You can ask it to verbatim produce training data and that takes very little compute for a lot of output tokens

i dont think you understand how these models operate.

You can burn kilowatts generating 10 tokens, and conversely produce millions of tokens burning very few watts. You're comment is horse shit