OpenAI: first custom inference chip Jalapeño, built with Broadcom

OpenAI unveiled Jalapeño, its first custom-built inference processor, co-designed and manufactured with Broadcom. The announcement came on June 24, 2026 — roughly eight months after the companies formally disclosed their partnership in October 2025.

The chip is purpose-built for inference: running a trained model against live user queries, as distinct from training. Early testing data cited by OpenAI claims “significantly better performance-per-watt than current state-of-the-art alternatives.” OpenAI emphasized that its own AI models participated in designing the architecture, positioning the chip as a product of the full-stack approach the company now calls “operating across the stack” — from silicon through the end-user product.

The strategic logic mirrors what Google (TPU), Amazon (Trainium), and Meta (MTIA) have each done over the past several years: reduce dependence on Nvidia’s H100 and H200 GPUs by building hardware optimized for the specific inference patterns of their own models. Custom silicon gives vertically integrated companies an architectural moat that competitors using commodity GPU infrastructure cannot easily replicate.

For product managers working with AI-powered products, this shift has two practical implications. First, as inference becomes cheaper for OpenAI, that pressure eventually flows downstream to API pricing — relevant for any team building products on top of OpenAI models. Second, the pattern of frontier labs owning the full hardware-to-product stack reinforces the structural concentration of the AI infrastructure market. Teams building on third-party APIs should account for the fact that the cost and capability economics of those APIs are now shaped by proprietary silicon decisions made by their suppliers.