OpenAI has disclosed benchmark results for its Jalapeño AI chip (co-designed with Broadcom and Celestica, launched June 2026) showing significant performance advantages. Testing on inference workloads with three open-source models (GPT-OSS 120B, DeepSeek R1, Kimi K2.5 1T) demonstrated 1.5–1.9× higher throughput per watt and 1.7–3.6× lower latency compared to alternatives—breaking the typical trade-off between capacity and speed. Jalapeño deployment will be modest by year-end 2026, scaling up in 2027, though OpenAI will not replace existing chips entirely.