The hidden environmental impact of AI: beyond the technological fascination
Artificial intelligence (AI) promises to solve some of humanity’s most complex problems, from drug discovery to combating climate change. Paradoxically, the development and deployment of these technologies themselves have a non-negligible, often overlooked environmental cost: this is the hidden environmental impact of AI. From the massive energy consumption of data centers for model training to the carbon footprint of manufacturing the necessary hardware, and water consumption for cooling, the rise of AI raises urgent questions about its sustainability and the need for a more sober and responsible approach.
Energy consumption of training and inference
Training large language models (LLMs) and complex generative AI models (like ChatGPT-4o, Claude 3.7, or image and video generation models such as Kling AI 2.0) is extraordinarily energy-intensive. These processes require thousands of graphics processing units (GPUs) or specialized accelerators (TPUs) running for days, weeks, or even months. Electricity consumption can reach several gigawatt-hours for training a single large model, equivalent to the annual consumption of hundreds of households. Inference, the daily use of these models to answer billions of queries, also represents a considerable and growing energy load, although less intensive than training. This energy largely comes from still carbon-intensive sources in many parts of the world, thus contributing to greenhouse gas emissions. Efforts are being made to optimize algorithms (frugal AI) and use more efficient data centers powered by renewable energy, but the current trend towards ever-larger models exacerbates the problem.
The hardware footprint: manufacturing and lifecycle
Beyond direct energy consumption, the hidden environmental impact of AI includes the environmental footprint of the hardware required for its operation. Semiconductor manufacturing (GPUs, CPUs, memory) is a complex process requiring huge amounts of water, energy, and often toxic chemicals. The extraction of necessary raw materials (rare earths, metals) also has significant ecological and social consequences in producing countries. Furthermore, the rapid hardware refresh cycle, driven by the AI performance race, generates a growing amount of electronic waste (e-waste), which is complex and often inadequately recycled. The entire lifecycle of AI hardware, from extraction to end-of-life, therefore has a significant carbon and environmental footprint.
Water consumption for cooling
An often-underestimated aspect of the hidden environmental impact of AI is water consumption. The data centers housing training and inference servers generate intense heat and require massive cooling systems, which frequently use large amounts of water (via evaporative cooling towers). Studies have estimated that training a model like GPT-3 could consume the equivalent of hundreds of thousands of liters of water, potentially millions for newer and larger models. In a context of increasing water stress in many regions of the world, this water consumption by the AI industry is becoming a major concern. Efforts are underway to develop more water-efficient cooling techniques (liquid cooling, use of non-potable water) or to locate data centers in colder climates.
Towards more sustainable AI: sobriety and responsibility
Faced with this hidden environmental impact of AI, awareness and a change in approach are necessary. This involves several axes:
- Measurement and transparency: Developing standardized methodologies to measure the full carbon and environmental footprint of AI models and systems, and publishing this information transparently.
- Optimization and frugality: Encouraging research into more efficient algorithms (frugal AI), smaller yet performant models (ChatGPT-4-mini?, Gemma 3 QAT), and less resource-intensive training techniques. Favoring edge AI inference (cf. Cloudflare and its AI Labyrinth) where relevant to reduce data transfers.
- Renewable energy and data center efficiency: Accelerating the transition of data centers to 100% renewable energy sources and continuously improving their energy efficiency and water management.
- Hardware eco-design: Designing more durable, repairable, and recyclable AI hardware that is less dependent on critical resources.
- Reasoned use: Questioning the relevance of using AI for every application. Do we always need the largest and most powerful model? Is AI the most sober solution to solve a given problem?
Brandeploy and responsible communication
Companies aware of the hidden environmental impact of AI may wish to communicate about their efforts towards more responsible AI. Brandeploy can help structure this communication. The platform allows centralizing validated information on the company’s AI-related sustainability initiatives (use of renewable energy, choice of optimized models, etc.). It ensures that messages communicated to the public or stakeholders are consistent, accurate, and based on verified data. By managing these CSR (Corporate Social Responsibility) communication elements within the same platform as other brand content, Brandeploy guarantees an integrated and credible approach to responsible communication.
AI has an environmental cost. How does your company integrate sustainability into its AI strategy and communication? Brandeploy helps you communicate transparently and consistently.
Manage your messages on responsible AI and ensure their alignment with your actions.
Discover how Brandeploy can support your CSR communication: request a demo.
