Probably from Nvidia's GTC keynote: https://www.youtube.com/live/USlE2huSI_w?t=2995.

In the keynote, Jensen uses 1.8T in an example and suggests that this is roughly the size of GPT-4 (if I remember correctly).