OpenAI has unveiled its first custom-built AI inference processor, developed in partnership with Broadcom and named Jalapeño, according to TechCrunch. The chip went from initial design to manufacturing in just nine months, and early test results show significantly better performance-per-watt than current alternatives, Business Today reports.
Jalapeño is designed specifically for LLM inference — the process of running pre-trained models in response to user queries — rather than the far more compute-intensive pre-training workloads. OpenAI president Greg Brockman framed the chip as targeting "specific workloads that are underserved" by existing hardware. The chip is the first in what Broadcom CEO Hock Tan described as a "multi-generation roadmap," with gigawatt-scale data centre deployments planned with Microsoft beginning later this year.
OpenAI's own AI models assisted in the chip's design, and the processor can run both OpenAI's models and third-party LLMs. The announcement formalises a partnership first revealed in October 2024, and places OpenAI alongside Google and Amazon — both of whom have built custom AI accelerators — in the growing club of hyperscalers designing their own silicon.
Analysis
Nvidia is the silent target of every sentence in this announcement. OpenAI currently spends an estimated $4-6 billion annually on Nvidia GPUs for inference alone, and Jalapeño is designed to slash that line item. The chip doesn't need to beat Nvidia on raw performance — it just needs to be cheaper per token on the specific inference workloads that dominate ChatGPT, Codex, and the API business. The real winner here is Broadcom, which positions itself as the arms dealer in the coming AI chip fragmentation war: it now has custom silicon deals with Google, Meta, and OpenAI, each designing chips that reduce their dependence on Nvidia. The loser isn't just Nvidia's revenue line — it's the entire thesis that general-purpose GPU compute is the only viable AI hardware strategy. If every major lab follows OpenAI into custom inference silicon, Nvidia's data centre margins, currently north of 70%, become a target rather than a moat.
