tl;dr: OpenAI unveiled "Jalapeño," a Broadcom-partnered inference ASIC that reportedly beats Nvidia Blackwell (and matches/exceeds Rubin) on perf/watt across DeepSeek R1, Kimi K2.5, and GPT-OSS, achieved without speculative decoding or prefill-decode disaggregation. Key architectural choices include a unified chip pool, out-of-order cores with L1 cache, HBM4 at 15.4TB/s, MXFP systolic arrays, and a 2,048-XPU scale-up network using Broadcom Tomahawk 6 switches. Taped out in Nov 2025 with production ramping through 2027, the chip leans heavily on Codex-generated kernels via OpenAI's Gluon language, suggesting AI-assisted design is meaningfully compressing ASIC timelines.