AI is big in GPU terms but in raw bytes the whole token economy is laughably tractable.
Case in point: OpenRouter is serving 60T tokens a week, but this is all human text and code (and cache hits!), which is almost certainly compressible enough that you could fit the whole week’s usage on a single hard drive.
Case in point: OpenRouter is serving 60T tokens a week, but this is all human text and code (and cache hits!), which is almost certainly compressible enough that you could fit the whole week’s usage on a single hard drive.