OpenAI in-house Jalapeno inference chip reportedly surpasses NVIDIA GB300 in tests
On August 25-26, OpenAI revealed Jalapeno, co-developed with Broadcom, outperforms NVIDIA GB300 in power efficiency and response speed.
On August 25-26, OpenAI revealed that Jalapeno, an inference chip co-developed with Broadcom, outperformed NVIDIA GB300 in two metrics: per-watt AI workload and response speed. Jalapeno is OpenAI's key in-house silicon project and could enter inference service later this year.
OpenAI stated it still considers NVIDIA a key supplier. The custom chip effort is aimed at reducing single-supplier dependency and tailoring silicon to OpenAI's own workloads. Combined with its Broadcom partnership, OpenAI is further diversifying its compute foundation.
The same reporting disclosed that OpenAI data center chief Chris Malone left the company on August 26, the latest in a string of recent executive departures. Malone joined OpenAI in March 2025 from Meta and Google and participated in the company's roughly $600B compute investment plan for 2030.