OpenAI announced on June 24 that it has completed the design of its first custom AI inference chip, Jalapeño, developed jointly with Broadcom [1, 2, 3]. The chip targets inference workloads that power services like ChatGPT responses, rather than AI model training [1, 2, 3].

Jalapeño aims to cut AI inference costs by roughly 50% compared to traditional GPUs, a significant step in reducing expenses for large-scale AI operations [1, 4, 3, 5]. OpenAI Hardware Lead Richard Ho said the chip "runs key AI workloads efficiently and its performance approaches the theoretical hardware limits" [2].

Testing shows Jalapeño delivers power efficiency and performance surpassing current leading AI chips on the market [1, 2, 3]. Broadcom CEO Hock Tan stated, "Jalapeño's performance is comparable to Nvidia's Blackwell chip and Google's TPUs, and this is just the beginning of a long-term collaboration" [3]. The chip's large compute die size of about 840 mm² pushes the limits of EUV lithography technology [6].

The design and tape-out process took an unusually fast nine months in the semiconductor industry, with OpenAI leveraging its own AI models to accelerate optimization [2, 6, 7, 8]. Jalapeño will be manufactured by Taiwan Semiconductor Manufacturing Company (TSMC) using advanced 3nm process technology [7, 8, 9].

Initial deployment will start by the end of 2026 in Microsoft and other OpenAI partner data centers, with plans to scale up in 2027 and full production by mid-2028 [1, 3, 10, 7, 8]. Canadian EMS company Celestica will assemble server systems, with potential future involvement from Taiwanese server ODMs Foxconn and Quanta [7, 9].

OpenAI aims for AI infrastructure consuming up to 10 gigawatts of power using custom chips like Jalapeño [2, 9]. Despite Jalapeño, OpenAI will continue to use Nvidia, AMD, Cerebras, and AWS Trainium chips to meet demand [1, 3, 10, 5].

OpenAI President Greg Brockman said Jalapeño is "part of our long-term strategy to build full-stack infrastructure to provide advanced AI services efficiently" [8]. The partnership with Broadcom targets multi-gigawatt scale deployments in data centers from 2026 onward [1, 2, 6, 5].

Jalapeño’s production is expected to boost Taiwan’s semiconductor supply chain, benefitting TSMC and server ODMs [7, 9].