In the future, an LLM will require between one and five gigawatts for training.

Matt Garman, CEO of AWS, predicts that training LLMs will soon use as much electricity as a major city. This corresponds to an output of between one and five gigawatts. To meet this demand, AWS is now investing heavily in energy projects.

Need for (much) more power

The development of LLMs is progressing rapidly. Meta released its first version of Llama in February 2023, followed by Llama 2 in June. Llama 3 was launched in April this year. OpenAI, on the other hand, took three years to bring GPT-4 to market, but also released the refined GPT-3.5 in 2022. The development of a new generation of LLM therefore seems to take one to two years.

The increasing complexity of LLMs requires enormous amounts of GPUs, sometimes more than 100,000 Nvidia H100 GPUs per model. OpenAI has even postponed the launch of ChatGPT-5 to 2025 due to a lack of computing power. At the current rate, five gigawatts of electricity consumption will be a reality in about five years.

Big players like Google and Microsoft are investing in sustainable energy sources and nuclear energy to ensure the energy supply of data centers. Oracle is considering building small nuclear reactors. Hyperscalers have no other choice as building and deploying new energy infrastructure takes time and traditional energy sources are not sustainable in the long term.