Why does it matter?
That much effective compute would make the next model generation feel qualitatively different, not just cheaper.
Direct quote
The upshot is that, in my best guess, we should expect another ~100,000x effective compute scaleup—resulting in another GPT-2-to-GPT-4-sized qualitative jump—over four years. Perhaps it will end up being more like 10,000x or 1,000,000x, but the basic picture seems right.
Leopold Aschenbrenner