OpenAI Built an Nvidia-Beating Inference Chip in Nine Months. Broadcom (AVGO) Helped Make It Happen

OpenAI has developed its own inference chip, dubbed Jalapeño, which can perform up to 1.9 times more AI tasks per watt than competing systems. The nine-month journey from design to manufacturing was made possible by Broadcom (NASDAQ:AVGO), which contributed the physical implementation and networking. Celestica handled the board, rack, and system integration, while Taiwan Semiconductor Manufacturing Company (TSMC) manufactured the chip.

Despite its impressive specifications, OpenAI notes that Jalapeño is not a direct replacement for Nvidia's GPUs and mainly focuses on inference tasks rather than training accelerators. The manufacturing cost remains unclear as the system has only been deployed for testing. Broadcom's potential expands with the planned deployment of 10 gigawatts of OpenAI-designed accelerators by 2029, but the company's AI semiconductor revenue grew 143% in the latest quarter.

Meanwhile, Nvidia remains a dominant player in the AI space, with a higher hedge-fund count and a larger proportion of short interest in its stock.