OpenAI's Jalapeño Inference Chip Nears Deployment

OpenAI reported first results from Jalapeño, its inference accelerator built for low-latency agent workloads, with deployment in its own infra targeted by year-end. AI was used to help design circuits and program kernels, and the system keeps prompt processing and token generation tightly coupled — a sign that inference cost/latency, not just training compute, is becoming the next infra battleground.