Enterprise

OpenAI and Broadcom unveil 'Jalapeño,' the company's first custom inference chip

OpenAI's first Intelligence Processor, co-developed with Broadcom and Celestica, targets gigawatt-scale deployment with Microsoft by the end of 2026 — and a sharp break from Nvidia dependence.

By Asher Glick · Enterprise software desk · 2026-06-24 · 3 min read

Photo: Unsplash / Igor Omilaev — Close-up render of an AI processor on a circuit board

OpenAI and Broadcom unveiled Jalapeño on June 24, an inference-optimized accelerator that OpenAI is calling its first Intelligence Processor and the opening piece of what the two companies describe as “the first AI accelerator in a multi-generation compute platform the companies are building together.” The reveal was staged across San Francisco and Palo Alto, with Broadcom’s Hock Tan and Charlie Kawwas handing a finished part to Sam Altman and Greg Brockman in front of cameras. The choreography is the message.

Jalapeño was designed inside OpenAI, with Broadcom and Celestica handling chip implementation, board and rack integration, and the high-performance networking that ties the racks together using Broadcom’s Tomahawk switch silicon. Richard Ho, who leads OpenAI’s hardware program, says the team optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier models. Engineering samples are already running ML workloads at production target frequency and power, including GPT-5.3-Codex-Spark.

The numbers are where the strategy becomes legible. Tan told Bloomberg that early testing shows roughly 50% cost savings versus typical AI GPUs, and OpenAI claims performance per watt substantially better than current state-of-the-art, with a detailed technical report promised in the coming months. Initial deployment is targeted for the end of 2026 at gigawatt scale alongside Microsoft. By 2029, OpenAI wants custom silicon powering 10 gigawatts of compute.

Until now, OpenAI has run almost exclusively on Nvidia, with some recent inference workloads moving to Cerebras. Jalapeño doesn’t end that dependence; pre-training stays on Nvidia hardware for the foreseeable future. But inference is where the bill compounds daily, and even modest unit-cost reductions there reshape gross margin at OpenAI’s scale.

Tan put the logic in unusually plain language: “at the end of the day, you cannot, should not rely on some other third-party GPU to do it for you, because it’s such a key part.” It’s the same argument Apple made about its CPUs in 2020 and Amazon made about Graviton: at sufficient scale, the most expensive input in your business can’t belong to someone else’s roadmap. OpenAI is now operating at that scale, and acting like it.

Sources