AI’s environmental footprint spans energy, water, and rare earths. Innovations in design and sustainability are key to reducing its impact.
As public and regulatory scrutiny of artificial intelligence grows, AI providers and adopters face a critical challenge: making AI both ethically sound and legally compliant.
Artificial intelligence comes with a hidden cost: its environmental impact. While discussions often focus on AI's capabilities, the resources it consumes throughout its lifecycle receive far less attention. So, just how significant is this cost?
In this interview, Damien Fovet, Architect and Head of Sustainable AI at Sopra Steria Group outlines the key stages of AI’s lifecycle, examines its environmental toll, and highlights emerging innovations that could drive a more sustainable future.
What exactly does AI’s environmental impact entail?
Damien Fovet: AI's environmental footprint spans every stage of its lifecycle—not just the commonly assumed usage phase. It starts with its creation and moves through data collection, training, fine-tuning, and, ultimately, disposal. Each stage relies on specialised hardware that consumes large amounts of energy, water, and rare earth metals while generating greenhouse gas emissions.
When it comes to training, models like GPT-3 are often mentioned for their high costs. Can you help put those into perspective for us?
Absolutely. The training phase is one of the most resource-intensive steps. Large models are trained on GPUs, which require vast amounts of energy and significant water resources for cooling. Even after training, models often go through fine-tuning for specific tasks, further increasing energy consumption.
Take GPT-3, for example. Training this model generate as much CO2eq as 205 round trips between Paris and New York. While this might seem manageable in isolation, the problem escalates with frequent updates. Just six months later, GPT-3.5 arrived, likely doubling the resource demands. Two and a half years later, GPT-4 was launched, with 48 times the training impact of GPT-3. This isn’t a one-time environmental cost—it’s a recurring and growing impact as models continue to evolve.
Does individual use also contribute to AI's environmental impact?
Yes, it does. Every interaction with AI—whether for customer service, data analysis, or generating images—uses energy. Large models require substantial energy even for routine tasks, making individual usage a critical piece of the global footprint puzzle. A single query might seem insignificant, but the impact grows rapidly at a global scale.
Models are trained on massive amounts of data, how does this impact AI's environmental footprint?
Data serves as the backbone of AI, and collecting it requires both human effort and energy. Storing and organising this data relies on additional hardware, which increases energy consumption and takes up physical space.
Water plays a critical role as well. Data centres use it for cooling, heating it up in the process but keeping it clean enough for reuse. In hardware production, however, water often becomes contaminated and cannot be reused.
Seawater isn't a viable alternative either, as it causes corrosion and salt buildup, reducing cooling efficiency and increasing energy consumption. For perspective, a single AI conversation uses about half a litre of water, while training models like GPT-3 can require an astounding 700,000 litres.
What role do rare earth metals play in AI systems?
AI hardware relies heavily on rare earth metals and specialised materials, which are challenging to source sustainably. These materials aren’t exclusive to AI; they’re essential for everyday devices like phones and laptops, drawing on nearly the entire periodic table.
The ethical and environmental costs of sourcing these materials are significant. Mining disrupts ecosystems and often involves exploitative labour practices, raising a critical question: how can we reduce our dependence on these materials?
Recycling offers a promising solution. Emerging technologies could enable the recovery and reuse of nano-scale particles from discarded hardware. But, while these methods hold significant potential, they remain in the early stages of development or are not yet cost-efficient compared to mining raw materials.
What steps can we take to reduce AI’s environmental footprint?
First, we need to rethink the materials AI systems require. Beyond sustainable sourcing, we must explore ways to design systems that require fewer resources. Addressing the root cause is more effective than simply shifting to “greener” materials.
Second, optimising computational loads is key. By operating hardware at its most efficient levels—not overused or underutilised—we can maximise performance while cutting energy consumption.
Finally, data sharing and mutualization present significant opportunities. When organisations independently collect, process, and train similar datasets or models, the environmental cost multiplies. Centralised datasets and shared, open AI models could eliminate redundancy. Collaborative efforts in model training could streamline processes and reduce the overall environmental impact.
Could quantum computing make AI more sustainable?
Quantum computing could be transformative, though it’s uncertain when—or if—stable quantum computers will become a reality. They have the potential to perform large-scale calculations with significantly less energy.
Quantum-inspired algorithms are already showing promise, compressing models by up to 85% while maintaining accuracy. This represents a major step toward reducing energy consumption. At Sopra Steria, we’re exploring this potential through a partnership with Multiverse Computing, a startup specialising in AI model compression.
Can AI go beyond reducing its own footprint to actively help mitigate climate change?
Absolutely. While reducing AI’s ecological impact is crucial, its potential to help mitigate climate change is extraordinary. AI’s ability to process and analyse massive datasets drives better environmental predictions, optimised energy consumption, and more accurate forecasting of climate patterns.
For example, AI processes vast environmental datasets to forecast extreme weather events. This helps governments and organisations prepare for natural disasters like floods, hurricanes, and droughts, minimising their impact. AI also enhances water resource management by forecasting shortages and identifying drought-prone areas, enabling proactive and sustainable actions.
AI is also a game-changer for energy optimization. By analysing grid data, it can reduce energy loss, balance supply and demand, and integrate renewable sources like solar and wind more effectively.
Any final thoughts you'd like to share?
AI’s effectiveness in mitigating climate change depends on how we design and deploy these systems. Training a large-scale AI model exclusively on sustainability data could yield recommendations and optimizations focused entirely on reducing our ecological footprint.
While AI alone cannot solve climate change, it serves as a powerful tool that, when applied strategically, promotes sustainable practices and offers practical solutions for both adaptation and mitigation.