I really wish these models were available via AWS or Azure. I understand strategically that this might not make sense for Google, but at a non-software-focused F500 company it would sure make it a lot easier to use Gemini.
Hence, a terabyte of text would cost about $42,000 to run through a pareto-frontier "cheap" model.
The most expensive cloud egress fee I could quickly find is $185 per terabyte (Azure South America Internet egress).
Hence, AI processing is 200x as expensive as bandwidth, or put another way, even a long-distance international egress at "exorbitant cloud retail pricing" is a mere 0.5% of the cost.
Petabytes, exabytes, etc... just adds digits to both the token cost and bandwidth cost in sync and won't significantly shift the ratio. If anything, bandwidth costs will go down and AI costs go up because: output tokens, smarter models, retries, multiple questions for the same data, etc...