Comment by billisonline

> If tpu's actually breakout as a viable alternative over the next few years

Why haven't they broken out yet, I wonder, if they're more efficient for inference and LLM costs are now weighted towards inference over training?