Enterprise

OpenAI and Broadcom unveil Jalapeño, a custom inference chip built in nine months

The ASIC accelerator, headed for gigawatt-scale deployment with Microsoft by year-end, is showing roughly 50% cost savings versus typical AI GPUs, Broadcom's CEO says.

By Asher Glick · Enterprise software desk · 2026-06-28 · 3 min read

Photo: Unsplash / Alexandre Debiève — Close-up of a printed circuit board with a central processor chip

OpenAI and Broadcom on Tuesday unveiled Jalapeño, OpenAI’s first in-house inference chip, an ASIC the companies say went from initial design to manufacturing tape-out in nine months. Broadcom CEO Hock Tan told Bloomberg the part delivers roughly 50% cost savings against typical AI GPUs, the clearest signal yet that the largest model lab in the world intends to wean itself off Nvidia silicon on a schedule of its own choosing.

The chip is purpose-built for serving ChatGPT, Codex, and the OpenAI API. Industry experts told CNBC that ASICs are less flexible than Nvidia’s general-purpose GPUs but cheaper and tunable to specific workloads, the classic trade once a buyer’s demand curve is steep enough and predictable enough to justify the engineering bill. OpenAI’s clearly is.

Richard Ho, who leads OpenAI’s hardware program, said the design was shaped “around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models.” Engineering samples are already running ML workloads in the lab at production target frequency and power, including a model the companies identified as GPT-5.3-Codex-Spark. A full technical report is expected in the coming months.

The nine-month cycle, which OpenAI and Broadcom suggest may be the fastest ever achieved in high-performance semiconductors, was accelerated in part by using OpenAI’s own models to assist the design and optimization work. The recursion is the story: a frontier lab using its frontier models to build the silicon that’ll serve the next generation of them.

Tan laid out the roadmap bluntly. “small prototype development” arrives in late 2026 before the program ramps in 2027 and runs “full tilt in first half ‘28,” he told CNBC, “enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026.” He described compute demand from Broadcom’s six unnamed customers as “simply insatiable.” OpenAI president Greg Brockman, speaking to CNBC’s David Faber, was simpler: the company “cannot get compute fast enough.”

The market has been pricing this in. Broadcom shares are up roughly 10% so far in 2026 and nearly sevenfold since the end of 2022. The partnership itself was officially announced last October. What changes today is the artifact: an actual chip, with a name, running a real model, on a schedule that ends in gigawatts. The dependency on Nvidia has a visible expiration date for the first time.

Sources