OpenAI custom chip Jalapeño beats Nvidia in tests

The Jalapeño announcement dominated the 37th Hot Chips conference, less than a year after OpenAI's Broadcom announcement. The chip is not an ASIC but a full alternative that reportedly beats Blackwell-class systems. SemiAnalysis framed it as unusually strong for a first-generation chip, comparing it directly against Blackwell and Rubin-class systems.

A second-order story is model-assisted systems optimization. GPT-Astra + Codex helped write and optimize low-level kernels, bringing three previously unplanned open-weight models to high performance in about two months. This suggests compiler and kernel work is increasingly being folded into the model improvement loop, not just application-layer coding.

Multiple technical reactions noted Jalapeño performed well even without tricks like aggressive prefill/decode disaggregation or speculative decoding in some setups, while beating systems that did use them. Several posts tie this to a larger industry transition where frontier labs may no longer be strictly downstream of NVIDIA for inference economics, though packaging and foundry capacity remain a hard bottleneck.