Power Hungry: The Hidden Cost of Artificial Intelligence

Interest in Artificial Intelligence (AI) surged with the announcement of OpenAI's ChatGPT in November 2022. While large language models had been in development for years, ChatGPT's ease of use, conversational style, and OpenAI's ability to position itself as a leader in the space captured the public's imagination. Additionally, despite now being considered a potential threat to OpenAI, Elon Musk (one of the company's original founders in 2015) played a significant role in bringing OpenAI and ChatGPT into the spotlight.

While WPF doesn't directly hold Open AI, we do hold positions in the below:

Our exposure to AI:

• Microsoft (£48.2m)
• Amazon (£43.4m)
• Apple (£29.6m)
• Alphabet (£34.9m)
• Meta (£5.7m)
• and Nvidia (£26.9m)
• all of which are either working directly with OpenAI or have made significant investments into AI

You can find out more on our Equity Holdings page where we publish quarterly updates. 

The buzz around AI has intensified as more companies make similar investments into AI, allocating increasing resources into developing AI models as well as the supporting hardware, resulting in a seemingly ever-expanding list of use cases. However, AI models and the hardware that supports them are power hungry, with an average ChatGPT query requiring ~10 times as much electricity to process as a Google search. To contextualise this, the power required to resolve a single ChatGPT query is enough to power an average LED bulb for roughly 20-30 minutes.

According to Goldman Sachs, data centre workloads tripled between 2015 to 2019 (just prior to the current AI boom), but these increases were met by increases in efficiency - this meant that power consumption by data centres stayed relatively stable. However, current estimates by the International Energy Agency suggest that the global electricity consumption of data centres could increase by ~128% by 2026 (when compared with 2022), and this has largely been attributed to AI and cryptocurrencies. Both Google and Microsoft have announced significant increases in their emissions, a notable portion of which are the result of these companies integrating AI more heavily into their products. Some researchers are concerned that the hardware that AI models are most reliant on (Graphics Processing Units) aren't likely to be able to offer the efficiency improvements required to address AI's increasing hunger for electricity.

It isn't all doom and gloom though. Start-ups, researchers and industry leaders are actively working to address the issue. Some, including Intel and Analog Devices (two companies which Wiltshire Pension Fund holds positions in, valued at £631k and £1.7m respectively), are looking towards analogue computing. Unlike digital devices, analogue devices don't just work in binaries - they work with the in-between - therefore they can store more data in a given area. Additionally, "because the difference between analogue states can be smaller than those between the widely separated [binary states], it takes less energy to switch between them." That being said, analogue devices have issues of their own, for example, they produce noise, it may be necessary to perform digital-to-analogue conversions, and they lack the clarity of signal that makes digital devices so reliable. Other approaches are being considered, although they all come with their own unique challenges and trade-offs. In another example, Google are exploring the possibility of using small nuclear reactors to generate the clean energy required to power its AI data centres.

Meanwhile, some datacentre operators are focusing on improving the efficiency of datacentres themselves. As part of our award-winning Climate Opportunities portfolio, Wiltshire Pension Fund is backing a project which is focusing on improving cooling, reducing energy use and CO2 output by up to 50% versus traditional air-cooled data centres.

AI has many potential benefits, such as helping us achieve a more precise understanding of dark energy, improving our general understanding of the world, enabling businesses to operate more efficiently, and AI may even help us to mitigate up to 10% of greenhouse gas emissions by 2030. While the opportunities are huge, and it's likely that we'll see AI models become more energy-efficient over time as their use cases and designs become more refined, it's reassuring to learn that some effort is being made to address the rise in emissions resulting from our increasing use of AI in the here and now.

As touched on earlier, Wiltshire Pension Fund holds positions in Microsoft, Amazon, Apple, Alphabet, and Meta, and so any increase in their emissions will impact the fund's carbon footprint. We will therefore be closely monitoring their progress and, if necessary, we'll engage with each company to ensure they're working to reduce emissions - in line with our goal for our investment portfolios to be net zero by 2050.

Written by Thobeka Kellett

Investment Analyst