What’s all the buzz about AI & energy?

Issue 72 l Eka’s Weekly Roundup (26 April 2024)

Apr 26, 2024

Chat GPT is reported to use more than half a million kw hours on a daily basis. To put that into context, the average US households uses just 29 kw hours in the same time period.

The industry press is picking it up. A hot paper from earlier this year predicted that the data centres which power AI development and operation could be using as much energy as Sweden by 2027. In 2022, Google reported that machine learning accounted for about 15% of its total energy use over the prior three years.

So what’s happening under the hood?

Breaking down the AI supply chain 🗞️

The Artificial Intelligence supply chain, from chip manufacture to compute, has a variety of environmental impacts.

Source

These broadly fall into three distinct categories: production, transportation, and operations.

On the production, estimates suggest that 15% of total data centres GHGs came from the production phase. However, this is potentially set to increase as a % of total data centre emissions the actually running of data centres transitions to renewable energy sources, making the absolute carbon emissions decrease.

On transportation, it is difficult to disaggregate the impact of ‘AI specific’ components relative to broader ICT components. The few estimates which exist put the transport of AI compute hardware at less than 5% of the total GHGs emissions over an AI systems lifetime. In other words, we don’t think this is the highest impact value segment, and will become less important as critical supply chains are near-shored across the world.

Operations is the largest impact area of AI compute. One example is the Irish market which has a significant amount of data center operations. Electricity consumption by data centres increased +144% from 2015 to 2020, accounting for 11% of metered electricity consumed in the country in 2011. Another example is a Greenpeace report which estimated Chinese data centre electricity consumption at 161 TWh in 2018, of which 75% came from coal-fired power stations (this will likely have changed with the growing % of renewables on their grid). This will most likely transition to renewables in the medium term, but will remain material in the short term.

Training AI models remains particularly costly (both financially and environmentally). We show this slightly outdated benchmarking below, outlining estimates for AI model trainings versus other CO2e footprints.

Source

The obvious pushback here is that AI can be used for highly impactful activities. Deepmind published an article in 2023 covering the material benefits which AI could introduce for weather, climate, and broader impact goals.

There are innovators in the UK like Oriole Networks or Literal Labs which are trying to target the model architecture of LLMs while remaining energy efficient. We believe there is a world where we can have our cake and eat it, when it comes to AI & Sustainability.