OpenAI introduced its latest inference ASIC, named Jalapeño, at the Hot Chips 2026 conference with benchmark comparisons showing promising performance and efficiency gains over existing Nvidia Blackwell systems. This presentation underlined advancements in ASIC technology aimed at optimizing AI inference.
At the Hot Chips 2026 symposium held from August 23 to August 25, OpenAI showcased its Jalapeño inference ASIC during the final day of the event. The presentation highlighted in-person benchmark results, emphasizing the performance of this new hardware design.
The talk included a detailed comparison using the SemiAnalysis InferenceX benchmark against Nvidia's Blackwell systems. OpenAI reported that their engineering samples demonstrate higher throughput-per-watt and reduced end-to-end latency relative to the competing systems.
This development marks a significant step in inference acceleration hardware, focusing on efficiency and performance improvements in artificial intelligence workloads. The Jalapeño ASIC insights provided valuable information to industry professionals attending the Hot Chips symposium organized by IEEE and other collaborative entities.