Digitalisation |
Squaring the AI energy consumption circle
How can we achieve a balance between the growing energy demands of artificial intelligence and its potential contributions to achieving the energy transition?
Graham Faiz Head of Digital Energy, UK and Ireland, Energy Systems at DNV
Debates about the impact of artificial intelligence on the energy transition are commonplace and show little sign of letting up, at least in the short term.
They also centre on one major conundrum, a circle that needs squaring. The “AI Age” will accelerate the shift to net zero by driving improvements in energy production, transmission, and usage, but at the same time it may slow this progress due to the high energy demands of data centres and networks supporting AI systems. How can the sector unlock this potential, without making a rod for its own back?
AI: not the main driver, not just renewables, not new
It is important to note though that policy, rather than AI, primarily drives the pace of the energy transition. Moreover, AI is not a tool exclusive to green energy developers; it is also being used in oil and gas production to improve efficiency, safety, and sustainability. That said, while AI- based tools and algorithms are often transferable across renewable energy sites, their application is more limited in oil and gas due to the unique nature of geological formations.
Much is currently being made of AI and its place in the energy sector. We are now at a pivotal moment where the large-scale adoption of AI into operating environments is materialising. Significant barriers – such as resistance to change, as well as concerns over technical and cybersecurity issues – mean this will inevitably take time. Additionally, it’s challenging to distinguish the specific impact of AI investment from broader digitalisation efforts by energy companies.
Despite this, DNV is closely monitoring developments and is actively involved in AI innovation for the energy industry, leading initiatives to ensure the safety and reliability of AI-powered systems. As we detailed in our most recent Energy Transition Outlook, AI is subject to Amara’s Law: its impact is often overestimated in the short term but underestimated in the long run. Much of today’s AI hype focuses on generative AI, particularly large language models (LLMs). Their direct use in energy is limited to low-risk applications like staff productivity tools and early experimental projects, although the risks increase as the outcomes desired by the associated use cases become safety-critical. AI that is relevant to energy systems generally falls under “discriminatory AI” – which
includes applications like computer vision, forecasting, predictive maintenance – and design optimisation, which is collectively referred to as “industrial AI”. These systems are already in use, and as AI methods improve, generative AI is expected to be increasingly applied to industrial settings.
The energy footprint of AI It is widely assumed that the hardware used to train AI models – particularly graphics processing units (GPUs) – consumes significant amounts of energy. Training GPT-3, a precursor to ChatGPT, required around 1300 MWh of electricity, equivalent to the monthly energy use of 1450 average US households. On the surface this seems substantial, but it is only a small fraction of the total energy consumption of companies like Microsoft, whose energy use in 2022 reached 18.6 TWh.
In the coming years the training and use of AI is expected to grow dramatically, and it could exceed $1 trillion in market value by 2030, roughly five to six times its current size. However, energy consumption is unlikely to rise at the same rate, as there are ongoing efforts to develop “responsible AI” that focuses on efficiencies. Beyond hardware efficiencies, advancements in machine learning architectures, better algorithms, and cloud computing - particularly in regions with cleaner, more efficient energy – are expected to further reduce AI’s energy demands.
Although research into AI’s energy growth versus efficiency improvements is still developing, current data centre energy usage offers insights. According to the International Energy Agency, in 2022, data centres, cryptocurrencies, and AI consumed approximately 460 TWh of electricity, roughly 2% of global electricity demand. By 2026, this figure is expected to rise to 800 TWh, or 2.4% of global demand. In our energy model, AI- related data centre usage falls under commercial appliances and lighting, representing about 7.5% of this category. Within that figure, regional differences in demand are inevitable, with higher growth in Europe, North America, and China, where AI adoption is accelerating.
In the broader context, this remains a relatively small portion of global electricity demand. By 2050, AI is forecast to account for just under 5% of electricity demand in the buildings sector and nearly 2% of overall global electricity usage. However, global electricity demand is expected to more than double by mid-century, meaning
38 | November/December 2024|
www.modernpowersystems.com
AI and data centre energy demand (source: DNV Energy Transition Outlook)
AI’s energy consumption could triple during this period, despite ongoing efficiency gains. These forecasts carry significant uncertainties, which will be ironed out with further research and actual evidence.
Current discussions primarily focus on the direct energy consumption of AI and the potential efficiency gains at data centres and in AI hardware. However, a larger and more meaningful conversation lies in comparing AI’s energy and emissions footprint with the many efficiencies it is expected to bring to daily life, including energy use.
A revolution that’s well underway Even at this early stage, AI technology is driving significant advancements in areas such as power grid management, materials science, renewable energy siting, and environmental impact assessments.
It has the potential to revolutionise transmission and distribution system operations
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45
