OpenAI unveiled detailed benchmark results for its Jalapeño inference ASIC at the Hot Chips conference, showing it delivers more tokens per user and higher throughput per kilowatt than current state-of-the-art Nvidia Blackwell systems. Developed with Broadcom and supported by OpenAI’s own models, Jalapeño is tailored for large-scale AI inference. Its full-stack design targets bottlenecks in prefill and communication phases, optimizing how compute, memory, and networking resources are orchestrated during processing. OpenAI plans limited Jalapeño deployment by late 2026, ramping in 2027 as part of a multigenerational platform strategy. The results suggest stronger in-house hardware capabilities, potentially reshaping competition in AI infrastructure and influencing future data center investment decisions.
This update represents a notable development in the Ai sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.