OpenAI has unveiled Jalapeño, its proprietary AI accelerator chip, which achieves up to 13.4 petaflops of 4-bit compute and utilizes 232 GB of memory with a 15.4 TB/s bandwidth. The chip reportedly reduces end-to-end latency by up to 3.6 times compared to Nvidia’s GB300. To achieve this, OpenAI utilized a rapid design cycle, moving from architecture concept to first silicon in under 20 months, with only nine months between RTL generation and tape-out. This speed was facilitated by a small team of fewer than 100 people leveraging large language models (LLMs) to accelerate front-end workflows. Specifically, the team used the Accelerated Hardware Synthesis (XLS) chain to convert programming languages like C++ and DSLX into Verilog, a process highly compatible with LLM capabilities. OpenAI also used internal, fine-tuned models and precursors to GPT-6 Astra to optimize software; for instance, they improved a multi-head latent attention kernel benchmark from 0.31% to 88.94% of the theoretical ceiling in approximately 40 hours. While Broadcom managed the physical backend design, OpenAI’s AI-guided optimization achieved a 10% area reduction in matrix multiplication units compared to human-optimized baselines. The company plans to deploy Jalapeño in pods containing 2,048 chips.