Green IT in the Age of AI: Less or More Energy Use?

26.03.2025

Green IT has not received the same level of attention in recent years as it once did. This is largely due to the rise of AI. The technology requires even more energy resources but could simultaneously be part of the solution to steer energy demand and consumption towards more sustainable paths.Fujitsu launched its own Green IT program as early as 1988. The Energy Star initiative by the US Environmental Protection Agency (EPA) further boosted the idea of more resource-efficient use of information technology in 1992. However, the concept only gained real traction at the end of the 2000s with the discovery of its economic dimension and the realization that sustainable IT systems could save both energy and money. Long before any potential image boost, this is the primary motivation for hyperscalers to establish climate-neutral data centers. This aspect is becoming increasingly important with the growing consumption in the era of AI and GenAI.

ChatGPT Consumes Less Energy Than Thought

Microsoft, for instance, has recently bolstered its AI strategy with three solar farms in the USA, which are set to deliver 389 megawatts of renewable energy, as reported by IT-Boltwise. The overarching goal is to achieve carbon negativity by 2030, meaning the company aims to remove more CO₂ equivalents from the atmosphere than it emits through its own activities. The plan includes direct air capture (DAC), reforestation, and hybrid installations that combine solar and wind energy with battery storage. This way, the GreenIT concept can be sustained long-term despite the energy demands of AI.

Meanwhile, there is a glimmer of hope regarding energy consumption for at least one AI model. ChatGPT, according to a study, requires less power than previously thought. The research institute Epoch AI has calculated that the power consumption per query of OpenAI’s language model is not 3 watt-hours, but merely 0.3 watt-hours. Analyst Joshua You attributes the previous estimates to outdated hardware data.

Total energy consumption includes training as well

This does not mean, however, that the growing energy demand of artificial intelligence is off the table. There are at least two aspects to consider:

1. The computing power and server operation require significant amounts of electricity and other resources such as water for cooling. The total energy consumption and environmental impact are therefore considerably higher.
2. Training language models is particularly compute- and energy-intensive because it involves billions of parameters and often takes weeks or months. This process frequently employs thousands, if not tens or hundreds of thousands, of GPUs, TPUs, and NPUs (Graphics, Tensor, or Neural Network Processing Units) in parallel, each consuming a substantial amount of power depending on the generation.

On the other hand, AI also has the potential to revolutionize energy generation, consumption, and efficiency. A significant amount of renewable energy produced in Germany is still lost because of insufficient transmission infrastructure and storage solutions. In industry, there are already initial approaches to combining AI and IoT into AIoT to better monitor and control resource usage and consumption. AI is thus both a curse and potentially a blessing when it comes to GreenIT.

Source image: Adobe Stock / Bundi

Read more

• CSR becomes a board issue: from mandatory report to lever
• CSRD Audit: Where the IT Data Chain Breaks
• DGX Cost Trap: Power OPEX Div

More on this topic: Further articles on mybusinessfuture

Frequently Asked Questions

What’s important when ChatGPT uses less energy than expected?

Microsoft, for instance, has recently bolstered its AI strategy with three solar farms in the USA, which are set to deliver 389 megawatts of renewable energy, as reported by IT-Boltwise. The overarching goal is to achieve carbon negativity by 2030, meaning to remove more CO2 equivalents from the atmosphere.

What’s important when total energy consumption includes training?

This doesn’t mean that the growing energy demands of artificial intelligence are off the table. There are at least two aspects to consider: the computational power and server operations.