OpenAI has built its first custom AI chip, called Jalapeño, and it could make ChatGPT faster and cheaper to run for users everywhere, including Nigeria. The chip was developed with Broadcom and Celestica, marking OpenAI’s first real step into building its own hardware instead of depending fully on companies like Nvidia.
OpenAI’s president, Greg Brockman, personally received an early model of the chip from Broadcom’s CEO, Hock Tan. The handoff was a symbolic moment showing how serious OpenAI is about controlling more of the technology that powers its AI models.
Running AI models like ChatGPT costs a huge amount of money and computing power. Most AI companies rent or buy chips from Nvidia, but demand for those chips is so high that companies cannot get enough, fast enough. Brockman explained that OpenAI “cannot get compute fast enough” to meet the demand for its AI products. Broadcom’s CEO backed this up, saying demand from Broadcom’s AI customers is far more than the company can currently supply, and that this pressure is expected to continue for years.
By designing a chip built specifically for the way ChatGPT and similar tools work, OpenAI hopes to cut costs and increase how many people it can serve at once. Brockman said the goal is to make AI “faster, more reliable, more affordable for people and businesses.”
Unlike a general-purpose chip, Jalapeño was built only for one job: running AI models after they have already been trained, a process called inference. This is the exact process that powers your ChatGPT replies every time you type a question.
Don't MIss This Article
OpenAI’s head of hardware, Richard Ho, said the chip was shaped around the specific patterns that matter most for running large language models, like memory movement and how data flows through the system. Early tests show the chip performs much better per unit of power compared to current top chips on the market, though OpenAI has not released full performance numbers yet.
Notably, Broadcom’s CEO told Reuters that the new chip performs at a level close to Nvidia’s top Blackwell chips and Google’s own AI chips, two of the most advanced processors available today. One of the most surprising details is how fast this chip came together. Designing a chip like this normally takes one and a half to two years. OpenAI says Jalapeño went from design to manufacturing in just nine months, something the company calls one of the fastest development cycles ever for this kind of advanced chip.
Part of that speed came from OpenAI using its own AI models to help design and improve parts of the chip itself. In simple terms, the AI helped build the hardware that will run future versions of itself. Broadcom is handling the chip’s core silicon work and networking technology, while Canada based company Celestica is helping put together the boards and systems that will house the chips in data centres.
Don't MIss This Article
For Nigerian users, the biggest impact would likely show up gradually rather than overnight. As AI use grows across schools, startups, and businesses in Nigeria, cheaper and faster AI infrastructure could mean quicker ChatGPT responses, fewer slowdowns during busy periods, and lower costs for businesses building tools on top of OpenAI’s systems. It is worth noting that OpenAI will still rely on Nvidia and AMD chips for many of its workloads, especially the heavier task of training new models from scratch. Jalapeño is meant to handle the lighter, repeated task of serving answers to users, not building new AI models.
OpenAI and Broadcom are aiming for the first real deployment of Jalapeño by the end of 2026, with bigger rollouts expected in the years after that. Early use will likely be small scale at first before expanding to larger data centres. OpenAI joins Google, Amazon, and Microsoft, who have all built their own AI chips already, as the AI industry shifts away from depending only on Nvidia. Whether Jalapeño actually delivers on its promises will only become clear once independent testing and real-world use begin later this year.
