Right, so my calculation only works if you assume 1000 reasoning requests per hour, which is admittedly unrealistic. And unfortunately I can't find the raw data for the paper, so I can't calculate the actual value.
> On 29 Jul 2026, at 12:46, Luite Stegeman via ghc-devs <[email protected]> > wrote: > > The table doesn't say 29kW, it says 29Wh for 1500 output tokens with > 10k input tokens, i.e. energy per token, not power. > > On Wed, Jul 29, 2026 at 12:12 PM Jaro Reinders via ghc-devs > <[email protected]> wrote: >> >> I've done a bit of research now, mainly based on Jegham et al. "How Hungry >> is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference" >> >> They list the energy consumption of DeepSeek-R1 (671B params) to be ~29 kW >> at 10k input token and 1.5k output token scales (Table 4). I think it is >> reasonable to expect current frontier models to at least match that since >> the models seem to have gotten larger and I think contexts are generally >> also larger in the use case we're considering (large chunks of GHC would >> have to get loaded into the context presumably). >> >>> From the same study, most datacenters use around 0.3 kg of CO2 equivalent >>> per kWh (Table 1), so DeepSeek R1 uses about 8.7 kgCO2e/h. >> >> That means using a single agent for ~400 hours is equivalent to a return >> flight from Austin Texas to Zurich (3500 kg CO2e; source: online tool), >> which I find an unjustifiable amount of emissions, but I guess some members >> of the community do already make such a trip once a year. >> >> I think this shows that it is likely that heavy use of LLMs causes >> significant emissions exceeding that of yearly transatlantic flights. >> >> Cheers, >> >> Jaro >> >>> On 27 Jul 2026, at 15:46, Tom Ellis via ghc-devs >>> <[email protected]> wrote: >>> >>> On Mon, Jul 27, 2026 at 03:32:45PM +0200, Magnus Viernickel via ghc-devs >>> wrote: >>>> Technological improvements increasing the efficiency of a resource do not >>>> lead to a fall but to a rise in total consumption of that resources. >>>> >>>> That's a well-known effect. See e.g. >>>> https://en.wikipedia.org/wiki/Jevons_paradox. >>> >>> Yes indeed. I do not dispute that the amount of energy going into AI >>> usages is increasing both in relative and absolute terms. >>> >>>>> But how much? >>>> >>>> There are countless projections and surveys like the one shared by Andrei. >>>> We fortunately do not have to rely on bloggers with no considerable >>>> qualification in that area. >>> >>> That's good to hear, and I would welcome someone making a numeric >>> claim and backing it up with citations. I think that would be a >>> valuable contribution to the discussion here. >>> >>> To be clear: *I* am not trying to make a claim, nor refute someone >>> else's claim. I am pointing out that no such claim has been made >>> precise nor substantiated in this discussion! If someone believes >>> strongly that AI energy use should be factored into GHC's LLM policy >>> then the responsibility is theirs to make a precise and substantiated >>> claim. If claims about AI energy usage are not precise and >>> substantiated then I don't see why they should factor into the policy >>> decision. >>> >>> Tom >>> _______________________________________________ >>> ghc-devs mailing list -- [email protected] >>> To unsubscribe send an email to [email protected] >>> >>> >> >> >> _______________________________________________ >> ghc-devs mailing list -- [email protected] >> To unsubscribe send an email to [email protected] > _______________________________________________ > ghc-devs mailing list -- [email protected] > To unsubscribe send an email to [email protected] > > _______________________________________________ ghc-devs mailing list -- [email protected] To unsubscribe send an email to [email protected]
