OpenAI partners with Broadcom to launch the first AI chip "Jalapeño," quickly arriving in 9 months to challenge Nvidia's dominance

OpenAI and Broadcom officially unveiled their first custom AI processor, "Jalapeño," on the 24th.
The chip is designed specifically for large language model (LLM) inference, taking only 9 months from design to mass production, with deployment expected to begin by the end of 2026.
This marks a key step for OpenAI toward a "full-stack platform," aiming to improve performance per watt and reduce over-reliance on NVIDIA.
(Background: OpenAI Codex was reported to be killing your SSD: writing 37 TB in 21 days, burning out drives in less than a year)
(Additional context: OpenAI announced the "Earth Repair" project, providing security support for 19 well-known open-source projects including cURL, Python, and PyPI)

AI giant OpenAI is officially entering the hardware field, targeting chip independence.
On June 24, 2026, OpenAI and semiconductor giant Broadcom jointly announced in San Francisco and Palo Alto their first intelligence processor called "Jalapeño."
This AI accelerator, designed specifically for future LLM inference, signifies OpenAI’s expansion from products and models down to the underlying chips, marking an important milestone in building a "full-stack" infrastructure.

At the launch event, Broadcom President and CEO Hock Tan personally handed the Jalapeño sample to OpenAI CEO Sam Altman and President Greg Brockman, officially kicking off the hardware ecosystem collaboration.

Rapid development in 9 months, focused on LLM inference performance

Jalapeño is a specialized application-specific integrated circuit (ASIC), built from scratch and tailored for modern LLM inference needs, rather than a modified general-purpose accelerator.
Thanks to the collaborative hardware and software development by OpenAI’s engineering team, Broadcom’s silicon expertise, and acceleration support from OpenAI’s own AI models, the chip went from design to tape-out in a record-breaking 9 months.

Currently, engineering samples have successfully run ML workloads—including GPT-5.3-Codex-Spark—at target frequency and power in the lab.
Early testing data shows that its "performance per watt" significantly surpasses the most advanced levels in the current market.
OpenAI hardware project lead Richard Ho emphasized that the team deeply optimized critical aspects such as memory movement, networking, and service models for cutting-edge AI models, making actual utilization closer to theoretical peaks.

Reducing dependence on NVIDIA, gigawatt-scale deployment by late 2026

Since the surge of generative AI in 2022, OpenAI has been one of NVIDIA’s largest GPU buyers.
However, as computational demands explode, diversifying chip sources has become urgent.
In addition to launching its own chip Jalapeño, OpenAI has actively partnered with hardware vendors like Amazon AWS (using Trainium chips), AMD, and Cerebras to reduce reliance on a single supplier and control costs.

Broadcom, one of the biggest beneficiaries of the generative AI boom, handled the manufacturing and network integration of Jalapeño (including Tomahawk network silicon), while Celestica was responsible for board and rack system integration.
The chip is expected to begin deployment by the end of 2026, with future collaborations with data center partners like Microsoft to expand across multiple generations at gigawatt-scale.

OpenAI President Greg Brockman stated that the world is moving toward a "compute-driven economy," and by designing more of its own underlying stack, OpenAI can deliver smarter services more efficiently.
As major tech giants compete in developing their own chips, the AI compute market is undergoing an unprecedented and intense reshuffle.