OpenAI Unveils Custom 'Jalapeño' Chip Delivering Industry-Leading Inference Throughput

OpenAI has revealed key performance figures for Jalapeño, demonstrating a 3.2x gain in inference throughput per watt over conventional GPUs.

OpenAI has officially disclosed operational metrics for Jalapeño, its custom-designed ASIC engineered specifically to serve foundation model inference at global scale.

Engineered in collaboration with TSMC on a custom 3nm process node, Jalapeño features high-bandwidth memory stacks integrated directly onto the interposer to eliminate memory bandwidth bottlenecks during autoregressive generation.

Initial production clusters deployed across OpenAI data centers indicate a 60% reduction in total cost of ownership per million tokens generated.