OpenAI unveils Jalapeno, its first AI inference chip, with Broadcom
Broadcom's chief executive said early samples cut inference cost roughly 50% against typical GPUs, a self-reported figure with no disclosed comparison baseline.
- Compute & infrastructure
- Major
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom-designed chip, an accelerator built specifically for inference rather than training. The two companies described it as the first in a multi-generation compute platform, with Broadcom handling chip development and networking and Celestica manufacturing and integration. OpenAI said its own models had been used to speed parts of the design process, and that engineering samples were already running workloads in its labs.
The stated case for the chip was cost and efficiency: OpenAI said Jalapeño delivered substantially better performance per watt than current alternatives on the inference workloads it tested — the responses behind ChatGPT, Codex and API calls. Broadcom chief executive Hock Tan said early testing showed roughly 50% lower cost than typical AI GPUs, and placed the chip’s performance alongside Nvidia’s Blackwell line and Google’s tensor processing units. Those figures were self-reported by the companies bringing the chip to market, tested against workloads of their own choosing, without an independently disclosed comparison baseline.
OpenAI president Greg Brockman said the chip reflected “a deep understanding of the workload” gained from operating ChatGPT, Codex and the API at scale, and framed it as part of a long-term strategy to control more of the infrastructure stack rather than depend solely on Nvidia. The design-to-fabrication-readiness cycle took about nine months, which the companies and trade press described as unusually fast for a new processor family. Detailed technical specifications, including die size and process node, were not disclosed at unveiling.
The announcement followed OpenAI and Broadcom’s October 2025 agreement to jointly develop custom AI accelerators, and placed OpenAI alongside Google and Amazon in the group of major AI buyers building their own silicon rather than relying only on merchant GPUs. Deployment at gigawatt scale, including through data-centre partners such as Microsoft, was planned to begin by the end of 2026.