OpenAI is currently utilizing its custom Jalapeno AI inference ASIC exclusively for internal compute needs. The company reports that internal demand for compute is high and growing, which will occupy its production capacity for the foreseeable future.
The Jalapeno chip is designed as a programmable, general-purpose ASIC rather than a hard-coded solution for OpenAI models. During the Hot Chips event, the company demonstrated the chip accelerating its own GPT-OSS model, as well as DeepSeek R1 and Kimi K2.5.
Engineers achieved compatibility with these models within two months between the A0 sample stage and the Hot Chips presentation. OpenAI emphasizes that the primary design goal for the ASIC was efficiency to address power constraints in modern AI data centers.
OpenAI opted to develop custom silicon to facilitate codesign with its internal models, citing concerns over potential intellectual property leaks when working with third-party silicon merchants. The company noted that it cannot share sensitive research widely.
While OpenAI currently focuses on internal deployment, it has not ruled out a broader market rollout. The company compared Jalapeno performance against Nvidia GB200 and GB300 platforms at Hot Chips and has conducted internal testing against the newer Vera Rubin platform.
