In a major bid to reduce reliance on NVIDIA Corp. and secure its own hardware pipeline, OpenAI has partnered with Broadcom Inc. to unveil its first custom artificial intelligence (AI) chip, codenamed Jalapeño.
The chip marks the ChatGPT maker’s first official entry into custom AI hardware. Designed as an application-specific integrated circuit (ASIC), Jalapeño is optimized specifically for inference, the compute-heavy process of running live AI models to answer user queries. Following Wednesday’s announcement, Broadcom shares rose about 2%.
While ASICs are typically less flexible than NVIDIA’s dominant graphics processing units (GPUs), they are cheaper to operate and can be tailored for highly specific tasks.
OpenAI executives and industry partners claim the new processor punches well above its weight. Broadcom CEO Hock Tan said early testing shows Jalapeño performs on par with NVIDIA’s high-end Blackwell chips and Alphabet’s Google Tensor Processing Units (TPUs).
OpenAI confirmed that physical samples are already running in its labs, hitting target power and performance metrics while processing the company’s advanced GPT-5.3-Codex-Spark AI model.
Additionally, OpenAI hardware chief Richard Ho noted the chip is engineered to remain performant across “all kinds of future iterations” of large language models.
“By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access,” OpenAI President Greg Brockman said in a statement.
The creation of Jalapeño highlights a remarkably fast turnaround for a custom processor. OpenAI engineers completed the chip design in just nine months, accelerating the process by utilizing AI design tools.
The manufacturing and deployment ecosystem relies on several tech heavyweights. The chips are fabricated by Taiwan Semiconductor Manufacturing Co. (TSMC). Canadian electronics manufacturer Celestica will build the proprietary server systems housing the chips. South Korea’s SK Hynix and Samsung Electronics are supplying the critical high-bandwidth memory (HBM) required for the processors.
Tan noted that the intense global demand for HBM is currently tightening Broadcom’s profit margins on custom silicon compared to traditional networking hardware, though the strategic value remains immense.
OpenAI’s shift toward in-house hardware reflects a broader industry trend among tech giants scrambling to bypass NVIDIA’s supply constraints and high prices. Meta Platforms Inc., Amazon.com Inc., Google, and Microsoft Corp. have all deployed proprietary AI processors, while rivals like Anthropic are reportedly weighing similar moves.
“All the major hyperscalers are doing their own optimized AI chips, some with Broadcom as well. OpenAI wants to have an optimized chip architecture that can best run their models, as well as have an option to reduce costs of compute by not having to pay the steep margins that NVIDIA (and others) make on their commodity chips,” tech analyst Jack Gold said. “This is right out of the hyperscaler’s playbook. By having their own custom chips running optimized architecture, they achieve lower cost of operations and can pass the savings on to users. This last point is increasingly important as the cost per token is a key metric for AI use these days by many companies, and especially as we move to agenticAI where the token use skyrockets.”
OpenAI remains a massive buyer of NVIDIA hardware, but it has steadily diversified its infrastructure portfolio. The startup has already secured partnerships to use Amazon’s Trainium chips, Advanced Micro Devices (AMD) hardware, and Cerebras Systems Inc.
Physical samples of the Jalapeño processor were delivered to OpenAI this week. The companies plan to begin initial rollout and deployment in data center racks by the end of 2026.
