Overview
- OpenAI and Broadcom publicly revealed the Jalapeño Intelligence Processor on Wednesday and delivered engineering samples that OpenAI says are already running workloads including GPT‑5.3‑Codex‑Spark at target frequency and power.
- Jalapeño is an application‑specific integrated circuit built for inference, which means it is optimized to generate model responses for chatbots and coding agents rather than to train new models.
- Broadcom CEO Hock Tan said early tests show roughly 50% lower operating cost versus typical AI GPUs, but independent benchmarks and large‑scale deployments have not yet confirmed that figure.
- OpenAI said it completed the chip design in about nine months, sent the design to TSMC for fabrication, and will work with Broadcom, Celestica and cloud partners such as Microsoft to move from prototypes late in 2026 to broader scaling in 2027–2028.
- If validated at scale, the move to custom ASICs could reduce OpenAI’s dependence on Nvidia GPUs, change the economics of serving LLMs, and increase demand for TSMC capacity and high‑bandwidth memory with real effects on cloud costs and data‑center planning.