Greening algorithms: How seriously is big tech taking carbon emissions?
At a time where the world is set to exceed the target global temperature increase of 1.5°C above pre-industrial levels outlined in the 2015 Paris Climate Agreement, artificial intelligence (AI) is booming. In this context, the acceleration of AI and the related data centre growth has raised further concerns over climate action.
As responsible investors this direction of travel is difficult to approach. AI appears to have the potential to drive economic efficiencies and innovate in ways that can help not only find solutions to environmental issues, but also ultimately help reduce emissions.
But these positive outcomes from AI are challenged by the energy demands of AI transformation and it is clear that that world is not slowing down. Squaring these two competing agendas is difficult to do and as such investors must use this proliferation to better understand the approach of their holdings.
On the surface, there does appear to be a positive direction of travel. From a shareholder perspective, some companies pivotal to AI expansion, particularly cloud hyperscalers like Amazon, Alphabet, and Microsoft, have set ambitious decarbonisation plans. For example, Microsoft aims to be carbon negative by 2030, and Alphabet targets net zero emissions across all operations and value chain by 2030.
But we must question whether these targets are genuinely realistic, or simply paying lip service to an issue that is ultimately difficult to measure?
AI’s net impact
Understanding AI’s net impact on emissions is complex, and therefore engagement is key to better navigate the evolving landscape and capitalise on emerging opportunities. For that reason, our senior responsible investment analyst, Greg Kearney, has been probing our holdings in this space, alongside our technology analyst.
The findings are interesting. AI solutions’ increased complexity raises the power needs for manufacturing the smaller, intricate semiconductors now prevalent in servers globally. Emissions from data centres are already estimated at around 3% of the global total, comparable to those from the aviation industry.
But, the energy requirements to train and use AI models are not equal. Some specific models like Intel’s TinyBERT, consumes a fraction of what OpenAI’s GPT-4 requires for similar queries. As generative AI solutions develop, and OpenAI increasingly looks to cement itself as the leader, it is likely that more specific task-based models will be developed for customised needs, and thus the energy demand will grow.
For many companies who are upstream in the AI value chain (particularly semiconductor/ equipment manufacturers) Scope 3 indirect emissions are a material part of the overall picture, totalling over 90% of all emissions. We observed some market leading programs from the companies like ASML, ASMI and TSMC in attempting to reduce emissions within their supply chains. These efforts to take responsibility for indirect emissions and influence the wider supply chain is potentially transformative for overall AI decarbonisation.
Improved efficiencies
Cloud service providers and data centre infrastructure companies have also seen material improvements in efficiency. In 2024 AWS data centres (Amazon’s cloud services business) reported a global ‘power use effectiveness’ (PUE) measure of 1.15. PUE is a metric that measures how efficiently a data centre uses energy. A PUE score of 1.0 is considered perfect, indicating that all energy used by the facility goes directly to computing. The best performing AWS site was in Europe with a PUE of 1.04.
Clearly such statistics, and wider climate targets, are facing heightened headwinds as demand for cloud-based services grow. This has been exacerbated by the scaling of demand for AI solutions and its implicit increased energy requirements.
As a result, we will need to see accelerated action from companies, including more support for low carbon energy procurement that materially reduces emissions, rather than a heavy reliance of renewable energy certificates (RECs). Furthermore, efforts will need take place at local and national levels to ensure more meaningful low carbon sources of energy are secured.
We are also stepping into the unknown in this regard with some of the energy investments proposed by large cloud providers being speculative. While we welcome the foresight around securing future energy technologies, present challenges demand a focus on today’s solutions if climate targets are to be achieved.
Consequently, company emissions trajectories are under strain now. But the potentially transformative opportunity should also not be downplayed. Many in both the technology and industrial sectors are already adopting AI to optimise energy use and drive down emissions.
The aggregate effect could outweigh the impact of a well-managed scaling of data centre expansion. None of this is certain, and all is to play for – but there is a pathway to an elusive ‘cake and eat it’ scenario