Balancing innovation: addressing AI's energy appetite

The substantial electricity demand of Artificial Intelligence highlights a growing and unique challenge for the technology and energy industries – reconciling its progress with sustainable consumption. The hope is that sooner rather than later AI itself will provide a meaningful aid to a balancing act that looks particularly complex.

It is now apparent that AI’s capabilities to create images or videos, summarise long notes and write papers are in fact the result of machine-learning models that in turn depend on massive data sets and on the costly transfers of their content between processing and storage units, both internally and across chips. All that leads to a massive use of energy.

The scale of such consumption is unprecedented in the technology industry and increasingly amounts to a reason for concern. In its annual forecasts, the International Energy Agency (IEA) said that based on the projected sales of servers, by 2026 the AI industry may grow so much “to consume at least ten times its demand in 2023.” That would translate to approximately 85-140 TWh annually – comparable to the entire electricity consumption of countries like the Netherlands.

One possible factor contributing to this rise is the transition to AI-driven online searches. While it is challenging to accurately determine the energy use of current AI algorithms, the IEA estimates that a single query to chatbot ChatGPT requires 10 kilojoules or about ten times the energy of a standard Google search.

Will the training and running of Gen-AI models be made more energy-efficient? Or will businesses at all levels follow the example of the AI giants that are opting for alternative energy solutions despite resistance and regulatory hurdles they encounter?

Nuclear option

Marco Annunziata, co-founder of Annunziata + Desai Advisors and a former chief economist at General Electric Company, argues that the latter scenario is more and more likely.

“AI is enormously power hungry, and all credible forecasts point to a massive increase in the energy consumption needed to further improve AI models and deploy AI solutions across industries,” he says. “Moreover, AI needs stable and reliable energy, something that the current energy transition already struggles to ensure given current limitations to storage technology -- given this, it’s hard to see how the AI revolution can move forward without greater recourse to nuclear power, even if the new US Administration takes a more pragmatic stance on fossil fuels.”

He notes that under pressure to secure reliable power for their future growth, vocal advocates of renewable energy like AI giants Microsoft, Amazon, and Google are embracing nuclear power by either committing to buying it from companies managing existing facilities or striking deals with those building new ones. “Despite nuclear being more expensive than solar and wind (both on- and off-shore), they are still opting for it,” he said in an interview.

Interestingly, when Annunziata asked various AI models – including the ones controlled by those tech giants - about addressing their own power consumption and become less resource intensive, they consistently suggested renewable energy solutions, optimisation strategies, and emergency planning – but none mentioned nuclear power in response to his ‘prompts.’  

“These are the AIs that are supposed to help us crack the hardest problems, so…either nuclear is a stupid choice, or these Gen-AI models are not worth all the extra money and emissions, or…there is something they are not telling us.”

Been there before

It’s worth noting that the tech and the energy industries have navigated similar challenges before. In the early 1990s, they faced what seemed like an insurmountable problem: the unsustainable power consumption of increasingly faster computer processors. The solution that emerged transformed computing architecture forever.

Manufacturers shifted their strategy by adding multiple processing cores and improving efficiency rather than raw power. IBM launched the first mainstream multicore processor in 2001, with Intel and AMD following suit. And they did not just solve the immediate crisis – they enabled the mobile computing revolution. Today's smartphones pack more computing power than the supercomputers of three decades ago while consuming a fraction of the energy.

Deploying AI

This time around the fears surrounding the AI’s power hunger may find an answer… in the AI itself. Recent advancements, such as optimising algorithms for energy efficiency and deploying AI to manage power consumption in real-time, present new opportunities.

For instance, data centers can adopt AI-driven cooling systems to cut down on energy use. Another promising solution is the use of memristor-based systems, which can drastically reduce energy consumption, making it even conceivable to create self-powered edge AI systems that do not require batteries and can instead harvest energy from the environment.

However, integrating AI into the energy management infrastructure is not without its own set of challenges. There are concerns about the initial costs, the complexity of implementation, and the reliability of AI-driven systems in critical environments. These factors need to be addressed to pave the way for widespread adoption.

Through the collective efforts of industry leaders, policymakers, and researchers, the energy sector can harness AI’s potential while ensuring sustainable energy practices for future generations. This balanced approach is crucial to maintaining commitment to a sustainable energy future aligned with global climate goals.