OpenAI and Broadcom unveil Jalapeño, a custom chip for LLM inference

OpenAI adds a custom inference chip to its AI stack

OpenAI and Broadcom have unveiled Jalapeño, a custom processor designed to accelerate large language model inference, according to announcements from OpenAI and Broadcom.

OpenAI describes Jalapeño as its first “Intelligence Processor,” purpose-built for LLM inference and intended to become the first chip in a multi-generation compute platform. Broadcom says the processor is built for current and future LLM workloads and was developed with OpenAI from design to production in nine months.

The Decoder reported that OpenAI expects the chip to run at scale by late 2026. TechCrunch also reported that Jalapeño is a custom-built inference processor designed and manufactured with Broadcom for OpenAI’s inference systems.

Focused on inference rather than training

The key point is that Jalapeño is aimed at inference: the process of running AI models after they have already been trained. For companies operating large language models at high volume, inference can be a major cost and energy concern because each user prompt requires compute.

OpenAI’s announcement says the processor is optimized for LLM inference. Broadcom’s release says early testing shows better performance per watt, though the company did not provide detailed public benchmark figures in the provided announcement excerpt.

That focus puts Jalapeño in a different category from general-purpose GPUs used across both training and inference. OpenAI’s move follows a broader industry pattern in which major AI companies seek more control over the hardware used to serve models, particularly as demand for AI services grows.

Broadcom’s role

Broadcom says it worked with OpenAI on the chip and frames Jalapeño as an LLM-optimized intelligence processor. TechCrunch reported that the chip was built by Broadcom for OpenAI’s inference systems.

The companies have not, in the provided source excerpts, disclosed full technical specifications such as process node, memory architecture, package design, interconnect details, or expected deployment volumes. The available statements emphasize the chip’s purpose, development timeline, and expected efficiency improvements rather than a complete hardware breakdown.

Why it matters

For OpenAI, a custom inference processor could help tailor hardware more closely to the way its models are served. OpenAI says Jalapeño is the first chip in a broader compute platform, suggesting the company does not see this as a one-off design.

For Broadcom, the announcement highlights its role as a partner for companies building specialized AI infrastructure. Broadcom’s investor announcement says the chip was developed from design to production in nine months, a timeline the company presents as part of the project’s significance.

Still, the most important questions remain open. The sources do not provide independent performance results, deployment scale, pricing implications, or comparisons against specific GPUs or other AI accelerators. Broadcom says early testing shows better performance per watt, but public evaluation will depend on more detailed data and real-world deployment.

For now, Jalapeño marks OpenAI’s clearest move into custom inference silicon, with Broadcom as its hardware partner and late 2026 cited by The Decoder as the expected point for scaled operation.