OpenAI Unveils Jalapeño, a Custom AI Inference Chip Claiming to Outperform Nvidia Blackwell

OpenAI has introduced its first custom silicon, a 700-watt inference chip codenamed Jalapeño, co-developed with Broadcom. Benchmarks presented at Hot Chips suggest the application-specific integrated circuit (ASIC) outperforms Nvidia's Blackwell accelerators in efficiency and latency for running AI models, a critical factor in reducing computing costs for AI services.