Yes, this assumes that a user uses the model 100% of the time, which may not be realistic. I do expect people to also use LLMs outside of working hours to hack on GHC. Is it uncommon to use LLMs overnight to churn away at some task?
Also, in an attempt to extrapolate to newer models, we can compare open model inference pricing. Kimi K2.7 beats GPT 5, so we can assume that as a baseline and compare it to Kimi K3 which is arguably on par with frontier models. If we compare the prices of these two models we get a 4.3x increase ($3.5/Mtok to $15/Mtok). If we assume that the margins are slim and most of the price increase comes down to a difference in energy use, we can conclude that frontier models require 4.3x more energy to run than GPT 5. If that is the case the number is down to 3500 hours of full time LLM use for 1 transatlantic return flight equivalent. Cheers, Jaro > On 29 Jul 2026, at 14:00, Tom Ellis via ghc-devs > <[email protected]> wrote: > > On Wed, Jul 29, 2026 at 11:42:50AM +0000, Jaro Reinders via ghc-devs wrote: >> It lists the data center efficiency (pue) as 1.12, gpu power usage >> (pgpu) as 5.6 kW, non gpu power usage (pngpu) as 4.6 kW, lower bound >> on gpu utilization (ugpu) as 0.055, and non gpu utilization (ungpu) >> as 0.0625. >> >> Using the formula E = pue * (pgpu * ugpu + pngpu * ungpu), we find >> that GPT 5 uses about 0.67 kW to run > > But how many users is it serving? Does "utilization" imply > "utilization per concurrent user"? Even if so, bear in mind that one > concurrent user is not one *actual* user! > >> but the CO2e/kW of GPT is also higher at 0.34. We come to 228 grams >> CO2e per hour, or about 15000 hours (less than two years of >> continuous use) for a transatlantic return flight equivalent. > > But people don't work 24 hours a day. A rough working year is 2000 > hours, so this is still 7.5 years' worth of work. But I suspect > that might be to serve *one concurrent user*. Most users are nowhere > near fully saturating the capacity of an LLM endpoint. > >> It is not quite a return flight every year, but also not a once in a >> lifetime thing > > Maybe, maybe not. I don't think the analysis has definitively > determined that yet. I understand you may want to err on the side of > caution. That is your right. You also may want others to err on the > side of caution. It would be preferable to have hard data so that we > don't have to err! > >> especially if you start using multiple agents at once. > > Indeed. The total energy use and carbon emissions of agentic coding > will rise. > > 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]
