OpenAI pulled back the curtain Wednesday on its first custom-built artificial intelligence chip, created alongside semiconductor partner Broadcom, as the company pushes to accelerate the growth of its computing infrastructure.
AI companies like OpenAI and Anthropic are facing a serious crunch when it comes to securing enough computing power to run their most advanced chatbots and coding tools. Some organizations, including OpenAI, have begun developing their own chips as a way to cut costs and establish an alternative to the graphics processing units made by Nvidia, which have become the go-to hardware for AI workloads.
The new chip, called Jalapeño, was engineered by OpenAI’s team working alongside Broadcom. It is built specifically to handle inference — the process by which an AI system processes data to respond to a user’s question on platforms like ChatGPT.
Broadcom CEO Hock Tan told Reuters the chip performs on par with Nvidia’s Blackwell processors and the tensor processing units developed by Alphabet’s Google.
OpenAI’s hardware chief Richard Ho said the Jalapeño chip was built to work quickly and efficiently with the large language models, or LLMs, that drive many of today’s AI applications. “It will be performant on, we think, all kind of future iterations of LLMs,” Ho told Reuters.
OpenAI said it plans to roll out Jalapeño before the close of this year, describing it as the opening move in a multi-generational chip development strategy.
Canadian electronics manufacturer Celestica will be responsible for assembling the server systems that house the chips. Both the chips and the server systems will be used exclusively by OpenAI.
The company said it already has chip samples running inside its labs, operating at the intended power and performance levels with its GPT-5.3-Codex-Spark AI model.
OpenAI’s engineers spent approximately nine months completing the chip’s design before sending it to Taiwan’s TSMC for manufacturing. The company said AI tools helped speed up certain parts of that design process.
Reuters first reported back in 2023 that OpenAI was looking into building its own chip.
Other major tech players, including Meta Platforms, Amazon, and Google, have similarly turned to companies like Broadcom and Marvell for chip design services and intellectual property that are difficult to develop entirely in-house.
Anthropic, meanwhile, is also weighing whether to develop its own AI chip, according to sources who spoke with Reuters in April.
Tan noted that, for now, Broadcom’s profit margins on custom AI chips are not as strong as on some of its other products, such as networking switches. He attributed this to surging demand for the high-bandwidth memory that AI chips require. Tan said South Korea’s SK Hynix and Samsung Electronics are among the suppliers providing Broadcom with that memory.
