How sustainable is AI?

Agents for Change uses AI to support sustainability teams. But is AI itself sustainable? It's now widely known that AI consumes vast amounts of energy and water. People who prioritize sustainability rightly have concerns about this.

We do too.

Just as we try to make choices in our daily lives that have a positive impact, we do the same in our work. That's why we only use AI where it's far more efficient than doing things manually, because manual work also consumes energy: from commuting to the office to heating your workspace and video calls.

So, it's high time to examine the actual impact of AI. How bad is it, really?

Alex de Vries, a PhD researcher at VU Amsterdam, wrote an interesting article about it: [Link to the article](https://www.sciencedirect.com/science/article/pii/S2542435125001424?dgcid=author)

He points out that without regulation, it's extremely difficult to obtain reliable data on AI's energy consumption; Big Tech and hardware companies don't report on it. He can only make estimates, but they indicate enormous energy use.

The total electricity consumption of all data centers combined (excluding crypto mining) was 415 terawatt-hours last year (according to an estimate by the International Energy Agency).

De Vries estimates that AI hardware alone could account for 11%–20% of that, more than the energy use of entire countries like Switzerland and Finland. And AI's share of energy consumption is growing rapidly. By the end of this year, it could already consume 50% of all data center capacity, even more than Bitcoin mining.

Data centers currently consume about 1.5% of global electricity. As usual, the U.S. is the biggest consumer at 45%, followed by China at 25% and Europe at 15%.

The IEA expects that by 2030, electricity demand for AI will quadruple.

Worse yet: in the rush to build more data centers, there's an expectation that they will increasingly rely on dirty energy sources like gas in the coming years. Data centers must run continuously, making it harder to use solar and wind energy. Big Tech is looking to nuclear power as a solution; Meta, Amazon, and Google are pushing for a tripling of nuclear capacity by 2050. Experts aren't taking this too seriously, as nuclear energy currently provides only a fraction of the energy needed for data centers, and developing new plants is expensive and time-consuming.

AI models improve when they have access to more data. But the larger the AI models, the more energy they consume. To give you an idea of scale: DeepSeek has an estimated 600 billion parameters, while OpenAI's next model is expected to exceed a quadrillion.

What makes a huge difference is how clean the energy grid supplying the data centers is. The more solar and wind energy, the better.

All the more reason to carefully consider AI applications, and only use AI when the use case is clear.

Hugging Face calculated that so-called "reasoning models" consume 43 times more energy for a simple multiplication than a calculator.

Fortunately, there are signs that AI could become less energy-intensive in the future. For example, more efficient training methods are being discovered that don't require feeding AI models all the data from the internet. Smaller, more specialized models are being developed. You don't always need an all-encompassing language model—sometimes, that even reduces reliability.

Smaller models also require less powerful computer chips. It would be great if AI models could run on home computers.

Additionally, work is being done on more efficient cooling. The greater the computing power, the more cooling is needed. Currently, up to 40% of energy consumption in data centers goes toward cooling servers. Smarter methods are being developed here, too. Plus, using less energy simply costs less.

In short, the same rule applies to AI as in general: avoid unnecessary applications and be aware of harmful side effects so you can make better choices.