Single medium-sized GPT-5 response can consume near 40 watt-hours of electricity
tomshardware.com
tomshardware.com
Seems a bit much? Even if I assume 1/3rd is the amortised result of the training runs and 1/6th is the cooling system, that only takes me down to the power draw of 2.4 times the H200s' 600W maximum TDP* on the mean inference run.
Semi-related to main theme: why are the links so weird? For me at least, "the published findings" links to chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://arxiv.org/pdf/2505.09598v2 — and I opened this with Safari.
Going to the corresponding arxiv.org page manually, that's not even "the published findings" for the 40 Wh claim, it pre-dates GPT-5.
And the dashboard they link to doesn't even seem to make the claim of 40 Wh for a "medium" length response for GPT-5, unless I'm misreading this somehow, it's claiming 19.7 Wh for a "medium" length query?: https://app.powerbi.com/view?r=eyJrIjoiZjVmOTI0MmMtY2U2Mi00Z...
And I can believe 19.7 Wh. Same maths as above, if that takes 50 seconds to get an answer then it's 1.2 x (H200 TDP).