OpenAI unveils its first custom chip, built by Broadcom
Points and comments are a snapshot, not live.
OpenAI unveiled Jalapeño, its first custom inference chip made with Broadcom.
OpenAI announced Jalapeño, a custom inference processor designed and manufactured with Broadcom. The chip targets inference workloads, emphasizing lower operating costs for real-time coding models. OpenAI's AI models assisted development. Early tests show better performance-per-watt than current alternatives. OpenAI president Greg Brockman said the company sought underserved workloads to accelerate. The move reduces reliance on Nvidia GPUs for inference, following similar efforts by Google and Amazon. Pre-training may still use Nvidia hardware. OpenAI said optimizing infrastructure across chip, kernel, memory, and networking layers aims to make models faster, more reliable, and more affordable.
What commenters are saying
Commenters noted Broadcom's strategy: becoming wealthy as Google's TPU hardware partner and now doing the same for OpenAI, sharing TSMC capacity. Some questioned the chip's focus on inference rather than training, reasoning inference costs now exceed training costs. One commenter argued memory bandwidth, not compute, is the bottleneck for small-form-factor AI hardware, using the M3 Ultra's wide bus as an example. Another criticized the "Jalapeño" name as cringe and California-centric, comparing it to Lord of the Rings-themed company names.