The Energy Paradox

Paris AI Action Summit

February 10, 11 of 2025, the Paris AI Action Summit happened.

This is the third summit

Observatory on E-AI-DC

Key outcome was an Observatory on Energy, AI and Data Centres, expected to launch in April 2025 under the International Energy Agency (IEA), which will compile global data on AI's electricity demand while tracking AI applications in energy sectors.

GPT's energy demand

Large scale AI models consume vast energy, adding to carbon emissions. AI applications rely on Generative Pre-trained Transformers (GPTs) models

5 phases of GPTs

These models requires power throughout their lifecycles, spanning five phases

Planning and data collection
Model development
Training
Deployment
Ongoing maintenance

2% of global demand

In 2022, Data centres consumed 460 terawatt hours (TWh) of electricity - about 2% of the global demand and contributed 1% of the world's energy related to carbon dioxide emissions.

Aviation's global CO2 emissions

Aviation accounted for 2% of global CO2 emissions in 2022.

Electricity consumption

Global data centre electricity consumption currently at 460 TWh, could exceed 1000 TWh by 2026 - matching Japan's consumption, according to IEA.

Water guzzler

AI is also a major water guzzler. Data centres rely on water intensive cooling systems to maintain optimal conditions for high performance computing.

Google - 29 billion litres

In 2023, Google's data centres withdrew 29 billion litres of water, of which 23 billion litres was consumed for on site cooling. Of this, 80% was potable.

Meta - 5.2 billion litres

Meta data centres consumed over 5.2 billion litres in 2023 marking a 50% rise since 2019.

DCs in water stressed regions

Many data centres are located in water stressed regions. A report by London based non profit Planet Tracker finds that 41% of data centres are in areas with extreme water stress, with another 22% in highly stressed zones. In Asian continent, Indonesia has the most data centres in water stressed regions, followed by India and Russia.

Power consumption of data centres

Between 2015 and 2019, global data centres power consumption remained stable, but with AI demand accelerating, these gains are showing.

Tripling global nuclear capacity

In March, Amazon, Google and Meta pledged to triple global nuclear capacity by 2050.

US data centres could exceed $20 billion

A 2024 pre print study by Ren and colleagues estimates that the health burden of US data centres could exceed $20 billion annually by 2030.

ChatGPT vs Environment

[!] ChatGPT Electricity Consumption is More Than 17,000 US Household
That’s ChatGPT for you – a brilliant conversationalist with an appetite for electricity that’s sparking conversations about sustainability and innovation.
The AI sector’s energy appetite is growing exponentially, and by 2027, it could consume between 85 to 134 terawatt-hours (TWh) annually equivalent to half a percent of global electricity consumption.