Compute Gap Debate: Stop Moving the Goalposts

teortaxesTex · x · 2026-07-18

The post responds to claims questioning why the US-China compute gap hasn't shown up, arguing a core point: US labs already have massive numbers of chips online and have invested billions in GPU R&D. These costs shouldn't simply be dismissed as something easily erased by distillation.

The author further challenges the narrative of "waiting until the end of the next year to see." Every time a prediction fails to materialize immediately, the reasoning shifts to "booked compute isn't online yet," "the gap will emerge after the next expansion," or "let's wait for the next training cycle." This continuous moving of the goalposts without ever clearly validating the gap makes the argument increasingly untenable.

The text in the attached image reinforces the same message: explanations regarding chip inventory, signed-but-not-online compute, RL phases, and chip yields cannot be postponed indefinitely, lest they devolve into a perpetual "just wait another 18 months" narrative.

Original post →

More from Infra

Infra channel →