@bhg
Thanks for sharing your thoughts.
Here are the actual numbers
Train Gemini Ultra ~150 GWh
Distill Nano from Ultra ~1–5 GWh / ~1–3%
One Gemini cloud query 0.24 Wh / 0.00000016%
1 year of Gemini cloud serving (~1B queries/day) ~90 GWh / ~60%
👉One Nano on-device query ~0 datacenter-side (phone battery) 0% 👈
Aggregate inference passes the training run in roughly 18–24 months at frontier-deployment scale and that's only counting the per-query number Google chose to publish. Query volume is a rough order-of-magnitude estimate
Gemini Ultra burn would have occurred regardless of Nano deployment.
Nano distill is a one off expenditure of SA grid demand for a day. After that free model with phone recharge.
Maths.
#AiEnergy #AIcosts #Ai