As generative AI continues to rise in popularity, so do questions about its contribution to global CO₂ emissions. While it’s true that modern GPUs—the hardware powering AI systems—consume substantial amounts of energy, discussions often lack broader context. Like the debate over plastic straws a few years ago, focusing solely on AI’s energy use without considering other aspects of energy consumption can lead to an incomplete picture. That said, it’s worth examining both the benefits of generative AI and its potential environmental impact with care.

A recent post illustrates one take, and in the following graph adapted from Founders Pledge (see here), they highlighted the energy consumption of generative AI to other everyday activities. To summarise the findings:

• Asking 50,000 questions on ChatGPT has a negligible climate impact compared to significant lifestyle changes, such as living without a car or avoiding a single transatlantic flight.
• For scale: Performing 300 ChatGPT queries consumes roughly the same amount of energy as:
  
  • Driving your car 1 kilometre or
  • Streaming 1 hour of Netflix.

This comparison reveals that while generative AI does consume energy, its impact is minimal compared to other high-carbon activities we regularly engage in:

A recent post illustrates this point perfectly in the following graph adapted from the source found here:

However, as the adoption of generative AI grows exponentially, so does its cumulative energy use. While this expansion won’t rival the emissions of the aviation or automotive industries anytime soon, researchers from institutions like the University of Massachusetts Amherst caution that the training of large AI models can result in carbon emissions that are immense (Forbes). Figures from these reports highlight the importance of improving energy efficiency as the technology scales. 

Beyond Energy: Water Usage and Broader Implications

The study also examines other environmental factors such as water consumption. It reached a similar conclusion that generative AI’s contribution to climate change or environmental degradation is not a significant threat compared to other industries or activities.

However, we can’t escape that cooling the massive data centres that power AI systems requires significant water usage. A 2023 report by the Allen Institute estimates that training large models consumed hundreds of thousands of litres of water. 

While the numbers are still smaller compared to other industrial uses, it underscores the need to consider a broader range of environmental factors when assessing AI’s sustainability.

Denmark’s AI Supercomputer: Balancing Innovation and Emissions

As a way of balancing and taking care of the environment, Denmark’s AI supercomputer, Gefion, provides a useful case study. 

With 1,528 NVIDIA H100 GPUs, Gefion represents a leap forward for research and innovation in the country. Its energy consumption, however, is notable:

• Energy usage: Each GPU consumes approximately 700W, amounting to 6,000 kWh per year.
• CO₂ emissions: The Danish grid emits 119 gCO₂ per kWh, resulting in annual emissions of 1,100 tonnes of CO₂ for the supercomputer’s GPUs.
• Perspective: This is equivalent to the annual emissions of 150 Danes.

These emissions can be argued to be relatively modest compared to Denmark’s overall output and compared to the vast benefits of advancing cutting-edge research, supporting sustainable technologies, and driving AI innovation in Denmark.

However, it is still important that we constantly optimise data centre efficiency and increase reliance on renewable energy, which also helps reduce this footprint further. This is where Denmark stays in front and makes sure the likes of the Gefion supercomputer are as close to zero emissions as possible. Utilising the benefits of AI can therefore be done in energy-efficient ways.

To conclude, Generative AI is already impacting numerous sectors. For example, Raffle Chat and Summary help users get more accurate answers to their questions, saving time and effort. As high accuracy is known to cost energy, we have developed ways to lower this footprint without jeopardising the essence of the solutions. To keep accuracy at its highest with as little energy used as possible, Raffle has customised small language models to narrow down the content from indexes (intelligent search) and thereby deliver, and ask for as little as possible, to energy-consuming large language models. Among many other hoops we do, to use as little power as possible, this is just one example.

While generative AI uses energy, it can therefore be utilised in optimised, effective ways so that it will not impact the global climate any more than old technologies were—maybe even less.

‍

The bottom line is that you can use generative AI guilt-free and focus your energy consumption on what truly matters, but only if you seek to choose providers that are constantly working to reduce their footprint. This can be hard to know and navigate in, so look out for suppliers that promote this and can actually prove the validity.
Reducing footprint to an absolute minimum is one of Raffle’s values and drivers. We are committed to delivering state-of-the-art AI search, chat, and assistants that are CO₂ neutral and encourage all providers of AI tools to do the same.