OpenAI unveiled Jalapeno, a custom LLM inference accelerator built with Broadcom. The silicon combines fixed‑function units and programmable cores to speed up the inference workload behind ChatGPT, Codex, the OpenAI API and upcoming agentic AI services. Unlike Google’s TPU, which handles both training and inference, Jalapeno focuses solely on inference, leaving model training to GPUs.

OpenAI and Broadcom claim the chip reached tape‑out in just nine months, the quickest ASIC development cycle among contemporary semiconductors. Jalapeno sits on a multi‑chip module with an interposer, a central logic tile and eight HBM3E memory stacks, positioning it for high‑bandwidth data access. The design is the first step in a generational compute platform slated for initial deployment toward the end of 2026.

The accelerator signals OpenAI’s push to own more of its inference stack, reducing reliance on third‑party GPUs and potentially significantly lowering latency for end users. By tailoring silicon to its own models, the company can optimize power efficiency and cost as demand for real‑time AI responses grows. Jalapeno’s rollout will give OpenAI a dedicated hardware edge in the competitive generative‑AI market.