OpenAI and Broadcom unveil custom silicon to tackle AI compute crunch
The partnership marks a strategic shift for OpenAI to reduce reliance on external chip suppliers, with deployment targeted for late 2026.
OpenAI and Broadcom have announced the Jalapeño, a custom application-specific integrated circuit (ASIC) designed specifically for large language model inference within data centres. Developed over a nine-month period, the chip was built from the ground up rather than repurposing existing architectures, drawing on detailed insights from OpenAI’s researchers and its roadmap for future models. Both companies state that the silicon will be deployed in data centres by the end of 2026, marking the first generation in what they describe as a long-term project to refine the technology.
The announcement comes amid intensifying competition for artificial intelligence compute capacity, often referred to as a global compute crunch. As hyperscalers and frontier model teams scramble for limited data centre space, custom silicon has emerged as a critical method for squeezing out additional efficiency and capacity. Broadcom, an established silicon supplier, has significantly expanded its business model to provide custom chips to these entities during the current AI boom, leveraging its position to support infrastructure build-outs.
OpenAI claims that early testing indicates the Jalapeño will deliver performance per watt substantially better than current state-of-the-art hardware. However, the company noted that performance measurement is not yet complete and that a detailed technical report will be presented in the coming months. Until those specifications are made public, the precise metrics of the chip’s capabilities remain unverified by independent, large-scale testing.
This development represents a strategic move by OpenAI to own the full stack behind its models and products, including its ChatGPT and Codex services. By developing its own silicon, the company aims to reduce its dependence on external suppliers such as Nvidia. The push for vertical integration is ostensibly intended to provide better performance and efficiency while mitigating the risks associated with reliance on third-party hardware vendors.
While the end-of-2026 deployment target has been set by both companies, the actual timeline and scale of the rollout remain subject to change. The industry continues to watch closely as the race for custom AI infrastructure accelerates, with the Jalapeño serving as a tangible example of the shift toward specialised hardware designed to meet the escalating demands of large language models.
