Overview
- OpenAI and Broadcom announced the chip on Wednesday, June 24, 2026, and said engineering samples are already running workloads in OpenAI’s labs.
- Jalapeño is an application‑specific integrated circuit built for LLM inference and OpenAI reports samples have run GPT‑5.3‑Codex‑Spark at target frequency and power.
- The companies claim the chip delivers substantially better performance per watt than current state‑of‑the‑art hardware but have not released full independent benchmarks.
- Broadcom handled silicon implementation and networking, Celestica is doing board and rack integration, TSMC will manufacture the chips, and partner deployments including Microsoft are planned to begin in late 2026 with scale‑up through 2027–28.
- The move signals a shift to a vertically integrated stack to reduce reliance on Nvidia for inference costs, though training workloads remain GPU‑dominated and key questions remain about volumes, pricing, and long‑term flexibility if model designs change.