Because we are not given accurate, fully transparent, reliable information by Big Tech, people like Hannah Ritchie are making guesstimates. I’m not sure what she is including here? The upfront or now emissions of the data centres, hardware and other infrastructure? Here is another guess from Elian Pusceddu
1kgCO2/day or 0.37 tCO2/year, ot >5% of per capita emissions in the west.
Can you think of anything else that will increase emissions that much? – Elian Pusceddu
Here is his maths
‘I think you are off by a factor of ~10, based on the sources you use. I have not cross-checked with others so maybe you know better?
The informal estimate on Medium referenced by your ‘Best estimate’ paper reviewed the original data landing on ~25 tCO2/day for 5m queries/day -> ~5g/query just from running the model.
The retraining estimate to be amortised should also doubled due to the revised 10m->5m queries/day, landing at ~3.6g/query.
So we are at ~8.6g/query.
And then again there are significant gaps:
– the embedded carbon in *all* of the assets (data centres, renewable generation, infrastructure etc.)
– much of the new-build energy generation to meet increasing demand for ML is based on methane. Marginal emissions intensity >500g/kWh
Let’s optimistically round it up to 10g/query.
And let’s acknowledge that when AI is embedded in all sorts of products and services, the number of equivalent queries will be way higher than 10/day. Let’s say 100 (optimistic I’d say, given how many people spend >5h/day in front of screens.
If the estimate is OK, that’s 1 kgCO2/day or 0.37 tCO2/year, ot >5% of per capita emissions in the west.
Can you think of anything else that will increase emissions that much?’ – Elian Pusceddu Decarbonisation | Cleantech | Strategy
AI is typically deployed in 20-30 cabinet clusters at or above 40 KW per cabinet. This represents a fourfold increase in KW/ cabinet with the deployment of AI. The difference is staggering.
A typical Chat-GPT query uses about 10 times more energy than a Google search – and that’s just for a basic generative AI function. More advanced queries require substantially more power that have to go through an AI Cluster Farm to process large-scale computing between multiple machines – Stephen Spinazzola
“Under the Climate Action Plan, renewables were meant to replace existing fossil fuel use by the country generally – not to feed new demand from an unsustainable industry. Ireland’s overall electricity demand grew by 2.6pc annually since 2015, but data centre demand grew by 22.6pc annually….Not only did this mean renewables were not displacing fossil fuels, but data centres were actually driving increased fossil fuel use.” – Hannah Daly
Resist AI 