OpenAI and Broadcom Unveil Chip Optimized for LLM Inference

OpenAI and Broadcom have unveiled Jalapeño, their first intelligence processor optimized for LLM inference. Unlike existing general-purpose accelerators, this chip adopts a 'blank-slate' design tailored to LLM kernels, memory movement, and networking patterns.

Key features of Jalapeño include:

• High power efficiency: Initial tests show substantially improved performance-per-watt compared to existing state-of-the-art (SOTA) accelerators.
• Full-stack strategy: OpenAI is pursuing vertical integration of its infrastructure that extends beyond model development into chip design.
• Scalability: By combining Broadcom's Tomahawk networking technology and other elements, the goal is gigawatt (GW)-scale data center deployment.

It is currently being tested on major workloads including GPT-5.3-Codex-Spark, and is set to be deployed in large-scale data centers starting in 2026 alongside Microsoft and other partners.