OpenAI and Broadcom Unveil Jalapeño, the Company's First Custom AI Chip

OpenAI and Broadcom have unveiled Jalapeño, OpenAI's first custom silicon — an "Intelligence Processor" architected specifically for large language model inference, and the opening move in what the companies describe as a multi-generation compute platform.

The chip was co-developed from initial design to manufacturing tape-out in just nine months, an unusually fast cycle that OpenAI attributes to deep software-hardware co-design, Broadcom's implementation expertise, and the use of OpenAI's own models to accelerate parts of the design and optimization work.

Built for inference

Jalapeño is a reticle-sized ASIC purpose-built for running pre-trained models rather than training new ones. Early testing, the companies say, shows performance-per-watt "substantially better" than current state-of-the-art accelerators. Engineering samples are already running machine-learning workloads in the lab at production target frequency and power — including OpenAI's GPT-5.3-Codex-Spark.

The companies are aiming for initial deployment by the end of 2026, "expanding in the years ahead," and frame the chip as enabling gigawatt-scale data centers alongside Microsoft and other partners.

The custom-silicon arms race

Jalapeño puts OpenAI in the same camp as Google (TPU), Amazon (Trainium and Inferentia), Meta (MTIA) and Microsoft (Maia) — hyperscalers designing their own chips to cut dependence on Nvidia and squeeze more inference out of every watt. It also deepens Broadcom's role as the go-to partner for custom AI accelerators, a business that has become one of the most valuable franchises in semiconductors.

For OpenAI, whose compute costs scale directly with usage, owning the inference layer is less a hedge than a necessity. Jalapeño is the first proof that it intends to control that layer itself.