The Energy Appetite of AI: A Challenge to Environmental Sustainability

5 Aug

Artificial intelligence (AI) has become a transformative force across industries as well as an integral part of our daily lives, revolutionising processes, enhancing decision-making, and driving innovation. However, as AI technologies advance and become more pervasive, the electricity consumed by AI raises significant environmental sustainability concerns.

Training: A One-Time, Resource-Intensive Process

“Training” refers to the development of an AI model such as ChatGPT. Training sophisticated models involves processing vast amounts of data, which requires substantial computational power. For instance, training Generative Pre-Trained Transformer 3 (GPT-3) produced 552 tCO2 emissions and consumed around 1,287 megawatt-hours (MWh) of electricity. [1] This is equivalent to the annual energy usage of 476 UK households. [2] The energy consumption translates directly into carbon emissions, particularly when the electricity is not sourced from renewable sources. Although the energy consumption can be reduced by utilising more efficient hardware and optimising training algorithms, the sheer scale of data and computational power required means that training remains a major energy consumer.

Inference: Continuous Energy Consumption

These models are subsequently operationalised for public use, in a process known as “inference”. The continuous activity of model inference further escalates energy consumption. Despite inference consuming less energy per operation than training, the high frequency of inferences aggregate substantial energy consumption. For example, a ChatGPT query demands almost 10 times as much electricity as one Google search query. [3] For generative large language models or those employing deep learning this energy consumption is amplified. For example, generating a single large 4k image using AI is equivalent to the electricity needed to charge a mobile phone. [4]

Despite these significant energy demands, there is currently no obligation for companies to declare these figures, which has made it challenging to communicate the issue effectively to the public. Some AI developers such as BLOOM are developing ethical and democratised AI. BLOOM emphasise the importance of consumer visibility.

Data centres that house AI servers and associated infrastructure currently consume 1-3.7% of the world's energy. This figure is equivalent to the global aviation industry's total energy consumption. Projections suggest that this energy usage is set to double by 2026, driven primarily by the development of more complex models and growing demand for AI. [5] Google and Microsoft have announced £800 million [6] and £2.5 billion [7] data centre investments respectively to support AI innovation in the UK. As companies invest in more data centres to support the proliferation of AI models, the environmental burden on national power grids increases. Goldman Sachs estimate that upwards of $1 trillion is needed to prepare European power grids for AI electricity demands. [3]

However, we should not underestimate the potential of the initiatives to green AI development. The EU has invested over €225 million in the Horizon Programme’s mission to reduce the energy consumption of powerful AI systems and optimise the resource management of data-intensive applications. [8]

Despite EU investment, governments need to commit to better regulation and discourse on the private sector’s AI responsibility. This includes incentivising the development of energy-efficient models, promoting the democratisation of AI, and allowing increased access to the societal benefits of AI.

Conclusion

AI models are inherently power-hungry and present a significant environmental burden throughout their lifecycle. As AI development and usage becomes more prevalent the need for considered sustainable solutions becomes ever more critical. Implementing better regulations, promoting corporate responsibility, and fostering clean energy investments is key to ensuring that AI development aligns with global sustainability goals.

By critically examining and addressing the environmental impact of AI, we can pave the way for a future in which technological advancement and ecological responsibility go hand in hand and the power of AI can be harnessed sustainably.