OpenAI Reveals AI Accelerator Chip 'Jalapeño' Designed Using Its Own LLM

OpenAI revealed details of the AI accelerator chip 'Jalapeño', designed using its own LLM. The design process centrally utilized OpenAI's internal fine-tuned LLM and Google's open-source high-level synthesis tool XLS. Initially, the o3 model was used, but in the latter stages of the project, access was gained to a predecessor model of GPT-6 Astra, advancing to the stage of directly generating Verilog code without XLS conversion. In the software optimization phase, internal AI models were used to improve the benchmark performance of DeepSeek's multi-head latent attention kernel from 0.31% of the theoretical upper bound to 88.94% in approximately 40 hours. OpenAI collaborated with Broadcom on the physical design (backend), and Ho and Leary cautioned against the misconception that full automation is possible with Codex alone, clarifying that AI is a tool for small teams to achieve high-quality results and shorten timelines.