OpenAI's Broadcom-Built Jalapeño Chip Beats Nvidia GB300 in Inference Tests as Cisco and Nvidia Expand AI Factory

OpenAI's custom Jalapeño chip, designed and built by Broadcom specifically for AI inference, outperformed Nvidia's GB300 processor in both throughput per watt and response latency in OpenAI's own testing, with the performance gap widening on larger workloads including Moonshot's Kimi model and unreleased OpenAI models. Running at around 700 watts, Jalapeño is not designed for model training. OpenAI plans to begin deploying the chip for its own models later in 2026, and a second-generation version is approaching tape-out while a third generation is already in development.

A key caveat is that testing was conducted against the GB300, not Nvidia's newer Vera Rubin generation, meaning the competitive picture could shift as Nvidia's next-generation hardware reaches customers. Separately, Cisco announced it is expanding its Secure AI Factory with Nvidia to include high-density liquid- and air-cooled compute solutions from Supermicro, targeting enterprises, neoclouds, and sovereign clouds. Those solutions will support Nvidia's Vera Rubin NVL72 and HGX Rubin NVL8 platforms, with availability set for October 2026.

Nvidia also introduced the Jetson Orin Nano 2, an entry-level edge AI computer delivering 78 TOPS of AI compute with 8GB of memory and twice the inference performance of the Orin Nano Super at 40 percent lower power consumption. Early adopters include Cognex, Doosan Bobcat, and Matic, while Alphabet's Wing plans to evaluate it for delivery drone perception; the module and developer kit are expected in the first half of 2027.