Broadcom, OpenAI debut Jalapeño LLM inference chip for hyperscale AI

Broadcom and OpenAI have jointly introduced Jalapeño, a custom large language model inference processor that marks OpenAI’s first dedicated AI chip and the first accelerator in a multi‑generation compute platform the two companies plan to deploy at hyperscale. OpenAI led the architecture and design of the chip around the specific behavior of its LLM workloads, while Broadcom is responsible for silicon implementation, manufacturing, advanced packaging, and networking integration, with Celestica providing board, rack, and system expertise. Engineering samples of Jalapeño are already running machine learning workloads such as GPT‑5.3‑Codex‑Spark at production‑target frequency and power in lab environments, and early testing indicates performance per watt substantially better than current state‑of‑the‑art AI accelerators.

Jalapeño is a purpose‑built ASIC optimized specifically for inference rather than training, designed from a blank slate to handle modern LLM kernels, memory movement, networking, and serving patterns at massive scale. The chip is intended to support OpenAI’s own services, including ChatGPT, coding assistants, and API workloads, but is architected with flexibility to run current and future LLMs across the industry. OpenAI designed Jalapeño using AI‑assisted tools and completed the design‑to‑tape‑out cycle in around nine months, an unusually fast timeline for a high‑performance ASIC.

The initial target market for Jalapeño is hyperscale data centers running LLM inference, with deployment planned alongside partners such as Microsoft at gigawatt scale over multiple generations. Broadcom has indicated that small prototype deployments will begin late 2026, followed by a significant ramp in 2027 and full‑scale production in the first half of 2028, and internal estimates suggest up to roughly 50% cost savings versus typical GPU‑based inference for some workloads. The chips are expected to be used primarily in OpenAI‑aligned infrastructure rather than sold broadly to third parties at launch, though the design’s ability to support a wide range of LLMs leaves open the possibility of wider availability later.

Jalapeño will be integrated into a broader platform that combines OpenAI‑designed accelerators with Broadcom’s Tomahawk networking silicon and Celestica’s system integration know‑how to build racks and data center systems optimized for AI inference at gigawatt scale. Initial deployment of the Jalapeño‑based platform is targeted by the end of 2026, with expansions across subsequent generations, meaning that for now AI customers will continue relying on existing GPU and accelerator vendors as Jalapeño progresses through validation, yield ramp, and system integration phases. This move represents a significant step in OpenAI’s vertical integration of its AI stack and positions Broadcom as a key silicon and networking partner for next‑generation LLM infrastructure.