OpenAI’s Jalapeño chip aims to cut LLM serving costs

OpenAI and Broadcom developed a custom application-specific chip called Jalapeño to reduce the cost of running large language models. TSMC will manufacture the silicon in Taiwan and Celestica will assemble the boards and rack systems. OpenAI reports early lab samples running frontier workloads, including an unreleased GPT-5.3-Codex-Spark model, at target production frequency and power.

Jalapeño is built specifically for inference. The design reduces data movement between compute, memory and networking to increase realized utilization. The processors integrate Broadcom’s Tomahawk networking silicon to speed communication across large clustered data centers. OpenAI officials describe the processor as optimized for interactive serving rather than repurposed from older accelerator designs.

OpenAI spent $8.4 billion on ChatGPT servers last year and projects about $14 billion in server costs this year. The company reports about $25 billion in annual revenue, a platform with roughly 900 million weekly users, and a commitment of about $1.4 trillion to computing power over the next eight years. High-end processors from Nvidia are estimated to carry roughly a 75% profit margin, while OpenAI retains about $0.33 of profit on each dollar after operating expenses.

OpenAI moved the Jalapeño design from a blank-slate concept to manufacturing tape-out in nine months. Engineers used OpenAI’s language models to automate and optimize portions of the hardware design process. Broadcom will scale the rollout with infrastructure partners, including Microsoft, to prepare for gigawatt-scale data center integration, Broadcom CEO Hock Tan confirmed.

Initial deployment of the hardware into data centers is scheduled to begin by the end of 2026. OpenAI positions Jalapeño as an “intelligence processor” intended to reduce capital and operating expenses associated with third-party accelerators and to better match hardware characteristics to future model designs. Greg Brockman, OpenAI’s president and co-founder, described Jalapeño as “part of our long-term full-stack infrastructure strategy to make compute more abundant.”

OpenAI controls chip architecture, software kernels, memory systems and network scheduling alongside the application layer. The company describes aligning those layers as a way to adjust hardware and software together to match model requirements.