OpenAI Unveils Jalapeño Chip to Curb Nvidia Reliance

The new silicon was created with the assistance of OpenAI’s own machine learning models, which helped refine the chip's architecture for specific inference tasks. While the hardware remains in testing, early internal data suggests a marked improvement in performance-per-watt compared to current market alternatives. President Greg Brockman noted that the project stems from a desire to address underserved workloads, focusing on the specific demands of running pre-built models rather than the compute-heavy pre-training phase.

This strategic pivot aligns OpenAI with tech giants like Google and Amazon, both of which have developed custom accelerators to manage their massive machine learning footprints. By controlling the stack from kernel and memory systems up to the product experience, OpenAI intends to make its models faster and more affordable. While Nvidia hardware will likely continue to power the most intensive training cycles, the transition to custom silicon for inference represents a vital shift in the company’s long-term economic strategy.