Richard Ho, OpenAI’s vice president of hardware, said the company’s goal was to have Jalapeño in production in the first half of 2027, with some users potentially running some models on the chip by then. Ho called the schedule a stretch goal, citing software work and other preparations still underway. He made the remarks in an interview with Richard Nieva, published by Forbes on October 12, 2026.
Jalapeño is OpenAI’s custom inference accelerator, developed with Broadcom. Its production target describes an early milestone, not the end of a reported year-long ramp.
Three separate milestones shape Jalapeño’s timeline
OpenAI said on August 25, 2026, that it planned to begin deploying Jalapeño within its compute infrastructure by the end of that year. The company’s post on its deployment plan also published its first benchmark results.
Ho’s first-half-2027 goal concerns having something in production, with possible use limited to some users and some models. Separately, SemiAnalysis reported a gradual production ramp during 2027, with most output scheduled for year-end. An initial deployment, an early production milestone and a later increase in output are distinct steps; the reported year-end schedule does not turn Ho’s goal into a full-ramp deadline.
What OpenAI’s InferenceX results measure
In results published August 25, OpenAI reported tests of GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. The workloads used nominal 8k/1k single-token prediction. For those comparisons, OpenAI reported 1.5–1.9× higher peak performance per watt and 1.7–3.6× lower end-to-end latency than the comparison systems. It also reported 2.1–4.1× higher performance for highly interactive workloads.
OpenAI lists Jalapeño’s rated chip power at 700 W and says sustained measured power was at or below 550 W on the tested workloads. These figures describe the company’s stated benchmark conditions, not a single universal performance gain across every model or use case.
Jalapeño adds capacity to OpenAI’s supplier mix
OpenAI and Broadcom announced the inference-chip program on June 24, 2026, describing it as part of a multi-generation compute platform. NeoTeo’s earlier coverage explains Jalapeño’s inference role. In the Forbes interview, Ho said NVIDIA would remain part of OpenAI’s compute fleet and described Jalapeño as additional capacity; he also named AMD and Cerebras among the company’s compute partners.
