vLLM claims 90x throughput on Rubin as analysts ignore software-driven gains

bookwormengr · x · 2026-11-17

The author argues stock analysts estimating AI supply/demand work too abstractly and miss how far software and architecture optimizations go within a single hardware generation. Citing the vLLM blog: with Rubin and recent optimizations, throughput can reach over 90x. He gives a hypothetical example of a model delivering 80x performance at 1/3 the size (model names are speculative, unverified), estimating up to 9000% gains at high interactivity and 300-400% at 100 tokens/sec, concluding no serious analyst models this and everything gets hand-waved with the Jevons paradox.

Original post →

More from Venture

Venture channel →