OpenAI has announced "Jalapeño," a generalized inference ASIC developed in partnership with Broadcom, designed to outperform NVIDIA and AMD hardware by prioritizing performance-per-watt (perf/W). Developed in an accelerated 16-month cycle, the chip utilizes HBM4 to achieve 15.4 TB/s of memory bandwidth, outperforming current HBM3E accelerators. Architecturally, Jalapeño eschews the common practice of prefill-decode disaggregation (PDD) in favor of a homogenous pool of cores to maintain high global utilization across fluctuating workloads. The design features an out-of-order (OoO) core with L1 cache and a weight-stationary systolic array matrix engine, utilizing MXFP4 numerical formats. Benchmarks on the InferenceX suite show Jalapeño significantly outperforming NVIDIA’s GB200 in token throughput per MW, even without Multi Token Prediction (MTP). In single-token prediction (STP) scenarios, the chip reached over 700 tokens per second per user on the DeepSeek R1 model at concurrency 1. Furthermore, the B0 stepping of the silicon, manufactured on TSMC’s N3P node, is projected to deliver 13.4 PFLOPs of MXFP4 with a 700W TDP. OpenAI’s software stack relies on "Gluon," a low-level programming language built on Triton, and "Codex" to automate the generation of highly optimized, hand-tuned kernels, effectively bypassing traditional compiler limitations.