Artificial intelligence is reshaping our world — but at what environmental price? Training a single large language model can emit as much CO₂ as five cars over their entire lifetimes. As models grow exponentially in size and adoption, so does their footprint. These six charts follow the full lifecycle of AI’s environmental impact — from the electricity needed to train models, to the water used to cool servers, to the mountains of discarded hardware left behind.
What this chart shows: Each dot represents a landmark AI model plotted by the year it was trained and the amount of computing power (FLOPs) required to train it. The y-axis uses a logarithmic scale — each gridline represents a tenfold increase — which is necessary because compute has grown by roughly a factor of one billion since 2012. Dot size reflects model parameter count (larger dot = more parameters).
Why it matters: Between 2012 and 2024, AI training compute grew approximately 10 billion-fold — far outpacing Moore’s Law, which doubles transistor density every ~two years. GPT-3 (2020) marked a sharp inflection point. Every joule of this compute draws electricity; every watt of electricity has a carbon cost unless sourced from renewables.
What this chart shows: A horizontal bar chart comparing the electricity demand of global AI data centres against the annual consumption of eight real countries. Countries were chosen to bracket the AI figures — some above, some below — to give meaningful scale.
Why it matters: By 2026, AI data centres are projected to consume roughly 500 TWh per year — more than entire nations like Sweden, Argentina, or the Netherlands. Australia, for context, uses about 272 TWh annually. This enormous and rapidly growing electricity appetite is the root cause of AI’s carbon footprint: if the grid supplying these data centres runs on coal or gas, every query carries a carbon cost.
What this chart shows: A lollipop chart comparing estimated water usage for AI queries against familiar everyday tasks. The x-axis is linear so the dramatic difference between one query and one hundred queries is visually apparent.
Why it matters: Water is rarely discussed in AI sustainability debates, yet it is a critical resource. A single ChatGPT query requires an estimated half a litre of water for server cooling — comparable to flushing a low-flow toilet. Scale that to the billions of daily queries handled by services like ChatGPT and Gemini, and the aggregate consumption rivals that of small cities. In water-stressed regions (like parts of Australia), this is a serious and underreported concern.
What this chart shows: Each bubble represents an AI model. The x-axis (log scale) shows how much CO₂ was emitted during training. The y-axis shows benchmark performance. Bubble size reflects total parameter count — a proxy for model complexity and cost.
Why it matters: The chart reveals a pattern of diminishing returns: as we move right (more emissions), performance gains shrink. GPT-3 emits over 800 times more CO₂ than T5-Large but scores only ~9 percentage points higher. Chinchilla — a smaller, more efficiently trained model — achieves comparable performance to Gopher at a fraction of the emissions. This challenges the “bigger is always better” assumption driving the AI arms race, and suggests that smarter training strategies can dramatically reduce environmental harm without sacrificing capability.
What this chart shows: A 100% stacked bar chart breaking down the reported energy mix — renewables, low-carbon (nuclear/hydro), and fossil fuels — for the four largest AI infrastructure companies. Clicking a legend item deselects that category so you can isolate and compare specific sources.
Why it matters: There is enormous variation between companies. Meta leads with 77% renewables; Amazon — by far the world’s largest data centre operator — sits at just 23%. All four companies have made bold net-zero pledges, but pledges are not the same as real-time clean power delivery. Microsoft, despite its high-profile partnership with OpenAI, still sources nearly half its energy from fossil fuels. At a time when AI compute demand is doubling roughly every six months, the pace of grid decarbonisation is not keeping up.
What this chart shows: Each bubble is a category of AI-related hardware. The x-axis shows how long devices typically last before retirement. The y-axis shows how much CO₂ was emitted just to manufacture each unit. Bubble size reflects the estimated number of units discarded globally each year, and colour encodes the hazardous materials risk index.
Why it matters: When we talk about AI’s environmental impact, we almost always focus on electricity. But the hardware powering AI systems must be manufactured — a process that consumes enormous amounts of energy, water, and rare earth materials — and eventually discarded. AI training GPUs are particularly concerning: they have short ~3-year lifespans due to rapid generational improvements (each new GPU generation offers dramatically more performance), yet manufacturing each one emits ~150 kg of CO₂ and involves toxic materials including lead, cadmium, and beryllium. The United Nations estimates global e-waste is growing five times faster than documented recycling. AI is quietly accelerating that trend.
The six charts above trace AI’s environmental footprint from computation to cooling water, from carbon emissions to discarded hardware. The scale of growth is staggering — and accelerating. Yet the story is not one of inevitable catastrophe. Chinchilla (Chart 4) showed that smarter, more efficient training can match the performance of far larger models at a fraction of the cost. Meta (Chart 5) demonstrated that a major AI operator can run on predominantly renewable energy. The question is whether the industry will make these choices at the pace and scale the climate requires — or whether the race for capability will continue to outrun the race for responsibility.
Epoch AI. (2024). Notable AI models dataset. https://epochai.org/data/notable-models
Goldman Sachs. (2024). AI is set to drive a surge in data center power demand. Goldman Sachs Research.
Gupta, U., Kim, Y. G., Lee, S., Tse, J., Lee, H. H. S., Wei, G. Y., Brooks, D., & Wu, C. J. (2022). Chasing carbon: The elusive environmental footprint of computing. IEEE Micro, 42(4), 37–47. https://doi.org/10.1109/MM.2022.3163226
International Data Corporation. (2023). IDC global data center tracker. IDC.
International Energy Agency. (2024). Electricity 2024: Analysis and forecast to 2026. IEA. https://www.iea.org/reports/electricity-2024
Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making AI less ‘thirsty’: Uncovering and addressing the secret water footprint of AI models. arXiv. https://arxiv.org/abs/2304.03271
Lottick, K., Susai, S., Sedlins, S., & Frankle, J. (2019). Energy usage reports: Environmental awareness as part of algorithmic accountability. NeurIPS Workshop.
Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M., & Dean, J. (2021). Carbon and the broad landscape of digital operations. arXiv. https://arxiv.org/abs/2104.10350
Prakash, S., Manhart, A., Amoyaw-Osei, Y., & Agyekum, O. A. (2022). Socioeconomic assessment and feasibility study on sustainable e-waste management in Ghana. Öko-Institut.
Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M., & Villalobos, P. (2022). Compute trends across three eras of machine learning. 2022 International Joint Conference on Neural Networks (IJCNN). https://doi.org/10.1109/IJCNN55064.2022.9891914
Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645–3650. https://doi.org/10.18653/v1/P19-1355