OpenAI’s First AI‑Designed Chip Jalapeño Hits Tape‑Out in Record 9 Months

OpenAI announced, together with Broadcom, the launch of its first custom inference chip named Jalapeño, a collaboration that produced a silicon wafer in just nine months.

OpenAI President Greg Brockman highlighted that the rapid tape‑out record stems largely from the extensive use of OpenAI’s internal AI models throughout the design process, demonstrating how AI can accelerate hardware R&D.

Jalapeño is a purpose‑built ASIC created from the ground up for modern large‑language‑model (LLM) inference, rather than being a repurposed accelerator for earlier AI workloads. It targets workloads such as ChatGPT, Codex, the OpenAI API, and future autonomous‑agent products, while also being applicable to broader industry LLM use cases.

Although detailed performance numbers have not yet been released, OpenAI says early tests show Jalapeño’s performance‑per‑watt will be significantly better than the most advanced existing technologies, with a full technical report expected in the coming months.

The chip’s architecture reduces data movement and balances compute, memory, and network resources, allowing actual utilization to approach theoretical peak performance. Broadcom’s implementation and its Tomahawk networking chip enable large‑scale production, while Celestica handles system integration, motherboard, and rack design.

Rising AI compute demand has made the need for dedicated chips urgent. Brockman noted OpenAI can no longer obtain “fast enough” compute from GPUs alone, a view echoed by Broadcom CEO Tan Hock Eng, who described customer demand as “seemingly endless” and extending well beyond 2026‑2027.

OpenAI and Broadcom plan to begin deploying Jalapeño chips by the end of 2026 and to expand the deployment over the next few years, ultimately targeting a 10‑gigawatt compute cluster. Earlier in the year, OpenAI also signed agreements with Amazon (Trainium), AMD, and Cerebras to diversify its hardware portfolio.

Jalapeño is positioned as a specialized ASIC—less flexible than Nvidia GPUs but cheaper and customizable for specific AI tasks. OpenAI calls it a “smart processor” and the first AI accelerator in its platform, intended to make advanced AI faster, more reliable, and more accessible.

Reference links: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/ ; https://www.cnbc.com/2026/06/24/openai-and-broadcom-reveal-jalapeno-first-ai-chip-in-partnership.html