OpenAI and Broadcom unveil Jalapeño inference ASIC to bypass NVIDIA GPU dependency

OpenAI and Broadcom jointly unveil Jalapeño, a custom ASIC for LLM inference. OpenAI handles architecture, Broadcom handles tape-out and networking (including Tomahawk 5 switch chips), Celestica provides board/rack integration. Tape-out achieved in 9 months, claimed fastest ASIC development cycle in AI accelerators. Engineering samples validated, running GPT-5.3-Codex-Spark workloads. Early tests show significantly better performance-per-watt than current state-of-the-art. President Greg Brockman calls it part of long-term full-stack infrastructure strategy. Hardware lead Richard Ho says optimization focuses on critical kernel, memory movement, networking, and service model. Jalapeño is first step in multi-generational compute platform, targeting large-scale deployment by end-2026 with gigawatt-scale datacenter clusters.