OpenAI and Broadcom unveil the Jalapeño AI chip

Published: 2026-06-27T04:00:00.000Z

OpenAI has spent years as the most prominent buyer of someone else's chips, and it is now trying to become a maker of its own. The company has unveiled Jalapeño, its first custom processor, designed and built in partnership with Broadcom. The chip, which OpenAI calls an Intelligence Processor, was delivered in physical form to chief executive Sam Altman and president Greg Brockman this week, marking the company's first real entry into AI silicon. It is purpose built for inference, the compute heavy process of actually serving AI models to users in ChatGPT and other products, rather than the separate work of training those models in the first place. The strategic logic is about money and dependence in roughly equal measure. OpenAI, like every major AI lab, has relied heavily on Nvidia's expensive GPUs, and the phrase that keeps recurring in the industry is that nobody wants to be beholden to a single supplier. Jalapeño is an application specific integrated circuit, or ASIC, which is less flexible than a general purpose GPU but cheaper and far more efficient when tuned to one job. Bloomberg reported that the chip could cut inference costs by roughly half, and OpenAI says early testing shows performance per watt substantially better than current state of the art. Since inference is what OpenAI does millions of times a day every time someone prompts ChatGPT, even modest efficiency gains translate into enormous savings and, the company argues, cheaper and broader access to its tools. The way the chip was built is arguably as notable as the chip itself. OpenAI and Broadcom say Jalapeño went from initial design to manufacturing tape-out in just nine months, which the companies describe as possibly the fastest ASIC development cycle ever achieved in high performance semiconductors, against a typical timeline of 18 months to two years. Part of that speed came from OpenAI using its own AI models to accelerate parts of the chip design and optimization, a striking loop where the models being served help design the hardware that will run future models. The companies have already tested it running a prior generation model, GPT-5.3-Codex-Spark, on production target workloads in the lab. The deployment plan is ambitious and explicitly multi generational. OpenAI aims for initial rollout by the end of 2026 through data center partners including Microsoft, scaling from small prototype batches to gigawatt scale over the years ahead, with Broadcom's networking silicon and manufacturing partner Celestica tying the systems together. Broadcom chief executive Hock Tan described compute demand from his AI customers as simply insatiable, stretching out to 2028 and beyond. The announcement also lands as OpenAI gears up for a possible IPO that could value it near a trillion dollars, sharpening the pressure to control costs. The regional read connects directly to the Gulf's ambitions. Saudi Arabia and the UAE are pouring billions into data centers and AI compute, and they are keen to diversify beyond a single chip supplier just as OpenAI is. Custom inference silicon that lowers the cost of running models is exactly what makes sovereign backed AI infrastructure projects economically viable at scale, and the region's growing compute hubs are precisely the kind of environment where chips like Jalapeño could eventually run.