OpenAI and Broadcom unveil "Jalapeño," a custom chip built for LLM inference

On June 24, 2026, OpenAI and Broadcom jointly announced the launch of their first self-developed chip, "Jalapeño," officially called the "Intelligence Processor." The chip is designed specifically for large language model inference, not modified from a general-purpose chip, but built from scratch optimized for modern LLM inference. Broadcom CEO Hock Tan and President Charlie Kawwas delivered the first wafer to OpenAI CEO Sam Altman and President Greg Brockman.

In the collaboration, OpenAI handles chip design, Broadcom provides silicon manufacturing and networking technology (including Tomahawk networking chips), and Celestica handles boards, racks, and system integration. Jalapeño is the first chip in a multi-generational platform plan between the two companies.

OpenAI states that early tests show Jalapeño's performance-per-watt is "significantly better" than current state-of-the-art hardware. The architecture reduces data movement, enabling utilization close to theoretical maximums. However, these figures are self-reported by OpenAI, not yet finalized, and no comparison chip models, test tasks, or conditions have been disclosed. A technical report will follow, so independent verification is currently impossible.

Engineering samples are already running ML workloads in the lab, including the GPT-5.3-Codex-Spark model. Notably, this model previously ran on Cerebras hardware—a company also focused on inference acceleration. This suggests Jalapeño may be competitive on certain inference tasks, but specific comparison results have not been made public.

OpenAI claims the design-to-tapeout cycle took only 9 months, the fastest known ASIC development cycle. Part of the design process was accelerated by OpenAI's own AI models. However, rumors about a chip plan have circulated since 2023, suggesting actual R&D may have started earlier. It is unclear whether the 9 months include the preliminary proof-of-concept phase.

Broadcom CEO Hock Tan stated that the first deployment is planned for late 2026 at gigawatt scale, in collaboration with Microsoft and other partners. According to reports, Broadcom required Microsoft to commit to purchasing 40% of the chips to secure first-phase capacity. This arrangement highlights Microsoft's strategic position as OpenAI's primary partner and investor, while also reflecting the high risk of custom chip mass production—requiring large customers to lock in orders in advance.

Jalapeño marks OpenAI's first expansion from a model and product company into hardware, with the core logic being: by controlling the full stack from chip to product, OpenAI can achieve faster model operation, higher reliability, and lower costs. This helps reduce dependence on external chip suppliers like Nvidia, especially amid AI chip shortages.

However, challenges are significant: performance data is unverified, mass deployment yield and actual performance are unknown; custom chip R&D and maintenance costs are high; and Microsoft as a major customer may affect OpenAI's independence. Additionally, competitors like Nvidia have deep expertise in AI chips, and whether Jalapeño can truly compete in ecosystem and performance remains to be seen.