I found the data and it includes newer models like GPT 5: https://github.com/Nidhal-Jegham/HowHungryisAIDashboard/blob/main/HHAI%20Snapshot%20Data/DataSnapshotOct26.csv
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 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. It is not quite a return flight every year, but also not a once in a lifetime thing, especially if you start using multiple agents at once. Cheers, Jaro > On 29 Jul 2026, at 13:31, Tom Ellis via ghc-devs > <[email protected]> wrote: > >> 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 > > On Wed, Jul 29, 2026 at 12:45:51PM +0200, Luite Stegeman via ghc-devs 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. > > That's right. > > Here's a different figure: median energy per query of 0.34 Wh[1] > > My *guess* is that an *extremely* intense user of a typical coding > agent with chain of thought might use about one query per 10 > seconds[2], which means a power of > > 0.34 Wh * 3600 s/h / 10s = 122.4Wh / h ~= 0.12kWh / h > > This is 200 times less than the figure you quoted, so even if it's off > by an order of magnitude the conclusion will still be that agentic > coding emits far less CO2 than your analysis. > >>>> 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. > > Using my 0.12kWh / h above, this would be > > 0.3 kg CO2e / kWh * 0.12 kWh / h = 0.036 kg CO2e / 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. > > Based on 3500 kg CO2e per flight, this is the equivalent of > > 3500 kg CO2e per flight / (0.036 kg CO2e / h) > = 97,000 hours per flight equivalent > = 48 years per flight equivalent (2000 hours per year) > = 1 long and hopefully productive career > >>> I think this shows that it is likely that heavy use of LLMs causes >>> significant emissions exceeding that of yearly transatlantic >>> flights. > > My analysis shows that it's nowhere near. It's more like one lifetime > of using agentic coding is equivalent to taking one transatlantic > flight in a lifetime. > > Now of course this may be nonetheless too much already. Energy use > and carbon emissions of agentic coding will rise. Some of my base > figures may be wrong (but they'd have to be *very* wrong for agentic > coding to be worse than other activities we take for granted). > > And I may have made a mistake in my calculations. I would welcome a > correction. > > Tom > > [1] > https://www.microsoft.com/en-us/research/publication/energy-use-of-ai-inference-efficiency-pathways-and-test-time-compute/ > > [2] Probably a massive overestimate (maybe by 10x?): I don't think > I've ever reached that sustained pace -- Moritz might have! -- let > alone the whole working day every day > _______________________________________________ > 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]
