AI systems could represent nearly half of data center power consumption by the end of 2025, according to research from Alex de Vries-Gao, founder of the Digiconomist tech sustainability website.

He estimates AI could consume 23 gigawatts (GW) of energy by year-end, nearly double the total energy usage of the Netherlands.

This projection excludes cryptocurrency mining and is based on the power needs of chips from Nvidia and AMD, used to train and operate AI models.

In 2024, data centers used 415 terawatt hours (TWh) of electricity, with AI already accounting for around 20% of that.

De Vries-Gao warned that while efficiency improvements reduce computational and energy costs, these gains might increase AI adoption and drive demand.

He also raised concerns over fossil fuel reliance, highlighting startups like Crusoe Energy securing gas-powered energy capacity.

🔗 Source: The Guardian

🧠 Food for thought

1️⃣ History suggests we should approach dramatic energy predictions cautiously

The current projections about AI consuming half of datacenter power by year-end follow a pattern of dramatic forecasts that haven’t always materialized as predicted.

Previous EPA reports from the early 2000s projected a 15% annual increase in datacenter energy consumption, creating fears of an “energy apocalypse” that didn’t materialize 1.

Instead, actual energy consumption growth slowed significantly to just 1.6% annually since 2010, despite massive growth in data services 1.

This historical context matters because early datacenter energy projections failed to adequately account for efficiency improvements that emerged in response to energy concerns.

De Vries-Gao acknowledges similar factors could impact current AI energy predictions, including efficiency innovations like those demonstrated by China’s DeepSeek R1 model that reportedly uses fewer chips.

These historical patterns suggest we should interpret today’s dramatic projections with appropriate context, recognizing that technological systems often evolve in response to resource constraints.

2️⃣ AI’s energy paradox: efficiency innovations compete with expanding applications