OpenAI Enters Final Stage of Second-Generation Chip Design

OpenAI is on the final leg of developing its second-generation AI semiconductor, Jalapeno, to decrease dependence on Nvidia. The chip has proven to outperform Nvidia’s premium offerings in power efficiency and response speed during performance tests. Richard Ho, OpenAI’s chief hardware engineer, stated that Jalapeno demonstrated superior performance compared to Nvidia’s "GB300" in tests measuring workload per power unit and response speed.

Ho, previously leading Google’s Tensor Processing Unit (TPU) project, described Jalapeno as a "really good chip that will significantly lower costs." Developed in tandem with Broadcom, OpenAI and Broadcom finished the chip's development at an exceptionally rapid speed, with the physical chip unveiled in June. SemiAnalysis reported that Jalapeno integrates Samsung Electronics’ HBM4 memory, providing 15.4TB/s of bandwidth per package.

The chip is designed to maintain high performance with a relatively low power consumption of 700 watts, making it an energy-efficient solution for data centers. OpenAI tested Jalapeno with various models, including DeepSeek and Moonshot AI, and found it particularly excelling with Moonshot’s "Kimi" model. Although Jalapeno is primarily for the "inference" stage of AI models, answering user queries after they have been trained, it has not yet been compared against Nvidia’s upcoming "Vera Rubin" chip.

OpenAI plans to continue collaborating with existing suppliers like Cerebras, while maintaining contracts with them. They are also utilizing AI models in the chip design process to expedite development cycles, with the second-generation chip expected to complete the "tape-out" stage within a few months, and work on a third-generation chip aimed at further reducing infrastructure costs already underway.