OpenAI Unveils "Jalapeño," Its First Custom LLM-Inference Chip Built with Broadcom

OpenAI Unveils "Jalapeño," Its First Custom LLM-Inference Chip Built with Broadcom

In a major milestone for its hardware-software vertical integration strategy, OpenAI officially unveiled Jalapeño, its first custom-designed AI chip. Developed in close partnership with Broadcom and Celestica, Jalapeño is an "Intelligence Processor" designed from the ground up specifically for large language model (LLM) inference rather than general-purpose training.

Key Specifications and Architecture

Jalapeño is a reticle-sized Application-Specific Integrated Circuit (ASIC). Unlike general-purpose accelerators adapted from earlier AI workloads, Jalapeño is optimized entirely for the kernels, memory movement, and high-speed networking that dictate LLM inference performance.

• Performance: Early laboratory testing shows that Jalapeño delivers performance-per-watt metrics substantially better than current state-of-the-art GPUs.
• Hardware-Software Co-Design: OpenAI's hardware team, led by Richard Ho, designed the chip's architecture around its roadmap of models, kernels, and serving systems. Broadcom handled the silicon implementation and integrated its high-performance Tomahawk networking silicon, while Celestica was responsible for board, rack, and system integration.
• Current Status: Engineering samples are actively running production workloads in the lab at target frequency and power, including early testing on GPT-5.3-Codex-Spark.

Record-Breaking Development Cycle

OpenAI and Broadcom completed the co-development of Jalapeño, from initial design to manufacturing tape-out, in just nine months. The companies believe this is the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. Notably, OpenAI used its own frontier AI models to accelerate and optimize parts of the physical design process, creating a self-improving hardware engineering loop.

Deployment and Strategy

Jalapeño is the foundational step in a multi-generation compute platform that OpenAI plans to deploy at gigawatt scale with data center partners (primarily Microsoft) beginning in late 2026. By designing its own silicon, OpenAI aims to drastically lower its inference costs, increase serving speeds, and bypass supply chain bottlenecks associated with Nvidia.¹² This development follows a period of intense talent competition in custom silicon, including the high-profile defection of OpenAI's custom hardware leader, Clive Chan, to Anthropic as detailed in OpenAI Custom Silicon Leader Clive Chan Defects to Anthropic as "Titan" Chip Reaches Tape-Out.