Artificial intelligence (AI) and its underlying compute infrastructure will define the 21st century. Beyond changing the ways we work and interact, AI will fundamentally transform how government leaders imagine the concept of the state and tackle seemingly impossible challenges – including climate change.
But, despite its transformative potential, training and using AI consumes significant amounts of energy – and, as a result, is highly carbon intensive. As companies race to develop increasingly advanced AI applications, demand for energy will only continue to expand. A higher demand for energy creates the risk of higher carbon emissions if new compute infrastructure is powered by carbon-intensive energy, potentially impacting countries’ and companies’ ability to meet their climate commitments.
A recent study projects that NVIDIA, a market leader in AI computing, will see its AI servers consume more than 85.4 terawatt-hours annually by 2027, exceeding the energy usage of countries such as Sweden and Argentina. In 2020, information-and-communication-technology (ICT) infrastructure and devices consumed approximately 4 to 6 per cent of global electricity use. In 2022, Google reported that machine learning had accounted for 15 per cent of its energy usage over the previous three years.
While improvements in the efficiency of IT hardware and cooling are expected to temper energy usage, significant increases in the workloads that are handled by data centres – which provide the compute infrastructure required to train and deploy AI – have caused energy use in this segment to rise, growing 20 to 40 per cent annually. The International Energy Agency’s recent electricity forecast outlines that data-centre electricity consumption could double by 2026. Currently, data centres are estimated to account for 1 to 3.7 per cent of global emissions. By comparison, air travel and shipping each comprise about 2 per cent of global emissions. Pressure from the growing demand for data centres is leading to the development of additional fossil-fuel power plants, potentially threatening climate targets.
To address these worrying trends – while continuing to expand and develop AI use – environmental impacts must be seriously considered in the development of AI models, systems and applications without hindering use or performance. This is why we at TBI are embarking on a project that looks at how governments can enable AI to be deployed and used more greenly.
A handful of companies are already beginning to tackle this challenge and experiment with renewables, batteries and carbon-dioxide-removal technologies. For example, Microsoft’s ambitious strategy to be carbon negative by 2030 includes significant investment in innovative approaches to conserve data-centre power, reduce emissions and even contribute energy back to the grid. Google DeepMind has applied machine learning to its own data centres, reducing the amount of energy needed for cooling by 40 per cent.
Google’s goal to operate its data centres on carbon-free energy by 2030 has also led the company to make considerable financial commitments to commercialise advancements in clean-electricity technologies, including a first-of-its-kind geothermal project that will play a key role in decarbonising the world’s electricity systems. Deep Green captures and repurposes waste heat from data centres, supplying a range of other companies and infrastructure such as swimming pools and district heating. This provides lower overall carbon emissions at the national level by replacing carbon-intensive means of heating homes and other buildings.
Elsewhere, AI’s immense energy usage is driving reinvention in computer-chip development through new materials and reversible computing, and is one of the moonshots highlighted by the Special Competitive Studies Project. These collective efforts are not only driving development in low- and zero-carbon compute infrastructure but also have potentially far-reaching implications across climate, energy and broader technology sectors.
Government policies and frameworks that incentivise and support initiatives such as those outlined above are required in order to unleash developments at speed and scale, drive innovation leaps and reduce the ever-growing emissions from AI use. At present governments generally have neither the data to accurately measure emissions from AI use nor the tools and frameworks to track and monitor its environmental impacts. For commercial-sensitivity reasons, data-centre providers are often hesitant to share this sort of information. This means measurements to date have been dependent on the goodwill of internal champions within technology companies, rather than formally anchored in regulatory frameworks. Without more substantive ways to measure the carbon emissions of AI, government efforts to “map the issue” and harness private-sector innovation at scale remain handicapped.
Furthermore, while government leaders are investing in both climate efforts and AI development, rarely are there national-level efforts that work at the intersection. For instance, there is currently little done across the public sector to map the technology’s carbon emissions or understand how sustainability measures may impact AI models and approaches.
Efforts by industry and professional associations (for example, the Institute of Electrical and Electronics Engineers) to convene best practices remain within their sectoral silos and divorced from mainstream policy discussions. That gap is reflected at the international and multilateral levels, where discussions on climate change are generally isolated from discussions about creating regulatory regimes for AI use.
This must change. As government leaders begin to develop AI regulatory frameworks, monitoring carbon emissions and environmental sustainability must become integrated and leaders must start to develop ways of measuring, standardising and incentivising the use of greener AI. It is crucial that governments, companies and end users think now about the environmental costs, impacts and benefits, rather than wait until the problem has been exacerbated sometime in the next decade.
While leveraging the benefits of greener AI use may seem like a Herculean effort, international leaders are not alone in this endeavour. Nor do they need to start from scratch. Some countries are already beginning to recognise the importance of this challenge and that there are emerging solutions from which they can learn.
Singapore is at the forefront, and outlines in its National AI Strategy 2.0 the imperative of pairing AI goals with sustainability commitments. In addition, the country launched a $30 million fund for research efforts that optimise software design and function for energy efficiency. The fund would also create green-software trials to let key industry players test the impact of carbon-reduction techniques. This builds on the country’s already pioneering efforts, including the world’s first standard to optimise the energy efficiency of data centres in tropical climates.
South Korea has similarly pursued strategies to green its ICT use, investing in research and development for green technology and increasing the energy efficiency of data centres, communications networks and ICT devices.
The European Union (EU) has also taken steps recently to address the green and digital transformation that is underway. As part of the Horizon Europe programme, which develops cutting-edge technology and research, the EU is making investments to reduce the energy consumption of powerful AI systems. In addition the EU has launched a digital partnership with Singapore, a component of which is focused on fostering investment in resilient and sustainable digital infrastructure.
All that said, green government AI strategies remain the exception rather than the norm. The private sector is already exploring ways to reach a balance between AI and climate protection. Governments now need to find the enabling frameworks and incentives to unleash innovation.