I put in a paper for #loco2026 (deadline 29 May EOA)
"Estimating the growth in emissions from AI data centres"
https://lancaster-university.github.io/loco2026/
(1/n)
Replies:
4
Boosts:
4
The aim of the paper is to model the emissions ensuing from the growth scenarios put forward by the IEA, McKinsey and the AI industry. We created a very detailed model for this. Here is a sneak preview of one of the graphs in the paper.
(2/n)
The graph shows an extension of the McKinsey high-growth scenario until 2045. This is exponential growth, 10x in 10 years, 100x in 20 years.
Note that Michael Dell said we'd need 100x in 10 years, and Altman said something similar about growth in AI servers.
So if the expectations of the AI industry would be realised, by 20245 AI alone would cause more emissions than the entire global carbon budget.
#FrugalComputing
(3/n)
I have also looked at the IEA scenarios. They use a bounded growth model, and their projections are only until 2035. But even with that model, in the worst case we'd see emissions of more that 1 GtCO2e/y (about 10% of the global carbon budget) purely from AI.
#FrugalComputing
(4/n)
The model takes into account the trends in embodied carbon (in fact, that is the main contribution of the research) and in global carbon intensity of electricity generation. For the latter I am assume we stay on trend, although it is conceivable that the energy hunger of AI would slow down the decarbonisation of energy production.
#FrugalComputing
(5/n=5)
