OpenAI Jalapeño NEW

OpenAI Jalapeño is the first accelerator OpenAI has designed itself, an inference-only chip built with Broadcom and unveiled in June 2026. OpenAI's Hot Chips 2026 specification slide gives it 13.4 PFLOPS of MXFP4 matrix compute and 3.4 PFLOPS of MXFP8, 216 GiB of HBM4 on six stacks at 15.4 TB/s, and a 700 W package. The FP4 rate is four times the FP8 rate rather than the usual two, which tells you where OpenAI spent its silicon: on low-precision inference, not training. The design keeps model state, including the KV cache, next to the compute that uses it, and scales over Ethernet through Broadcom Tomahawk 6 switches rather than NVLink: 128 chips share a 600 GB/s local domain and up to 2,048 chips form one system. OpenAI's own InferenceX results put it at 1.5 to 1.9 times the throughput per watt of NVIDIA GB200 and GB300 systems on open models. Those are OpenAI's numbers on early silicon. You cannot rent or buy Jalapeño; it will serve ChatGPT, Codex and the OpenAI API, and its effect on buyers is through what OpenAI charges for inference.