OpenAI on Wednesday, June 24, unveiled its first custom AI chip developed in partnership with Broadcom, marking a significant step forward in the ChatGPT-maker’s efforts to reduce its reliance on Nvidia chips and take greater control over the infrastructure powering its AI products.
Named ‘Jalapeño’, the custom Application-Specific Integrated Circuit (ASIC) has been designed for inference workloads, which involves running large language models (LLMs) on previously unseen data to improve the overall quality of their responses.
Jalapeño is part of OpenAI’s broader efforts to build the full stack behind its models and products. It marks the Microsoft-backed startup’s foray into AI silicon, targeting the booming inference market where rivals such as Google, Amazon, and Groq already offer their own custom processors.
For Broadcom, the launch reinforces its position as the leading partner for hyperscalers, frontier labs, and other companies to create their own custom chips for AI training and inference.
“The world is moving to a compute-powered economy. Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems. By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access,” Greg Brockman, president and co-founder of OpenAI, said in a statement.
“Our collaboration with OpenAI represents a fundamental commitment to scaling the physical infrastructure required for the next decade of AI. This is just the beginning of a multi-generation roadmap. By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026,” Hock Tan, president and CEO, Broadcom, said.
What do we know about Jalapeño?
Nvidia’s graphics processing units (GPUs) have dominated the process of AI model training in recent years. Major hyperscalers such as Google, Microsoft, Amazon, Meta, and Oracle have come to rely heavily on Nvidia’s GPUs to train and develop their own AI models as well as rent them out to AI startups like OpenAI and Anthropic.
However, these companies have also been looking to develop in-house, custom-built AI chips to reduce their dependence on chipmakers and cut costs — at least in the long run since custom chips have a steep upfront cost that can run into tens of millions of dollars.
GPUs are said to work well for training AI models, but as the technology matures and competition increases, companies are looking to optimise AI model inferencing.
This is where OpenAI’s Jalapeño comes in. It falls in the category of custom application-specific integrated circuits (ASICs) designed for specific AI tasks. Other custom ASICs include Google’s TPUs, Groq’s LPUs, Amazon’s Trainium, Microsoft’s Maia 100, and Tesla’s AI5 chips. Even Nvidia has its own AI inference chip developed by licensing Groq’s technology.
OpenAI and Broadcom have described Jalapeño as an “intelligence processor” that is part of a larger “AI accelerator” platform that they are currently building “to make advanced AI faster, more reliable, and more accessible to more people.”
While Jalapeño is expected to be deployed in data centres owned or leased by OpenAI, the company claimed that the chip is flexible enough to work with all LLMs. Its architecture reduces data movement and balances compute, memory, and networking resources to achieve realised utilization much closer to theoretical peak performance, as per the company.
It has been designed for large-scale production thanks to Broadcom’s silicon implementation and networking technologies, including the Tomahawk networking silicon.
How was it designed?
OpenAI said that the chip was designed from end to end in nine months. The company leveraged its deep understanding of LLM fundamentals as well as its own roadmap of models, kernels, serving systems, and other product needs in order to design the chip from the ground up.
“We optimised the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models. Based on early testing, Jalapeño will efficiently execute our most important workloads close to the hardware’s theoretical limits,” Richard Ho, who leads OpenAI’s hardware programme, said.
The company also used its own AI models in the design process, following which Broadcom and Celestica helped industrialise the platform through chip implementation, board, rack system integration, high-performance networking, and scalable production systems.
In terms of performance, OpenAI said it is still measuring the final figures. However, early testing of Jalapeño shows that it is capable of delivering performance per watt substantially better than current state-of-the-art. The company is expected to publish a detailed technical report on the chip’s performance in the coming months.
What is OpenAI’s chip strategy?
Since the launch of ChatGPT in 2022 that kick-started the generative AI boom, OpenAI has been one of the biggest customers for Nvidia’s GPUs which are still considered to be a key piece of infrastructure for building LLMs and running AI workloads.
As demand for AI tools and services has soared, OpenAI and competitors such as Anthropic have been forced to secure alternative sources of advanced silicon to keep pace with their computing needs.
Earlier this year, OpenAI inked a deal with Amazon Web Services (AWS) that includes use of the company’s Trainium AI chips. OpenAI has also signed similar agreements with Nvidia rival Advanced Micro Devices (AMD) and with AI chipmaker Cerebras.
In October 2025, OpenAI and Broadcom announced plans to jointly develop and deploy racks of custom chips with the goal of expanding compute capacity to 10 gigawatts of power. However, the multibillion-dollar partnership is structured so that Broadcom funds production of Jalapeño only once Microsoft or another buyer commits to purchasing 40 per cent of the chips, according to a report by The Information.
To build a chip rivaling Google’s TPUs, OpenAI recruited several former TPU engineers from Google, including Richard Ho, who led the effort as senior director of engineering. Broadcom was also heavily involved in the development of Google’s TPUs.
Moving forward, OpenAI-designed chips could give the company full-stack advantage. It could also serve as a positive signal to investors ahead of the company’s planned IPO sometime next year. “OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience,” the company said.
While engineering samples of the Jalapeño chip are currently running machine-learning workloads in the OpenAI lab at production target frequency and power, including GPT‑5.3‑Codex‑Spark, the companies said that they are aiming for initial deployment of the chips by the end of 2026.
OpenAI's Sam Altman (left) and Broadcom CEO Hock Tan pose with the new Jalapeño AI inference chip. (Image: OpenAI) 
Share
