Comment by ktzar

A good way to think about it is finding how much it'd cost to buy and run a GPU that runs a model at around 100tk/s ("thinking" agents are not viable otherwise).

The figure mentioned in the video is not far off