With around one in six companies in the UK choosing to embrace at least one Artificial Intelligence (AI) technology, it comes as no surprise that this technology has become an integral part of every industry.
The capabilities of AI to mitigate the climate crisis are well-versed, with reports suggesting that AI has the potential to reduce 4% of GHG emissions by 2030. However, as more companies use AI to manage their Scope 3 emissions, a growing concern is emerging surrounding the carbon emissions associated with the training and operating of these large models.
In January 2024, at the World Economic Forum’s annual meeting in Davos, the chief executive of OpenAI, Sam Altman, admitted that the AI industry is heading for an energy crisis. Despite the persistent downplaying and denial about the environmental impact of AI – which is, unfortunately, the case with most digital technologies – Altman issued a warning that the next generation of AI models will require unprecedented levels of energy, potentially straining global power grids. Altman urged that “there’s no way to get there without a breakthrough”, later underlining the importance of nuclear fusion to overcome this challenge.
So why is AI putting more pressure on our climate? AI systems undergo a four-stage lifecycle: production, transport, operation, and end-of-life. While all stages contribute to the carbon footprint, the operational phase is largely responsible for 70-80% of emissions. This is because this stage demands a substantial amount of energy to run physical infrastructure such as data centres, processors and other specialised computing hardware.
As AI models become increasingly complex, especially generative AI models, their energy consumption for training and operating is likely to grow exponentially, posing a substantial challenge to global energy systems. For instance, it’s estimated that GPT-4 consumed over 52,000 MWh, over 40 times higher than GPT-3. This is the equivalent of the energy consumption over 5-6 years of 1,000 average US homes.
In addition to consuming energy, AI training generates significant amounts of heat, which requires enormous amounts of fresh water to cool the data centres. The average modern data centre uses around 40% of its power for cooling. This is problematic because as more businesses adopt AI technologies, the demand for data centres and their water consumption will increase.
As with any tool, AI can be used responsibly or irresponsibly. While AI offers the potential to address global challenges, companies cannot ignore its environmental impact. A responsible and sustainable approach is essential to harnessing AI’s benefits while mitigating its environmental drawbacks.
Sustainability should be a core metric when developing AI systems, and trade-offs between performance and energy consumption must be carefully considered.
Before integrating AI in business operations, first conduct an environmental impact assessment to evaluate whether the implementation of AI is the most suitable option.
Often, simpler, less computationally intensive methods can achieve similar results with a significantly smaller environmental impact. For example, in a day-to-day business application, every 20-50 questions asked in ChatGPT require 500 ml of water. Therefore, a traditional Google search is far more energy-efficient than using AI to find information.
It’s common for businesses to look to AI solutions to enhance efficiency and reduce manual intervention. However, it’s important to consider alternative methods such as automation. Automation, which relies on predefined rules, can streamline processes without the energy consumption associated with AI. By starting with automation, businesses can identify areas where AI may add value, ensuring that AI is used strategically and responsibly.
To minimise the environmental impact of AI, businesses must prioritise energy efficiency and renewable energy sources. Scheduling AI workloads to coincide with peak renewable energy generation, such as during sunny periods, can significantly reduce carbon emissions.
Additionally, investing in energy-efficient hardware and optimising AI algorithms is crucial for reducing energy consumption. This can be achieved by training AI models off-grid in locations with an abundance of renewable energy.
Quantifying the exact energy consumption and carbon emissions of AI systems remains a significant hurdle. Without standardised measurement and reporting frameworks, companies can easily overstate their efforts or underreport their impact. To ensure transparency and accountability, clear regulations and industry standards must be established.
As the driving force behind AI development, the tech sector plays a pivotal role in addressing the environmental impact of AI. By pioneering sustainable practices and setting industry standards, tech companies can set a precedent for other industries.