GenAI – How sustainable is this US$1.3 trillion energy guzzling new tech?

A new dawn for technological innovation

ChatGPT has alerted the world to the transformative potential of Generative AI (GenAI), capturing global media attention and sparking a massive AI arms race by companies. The global GenAI market is expected to grow to US$1.3 trillion in 2032 from just US$40 billion in 2022(1), an increase of more than 30 times in size. GenAI is unlike traditional AI in that traditional AI focuses on analysis and prediction, while GenAI excels in generating new, original content. This content could range from text and images to music and even entire virtual environments, unlocking unprecedented opportunities and improving efficiency across multiple industries.

Even though GenAI is transformative in several ways, its rapid ascent raises critical sustainability considerations. Chief among these is surging energy demand driven by the increasing adoption of GenAI – this demand is likely to come from energy-intensive data centres supporting GenAI infrastructure. This poses substantial challenges for sustainability given the need to balance technological progress with the environmental impact of AI’s energy usage. Addressing these challenges involves developing more energy-efficient computing technologies, integrating renewable energy sources, and implementing robust policies to ensure that AI’s growth contributes positively to a sustainable future.

From machine learning to LLMs

GenAI has transformed from a theoretical concept into a powerful tool driving innovation across various fields. As illustrated below, this journey highlights the rapid advancement of AI technologies and their impact on industry and society at large.

The early foundations of AI in the 1950s were built on symbolic AI and rule-based systems. Researchers sought to create machines which could mimic human intelligence through predefined rules and logical operations. However, these rule-based systems were limited by their inability to learn from data and adapt to new situations, which limited their practical applications.

In the 1980s, a shift began to occur with the advent of neural networks, inspired by the human brain. Neural networks were capable of learning from data, and represented a departure from rigid rule-based systems. The early 2000s marked the resurgence of neural networks with the advent of deep learning. Aided by increased computational power and the availability of large datasets, these versions of neural networks were characterised by multiple layers which were capable of learning hierarchical representations of data

More recently, the development of large-scale language models (LLM's), such as OpenAI’s GPT (Generative Pre-trained Transformer) series has marked a milestone in the AI journey. GPT-3, launched in 2020, was a groundbreaking model due to its size and versatility. With 175 billion parameters, GPT-3 demonstrated unprecedented capabilities in generating human-like text, performing complex language tasks, and even displaying rudimentary reasoning skills. The subsequent launches of GPT 3.5 and 4.0 have enabled the platform to provide accurate and contextually aware responses. This enhanced performance made ChatGPT a more reliable tool for a wide range of applications, from customer service to creative writing. However, as LLMs gain scale and become all pervasive from a usage point of view, there is a shadow of doubt over GenAI’s sustainability credentials given the demand for computational power driving higher energy consumption.

The unprecedented power demands of GenAI – are we in for a nasty shock?

GenAI’s exponential growth is generating increased attention on its environmental impact. This is no surprise as the training and development of GenAI models is an energy-intensive exercise. With the growing adoption of GenAI technologies has come an insatiable hunger for electricity to power it.

The majority of AI systems are trained and deployed on servers housed in large data centres – these facilities use an enormous amount of electricity, amplifying the effects of climate change globally. Let’s look at a few datapoints that highlight the scale of the problem below:

• Google revealed that interacting with an LLM consumes nearly 10x more electricity than a single Google search, which uses 0.3 Wh of electricity(2)
• The International Energy Agency (IEA) predicts that electricity demand will increase by 10 terawatt-hours annually if ChatGPT were to be incorporated into the 9 billion Google searches done per day, which is the equal to the consumption of 1.5 million EU residents(3)
• Training an AI model generated a similar volume of carbon emissions as 300 flights between New York and San Francisco(4)
• Training one AI model with neural architecture search is estimated to emit as much CO2 as five cars during their lifetime(5)

Show me the ‘Energy’

Morgan Stanley estimates that GenAI’s power demands will skyrocket 70% annually through to 2027, at which point the technology could use as much energy as the entire country of Spain used in 2022(6).

AI systems, especially those employing deep learning and large-scale neural networks, require vast amounts of computational power. For instance, GPT-4, which is at the forefront of natural language processing, requires substantial computational resources for development and its ongoing operation. This has led to data centres emerging as significant consumers of energy on a global scale. Goldman Sachs research estimates that data centres in the US will consume 8% of the country’s electricity by 2030 – the equivalent of powering more than 4 million American households(7).

It is likely that this trend will play out across the world leading to a surge in energy consumption driven by the rising adoption of GenAI.

Pleasingly, research suggests that a large portion of the incremental power demand of GenAI could be sourced from zero or low-carbon technologies(8). AI-driven power demand could serve as an underappreciated driver for the manufacturers of large-scale and distributed clean energy technology and developers of wind, solar, and energy storage. However, timing is key – the question remains whether renewable energy capacity can be ramped up quickly enough to match the pace of GenAI development and adoption. It is likely that natural gas is used to power data centres over the short- to medium-term which could result in increased emissions. On the flip side, we could see key players in the AI value chain – i.e. the technology majors – put their weight behind renewable energy investments, thereby driving an investment super cycle in renewables and related technologies.

Pathbreaking tech that could transform or derail the climate transition

The environmental impact of AI’s energy usage could be significant, with increased energy consumption potentially driving up carbon emissions in the short-term. Additionally, data centres require substantial water resources for cooling, exacerbating water scarcity issues in certain regions. The extraction and manufacturing processes involved in building these data centres, including mining rare earth elements, add another layer of environmental concerns.

Addressing these challenges requires a multifaceted approach. Innovations such as specialised AI chips designed to optimise power consumption and advanced algorithms that reduce computing power are essential in mitigating AI’s environmental footprint. Moreover, integrating renewable energy sources into data centre operations can significantly reduce emissions. Alphabet and Microsoft have made strides in this direction, powering their data centres with solar, wind, and hydroelectric energy. But on the flip side, the enormous energy requirements of developing GenAI solutions have led to a big increase in the carbon emissions for these technology companies. It’s been reported that Google has seen its carbon emissions increase by 48%(9) vs 2019 while Microsoft has seen its carbon emissions increase by almost 30% in 2023 alone(10).

Furthermore, policymakers and industry leaders should adopt a collaborative approach aimed at establishing guidelines and best practice for sustainable AI development. This could encompass setting energy efficiency standards for data centres and promoting renewable energy adoption. Additionally, investing in research and development for green technologies and encouraging policies that incentivise sustainability are crucial steps toward balancing AI’s growth with environmental responsibility.

While the rise of AI heralds an era of technological progress and innovation, it also brings challenges relating to energy demand and sustainability. By developing energy-efficient technologies, integrating renewable energy sources into data centres, and implementing robust policies, we can ensure that AI’s growth contributes positively to a sustainable future. As we navigate this fast-evolving landscape, it is crucial to balance technological advancement and environmental responsibility to harness AI’s full potential while safeguarding our planet.