OpenAI Used Unreleased Model Astra to Design Its Jalapeño Inference Chip in 9 Months

jarrodwatts · x · 2026-08-26

OpenAI leveraged its unreleased model Astra to help develop its own inference chip, Jalapeño, finishing the design in just 9 months. They claim 1.5-1.9× more tokens per kilowatt and 1.7-3.6× lower end-to-end latency.

Original post →

More from Infra

Infra channel →