OpenAI and Broadcom have announced Jalapeño, a custom “Intelligence Processor” designed from the ground up to run large language models more efficiently, marking a new phase in OpenAI’s push to build its own infrastructure stack. The first-generation accelerator is positioned as the cornerstone of a multi‑generation compute platform intended to make advanced AI faster, more reliable and more accessible.
The chip has been architected specifically around OpenAI’s view of LLM inference, drawing on the company’s experience operating models behind products such as ChatGPT, Codex and its API. OpenAI worked with Broadcom and Celestica to take Jalapeño from initial design to manufacturing tape-out in just nine months, a timeline the companies describe as the fastest ASIC development cycle they believe has been achieved in high‑performance advanced semiconductors. Engineering samples are already running machine learning workloads in the lab at production‑target frequency and power, including a frontier model referred to as GPT‑5.3‑Codex‑Spark.
Early testing indicates that Jalapeño delivers performance per watt substantially better than current state‑of‑the‑art accelerators, although detailed benchmarks will be released in a technical report in the coming months. The architecture is built to reduce data movement and balance compute, memory and networking resources so that realized utilization approaches the chip’s theoretical peak. Broadcom’s silicon implementation and networking technologies, including its Tomahawk line, are being used to scale the platform to large‑scale production
OpenAI frames Jalapeño as integral to a broader “full‑stack” strategy that spans chip architecture, kernels, memory systems, networking, scheduling, deployment and product experience. By designing more of this stack internally, the company aims to improve compute efficiency and reinvest resulting gains into training and serving more capable models. The companies see Jalapeño as the first step in a roadmap that will support gigawatt‑scale data centers with partners such as Microsoft and others beginning in 2026, with deployments expected to expand over multiple generations.
The initiative is explicitly tied to making advanced AI more broadly available by improving the cost, speed and reliability of inference, the stage at which users interact with models. OpenAI suggests that infrastructure gains from Jalapeño could translate into faster ChatGPT responses, more capable Codex tasks, cheaper API products and more dependable access during periods of high demand. The company positions these improvements as part of a longer‑term effort to democratize AI for students, developers, small businesses, enterprises and researchers.
Read Article: RBI proposes AI risk framework with kill switch mandate for banks
