AI and sustainability — how much energy does AI consume?

The energy problem of AI

Training a large language model like GPT-4 is estimated to consume as much energy as the annual consumption of hundreds of households. The data center infrastructure behind AI is growing rapidly: Microsoft, Google, and Amazon are investing hundreds of billions in new data centers, all of which need electricity.

The IEA (International Energy Agency) predicts that data centers will consume as much electricity in 2026 as all of Japan. AI is a significant part of that growth.

Illustration created with Canva AI

Inference vs. training

Training is the most energy-intensive phase, but only happens once. Inference — running the model to generate answers — is cheaper per query, but scales with usage. As AI becomes more widespread, the total energy consumption of inference increases rapidly.

Water scarcity

Data centers cool their servers with water. Microsoft's data centers already consumed 6.4 million m³ of water in 2022. This is a growing problem in regions with water scarcity, such as the western US.

AI as a climate solution

The same technology that consumes energy can also help reduce the problem:

• Google DeepMind reduced data center cooling energy consumption by 40% using reinforcement learning
• AI optimizes energy grids and predicts power demand
• Climate modeling is faster and more accurate with AI
• Discovery of new battery and solar panel materials via AI simulations

More efficient models

There is a counterforce: the pressure to make models more efficient. Models like Mistral 7B and Phi-3 offer surprisingly good performance at a fraction of the compute of GPT-4. This 'efficiency race' can limit the net impact.