OpenAI and Broadcom today unveiled OpenAI\u2019s first in-house AI chip. The chip, named Jalape\u00f1o, is what\u2019s known as an Intelligence Processor\u2014in other words, an accelerator designed from the ground up for LLM inference. The launch marks the first step in a multi-year hardware platform that positions the AI giant to mount a direct challenge to the established order.<\/p>\n
OpenAI based the architecture on insights gained from the day-to-day operation of ChatGPT, Codex, and the API. Initial test results show that the chip delivers significantly higher performance per watt than the current state of the art, though a detailed technical report will not be released until the coming months. According to Broadcom CEO Hock Tan<\/a>, the chip is, in practice, \u201cjust as good\u201d as Nvidia\u2019s Blackwell GPUs<\/a> and Google\u2019s TPU data center tools. However, specific performance figures have not yet been shared.<\/p>\n Specialist versus generalist<\/p>\n Although Tan dares to compare it to Nvidia\u2019s Blackwell architecture, that claim must be viewed with the necessary nuance. The Blackwell GPUs owe their dominance to their broad versatility and the deeply entrenched CUDA software platform. As a result, they are the gold standard for both training and inference.<\/p>\n Jalape\u00f1o, on the other hand, offers pure performance gains for specialized use, but that also means OpenAI is committing to a specific type of workload. The chip is extremely efficient for current-generation models, but it offers less flexibility should the underlying AI architecture change radically in the future.<\/p>\n \u201cJalape\u00f1o is part of our long-term strategy for the entire infrastructure stack. Our goal is to make computing power more widely available, leading to faster, more reliable, and more affordable AI for both consumers and businesses,\u201d said Greg Brockman, president and co-founder of OpenAI.<\/p>\n The partnership with Broadcom and Celestica lays out the blueprint for OpenAI\u2019s hardware ambitions. Instead of investing billions in its own chip factories, the company is opting for the tried-and-true \u201chyperscaler route.\u201d This is a strategy that Google also employs in developing its TPU chips.<\/p>\n Within this trio, the hardware was developed in collaboration with Broadcom<\/a>, which provides the crucial IP for the interconnects and the Tomahawk network silicon, enabling thousands of chips to communicate efficiently with one another. The architecture minimizes data transfer and tightly aligns computing power, memory, and network capacity, so that utilization approaches the hardware\u2019s theoretical peak capacity. Manufacturing partner Celestica then acts as the ODM, integrating the chips into customer-specific server racks, ready for immediate deployment in Microsoft Azure data centers.<\/p>\n Nine months from design to production<\/p>\n According to OpenAI, it took only nine months from the initial design to a production-ready chip. This marks the fastest ASIC development cycle ever achieved in high-performance semiconductors. That pace was made possible by close collaboration between the engineering teams, supplemented by the use of OpenAI\u2019s own AI models in the design and optimization process.<\/p>\n The same models that are available to users every day have thus helped build the infrastructure for the next generation. Engineering samples of the Jalape\u00f1o chip are already running heavy ML workloads in the lab at the planned production speed and corresponding power consumption, including GPT-5.3-Codex-Spark<\/a>.<\/p>\n Gigawatt-Scale deployment and IPO Plans<\/p>\n Hock Tan has confirmed that the collaboration extends beyond a single chip: \u201cThis is just the starting point of a multi-year tech roadmap. Thanks to the direct co-development of our industry-leading chips with OpenAI, we will be able to roll out gigawatt-scale data centers starting in 2026, together with Microsoft and other partners.\u201d<\/p>\n With the launch of its own hardware, OpenAI is taking a crucial step in its full-stack strategy. The AI pioneer has long since moved beyond building only smart software models and is now also taking control of the underlying chip infrastructure. Through this shift, the company is reducing its dependence on chip giants such as Nvidia. It is also definitely emerging as a world-class tech conglomerate.<\/p>\n