OpenAI and Broadcom unveiled Jalapeño on June 25 and called it OpenAI’s first “Intelligence Processor,” an inference accelerator for large language models.

The US chip designer and OpenAI said they are targeting an initial deployment by the end of 2026 as OpenAI builds more of its computing stack.

OpenAI said it designed the chip with Broadcom and Celestica, a Canada-based electronics manufacturer.

Broadcom handled silicon implementation and networking, while Celestica contributed board, rack, and system expertise.

OpenAI said engineering samples are already running machine learning workloads in the lab, including GPT-5.3-Codex-Spark.

The program follows OpenAI’s October 2025 collaboration with Broadcom for 10 gigawatts of custom accelerators, as hyperscalers pursue inference chips to cut costs and power use.

Broadcom also has an established custom-silicon business, with a projected 60% share of the AI server compute application-specific integrated circuit market by 2027, according to an external analysis.

OpenAI has not released standardized Jalapeño benchmarks and said it is still measuring final performance.

A large-scale rollout will also depend on data center power availability.

🔗 Source: OpenAI

Recent OpenAI developments