Amazon CEO: Energy Consumption for Future Large Language Model Training Will Be Astonishing

Amazon Web Services (AWS) Chief Executive Officer Matt Garman recently projected that the energy required to train the next two or three generations of large language models (LLMs) will rival the energy consumption of major metropolitan areas. According to a report by The Transcript on X (formerly Twitter), Garman disclosed in an interview with The Wall Street Journal (WSJ) that training a single large language model could demand between 1,000 to 5,000 gigawatts of electricity. To meet this escalating demand, AWS is actively investing in energy projects.

For instance, Meta’s Llama LLM made its debut in February 2023, followed by Llama 2 in July and Llama 3 in mid-April of this year. If other LLMs maintain a similar training cadence, we might witness the emergence of new model generations approximately every seven months on average. In contrast, OpenAI, a frontrunner in the LLM domain, released GPT-3 in June 2020 and GPT-4 only in March 2023. Although OpenAI introduced GPT-3.5 in 2022, it was primarily an enhancement of GPT-3 rather than an entirely new generation. Consequently, OpenAI took nearly three years to launch its next-generation model.

From these insights, it can be deduced that, given current hardware capabilities, training a typical next-generation LLM model takes about one to two years. While AI firms are deploying more AI-specific GPUs for model training, these LLMs, such as the prospective Llama-4, are becoming increasingly intricate and may necessitate clusters exceeding 100,000 NVIDIA H100 GPUs for support. Due to computational constraints, OpenAI has postponed the release of its ChatGPT-5 model to 2025. It is anticipated that within approximately five years, the requisite 5,000 gigawatts of energy demand may be met.

This timeframe offers tech giants like OpenAI, Microsoft, Amazon, Google, and Oracle an opportunity to expand their energy production capacities. Garman noted that AWS is sponsoring over 500 projects aimed at integrating renewable energy into the grid, which is vital for data centers. The deployment of renewable energy sources is time-consuming, whereas traditional energy sources such as coal and natural gas generate significant greenhouse gas emissions. In the competition for AI supremacy, energy has emerged as a pivotal issue. Even Google has not met its climate objectives, with the energy demands of its data centers increasing emissions by 48% since 2019. Some former Google executives have even recommended abandoning climate goals to focus entirely on AI development, hoping to address the climate crisis in the future.

Nevertheless, these AI behemoths recognize the threats posed by potential energy supply shortages. In addition to medium-term investments in renewable energy, several companies are initiating investments in nuclear energy advancements. Microsoft has entered into an agreement to restart the Three Mile Island nuclear reactor to satisfy its data center requirements, while Google and Oracle plan to construct their own small-scale nuclear reactors. Furthermore, established entities in the traditional nuclear power sector, such as Westinghouse, are developing easily deployable microreactors designed to supply energy to next-generation AI data centers.

At present, energy supply limitations have become a bottleneck for AI advancement, particularly because the establishment of new infrastructure like power plants, transmission lines, and transformers is time-intensive. Although AI companies can resort to portable generators and other non-renewable energy sources to fulfill short-term electricity needs, this is not a viable long-term strategy. Therefore, the continued progression of AI hinges on the rapid deployment and operationalization of these alternative renewable energy sources.