OpenAI and Broadcom reveal a custom chip for ChatGPT and Codex
OpenAI and Broadcom have introduced Jalapeño, OpenAI’s first custom AI chip. It is built for inference, the work of generating answers in services like ChatGPT, Codex, and the API. OpenAI says the uses what it has learned from running models, products, servers, s, networking, and scheduling at large scale.
Test chips are already running machine-learning workloads in the lab at the target power and speed, including GPT-5.3-Codex-Spark. OpenAI has not released final numbers yet, but early tests show much better than today’s leading systems. Jalapeño will be part of a multi-generation platform with Broadcom’s chip and networking technology, including Tomahawk networking silicon, plus Celestica’s board, rack, and system work.
The first deployment is planned by the end of 2026, with Microsoft and other partners helping expand it to gigawatt scale. OpenAI says the goal is faster ChatGPT replies, Codex tasks with less waiting, cheaper API products, and more reliable access when demand is high.
Key points
- Jalapeño is OpenAI’s first custom AI chip, built with Broadcom.
- It is designed for inference in products such as ChatGPT, Codex, and the API.
- OpenAI says early tests show better than current top systems, but final numbers are not public yet.
- Test chips are already running lab workloads, including GPT-5.3-Codex-Spark.
- The first rollout is planned by the end of 2026 and is meant to scale through large .