OpenAI and Broadcom have unveiled Jalapeño, a custom ASIC purpose-built for large language model inference, marking a strategic push toward vertical integration in AI infrastructure.
Design and development
Jalapeño is an application-specific integrated circuit (ASIC) architected from the ground up for LLM inference workloads. Broadcom developed the chip over nine months, leveraging detailed architectural insights from OpenAI researchers and incorporating OpenAI’s roadmap for future models.
The chip targets data center deployments, with both companies stating it is the first generation of a long-term silicon project that will be refined over successive iterations. Broadcom’s existing custom-chip business for hyperscalers provided the manufacturing and integration expertise required for this effort.
Performance claims and testing
OpenAI reports that early testing indicates Jalapeño will deliver “substantially better” performance per watt than current state-of-the-art inference hardware. However, the company has not completed benchmarking and will release a detailed technical report in the coming months.
Until that data is published, independent validation of the chip’s efficiency gains remains unavailable. The specialization for LLM inference—as opposed to general-purpose compute—is the primary differentiator versus existing data center hardware.
Strategic implications
The partnership reduces OpenAI’s dependence on external suppliers like NVIDIA, a key driver as the company seeks to own the full stack from silicon to service. This vertical integration strategy mirrors moves by other frontier-model builders and hyperscalers facing a global compute crunch.
For Broadcom, the deal expands its custom-chip business beyond traditional networking and storage into the high-growth AI inference segment. Both companies expect Jalapeño chips to be deployed in data centers by the end of this year.
Forward outlook
Jalapeño represents a concrete step toward specialized, energy-efficient inference hardware at scale. If the claimed performance-per-watt advantages hold, it could pressure incumbent GPU suppliers and accelerate the industry shift toward domain-specific AI accelerators. The coming technical report will be critical for assessing whether this custom silicon truly delivers on its promise.
