Broadcom and OpenAI Unveil Jalapeño, a Custom Inference ASIC Slated for 2026

Broadcom and OpenAI revealed Jalapeño, a custom reticle-sized inference ASIC designed to optimize LLM data movement and latency, scheduled for gigawatt-scale deployment in 2026.

Chips & Climate Desk
Updated 5d ago 2 min read
Broadcom and OpenAI Unveil Jalapeño, a Custom Inference ASIC Slated for 2026

OpenAI and Broadcom have unveiled Jalapeño, a custom-designed inference ASIC tailored for modern large language models and agentic AI workloads. According to Tom's Hardware, the chip represents OpenAI's initial foray into proprietary silicon, skipping repurposed training architectures to address inference-specific bottlenecks such as data movement, memory bandwidth, and networking efficiency. Jalapeño integrates a single large compute chiplet with high-bandwidth memory (HBM) rather than lower-cost DRAM, aiming to optimize throughput and latency for complex reasoning and multi-step agentic execution.

The development cycle for Jalapeño achieved tape-out in nine months, a notable acceleration from the standard 1.5 to 2-year timeline for custom ASIC development. Broadcom and OpenAI credited this speed to logic reuse across Broadcom’s design catalog and the application of OpenAI’s own machine learning models to streamline architecture optimization. Engineering samples are currently running in lab environments at target clock speeds and power envelopes, executing models including GPT-5.3-Codex-Spark.

Physical analysis of package photos reported by Tom's Hardware reveals a reticle-sized compute die measuring approximately 25.46 mm by 33 mm, yielding an estimated area of 840 mm²—approaching the 858 mm² exposure limit of standard EUV lithography tools. The package incorporates six HBM stacks surrounding the primary compute chiplet, along with an input/output interface chiplet flanked by two structural dummy dies. The silicon floorplan displays a highly repetitive, columnar tiled layout consistent with systolic-array inference architectures.

While Broadcom and OpenAI have not released formal benchmark figures, early internal testing indicates the ASIC delivers performance-per-watt superior to existing hardware like Nvidia’s Blackwell architecture and AMD’s Instinct MI350 series. However, its competitive standing against future architectures, including Nvidia’s Rubin and AMD’s Instinct MI400, remains unestablished. According to Richard Ho, leader of OpenAI’s hardware program, the design targets kernel execution and serving patterns specific to frontier models to operate near theoretical hardware limits.

Deployment of Jalapeño is scheduled to begin in late 2026. Broadcom President and CEO Hock Tan stated that the custom processor will power gigawatt-scale data center infrastructure in collaboration with Microsoft and other cloud partners. Furthermore, the architecture is engineered to support third-party language models beyond OpenAI’s proprietary suite, opening possibilities for broader enterprise deployment provided sufficient foundry capacity is secured through Broadcom and manufacturing partner TSMC.

Sources

  1. 01 Broadcom and OpenAI unveil custom-built Jalapeño inference processor — OpenAI's first chip is a massive reticle-sized ASIC built in an ultra-fast nine-month development cycle — Tom's Hardware
  2. 02 AI chipmaker Groq confirms $650M raise, re-staffs after Nvidia’s $20B not-acqui-hire deal — TechCrunch — Hardware