OpenAI and Broadcom Unveil Jalapeño, an LLM Inference Chip

OpenAI and Broadcom have introduced Jalapeño, OpenAI's first custom-built inference accelerator designed specifically for large language models rather than adapted from general-purpose AI hardware. Architected around OpenAI's own model roadmap and serving needs, the chip was brought to production with Broadcom's silicon and networking expertise (including Tomahawk switching silicon) and Celestica's board and rack integration. Early lab testing shows substantially better performance-per-watt than current leading accelerators, with a full technical report expected in coming months.

Notably, the chip went from design to tape-out in just nine months, a timeline OpenAI credits partly to using its own models to accelerate design and optimization work — framed as one of the fastest ASIC development cycles in advanced semiconductors. This illustrates a feedback loop where AI models help improve the very infrastructure that runs future models.

Jalapeño extends OpenAI's full-stack ambitions beyond models and products into custom silicon, giving the company tighter control over architecture, memory, networking, and deployment to push down cost and latency. Initial gigawatt-scale rollout with Microsoft and other data center partners is planned for late 2026, signaling meaningful shifts ahead in inference economics and responsiveness for engineers building on top of these systems.