study (via Akepa) suggests wisely applied AI could help mitigate 5–10 percent of global greenhouse gas emissions.
But the infrastructure supporting AI has a dark side. Training and deploying generative models demands vast amounts of electricity and cooling water. MIT researchers report that training large AI models like GPT‑4 can consume staggering amounts of electricity, increasing carbon emissions and straining power grids. Data centres built to house these models have seen their power demand in North America almost double from 2022 to 2023, and global electricity use could exceed 1,000 terawatt‑hours by 2026. Each inference – every time you ask a chatbot a question – continues to use energy, and some estimates suggest a ChatGPT query consumes around ten times the electricity of a Google search.
On top of energy, generative AI requires large volumes of water to cool servers. MIT’s Noman Bashir notes that each kilowatt hour of data centre energy can demand roughly two litres of water, with implications for local ecosystems and biodiversity. UNEP warns that data centres also generate electronic waste and rely on critical minerals and rare elements, which are often mined unsustainably.
Short shelf‑life models and the tsunami of inference
One of the surprising findings from recent research is that training isn’t the only problem. The MIT team observes that while a model like GPT‑3 consumed about 1,287 megawatt hours of electricity and produced roughly 552 tons of CO₂ during training, inference – using the model to answer questions – can quickly eclipse those training emissions. As models grow larger and more people use them, the energy footprint of inference will dominate.
Generative models also have a short shelf‑life. Companies release new versions every few weeks, meaning the energy and materials used to train earlier models are soon wasted. Meanwhile, the number of data centres worldwide has surged from 500,000 in 2012 to 8 million today, and Ireland’s energy regulator predicts data centres could account for 35 percent of the country’s electricity use by 2026. These trends raise questions about sustainability at scale.
A wildcard technology: unintended consequences
UNEP calls AI a “wildcard” for the environment. Beyond its direct footprint, AI can have unintended higher‑order effects. For example, self‑driving cars could encourage more people to drive instead of taking public transport, increasing emissions. Misinformation generated by AI could undermine public support for climate action. The technology’s rapid growth makes it hard to anticipate every impact.
Building a greener AI future
Despite these concerns, I believe AI can be part of the solution if we make responsible choices:
- Measure and disclose impact. UNEP recommends standardised procedures for measuring AI’s environmental footprint and requiring companies to report the direct impacts of AI‑based products. Transparent data helps everyone make informed decisions.
- Optimise for efficiency. Researchers and vendors are working on lightweight models and techniques like quantisation and distillation to reduce energy use. Choosing task‑specific models rather than giant, general models saves power.
- Green your hosting. Powering data centres with renewable energy, reusing waste heat and recycling water can significantly reduce emissions. At Versantus we partner with hosting providers committed to sustainability and encourage clients to consider environmental factors when selecting infrastructure.
- Reuse and reduce. Building on existing platforms (hello, Drupal and WordPress!) and reusing code reduces duplicate training and inference cycles. Our AI for Drupal and AI for WordPress services emphasise reusability and sustainability.
- Educate and empower. The environmental cost of AI isn’t widely understood. Our AI training & support programmes help your team become AI power‑users and make informed choices that balance speed with sustainability.
Final thoughts – hope with caution
We’re in awe of what AI can do. We’re also concerned about its unseen costs. That tension doesn’t mean we should halt progress; it means we should move forward mindfully. Every technology has a footprint; AI’s is unusually visible because of its rapid ascent and the hype around it. Rather than dismiss AI or blindly embrace it, let’s choose a path that maximises benefits and minimises harm.