OpenAI's custom inference chip Jalapeño, developed in collaboration with Broadcom, has outperformed NVIDIA's Blackwell GPU in early benchmark tests, showcasing higher inference throughput and energy efficiency. The results are based on FP4 precision and inference-specific metrics. OpenAI's Head of Hardware, Richard Ho, described the performance improvement as 'very, very significant.' The Jalapeño die size is similar to NVIDIA's Rubin, but with NVFP4 peak compute power at about one-third, it still achieved a lead.