OpenAI Deploys Custom Jalapeño AI Chip Paired with AMD EPYC Turin CPUs
OpenAI has internally deployed its custom Jalapeño ASIC, an AI chip tailored for large language model inference co-developed with Broadcom. The company chose AMD EPYC Turin server processors paired with 1.5TB of memory as host CPUs due to pragmatic considerations over Nvidia's newer Vera CPU architecture. This deployment marks a major milestone in OpenAI's efforts to reduce hardware reliance on Nvidia while highlighting AMD's expanding role as a host provider in custom AI infrastructure. By prioritizing ecosystem maturity and deployment speed, OpenAI demonstrates a practical approach to scaling its internal compute capacity. According to Richard Ho, OpenAI's VP of Hardware, AMD's Turin platform provided the required performance and mature partner support needed to de-risk the Jalapeño deployment. The Jalapeño ASIC is currently restricted to OpenAI's internal workloads, with no present plans for commercial sale.
## BACKGROUND
An Application-Specific Integrated Circuit (ASIC) is a microchip customized for a dedicated task, offering higher efficiency and lower cost for AI inference compared to general-purpose GPUs. AMD's 5th Generation EPYC Turin server processors are built on the Zen 5 microarchitecture, offering high core counts and broad enterprise ecosystem support. In contrast, Nvidia's Vera CPU is a specialized processor platform designed to power agentic AI workloads alongside Nvidia GPUs.