AWS CEO: ‘AI model training requires more power than ever’

AWS CEO: ‘AI model training requires more power than ever’

One LLM needs between one and five gigawatts in the future to complete training.

CEO of AWS Matt Garman predicts that training LLMs will soon consume as much power as a major city. That equates to between one and five gigawatts of power. To meet that demand, AWS is now investing heavily in power projects.

Need for (much) more power

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

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

Major players such as Google and Microsoft are investing in renewable energy sources and nuclear power to secure power for data centers. Oracle is considering building small nuclear reactors. The hyperscalers have almost no choice, because building and commissioning new energy infrastructure takes time, and traditional energy sources are not sustainable in the long run.

read also

AWS CEO: ‘AI model training requires more power than ever’