OpenAI hardware chief discusses Jalapeño inference chip

OpenAI said in June 2026 that it was building its own silicon with Broadcom as design partner and Celestica handling board and rack integration. In this interview, Richard Ho said the part was called Jalapeño and that it was meant for inference rather than training.

Ho said the chip carried 216 GiB of HBM4 at 15.4 TB/s, alongside a compute die and an I/O chiplet. He said peak power was 700 W and sustained draw was closer to 550 W.

At HotChips 2026, Ho presented measured results from working silicon with Ravi Narayanaswami and Chris Leary. He said the chip outperformed NVIDIA GB200 and GB300 on performance per watt, latency and some operating points where NVIDIA struggled to run.

Ho said OpenAI had tried to build one balanced part for the full inference workload rather than split the job across several specialised accelerators. He also said OpenAI had used its own models for compute design, floorplanning and kernel tuning, and that one attention implementation rose from under 1 per cent of the roofline to nearly 90 per cent in about forty hours.

The interview also covered Ho’s earlier work at D. E. Shaw Research, Google and Lightmatter, and his view that OpenAI’s hardware effort started with people before deciding on the exact idea. He said Sam Altman had shown support for more fabs and more tokens during an Asia tour that included visits to TSMC and memory suppliers.