OpenAI has officially entered the custom AI semiconductor arena by revealing the first performance benchmark results for its proprietary AI chip, "Jalapeno." Developed in partnership with Broadcom, Jalapeno is a custom application-specific integrated circuit (ASIC) built specifically for AI inference—the phase where trained models process queries and generate real-time responses.

Unlike chips built for model training, Jalapeno focuses entirely on inference efficiency. At recent benchmark tests, OpenAI demonstrated that Jalapeno outperforms Nvidia’s advanced Blackwell-based systems, such as the GB200 and GB300. It delivered up to 1.9 times higher throughput per kilowatt while consuming roughly 700 watts of power. Furthermore, it achieved a 1.7 to 3.6 times reduction in end-to-end latency across prominent models like GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.

This simultaneous improvement in throughput and low latency is vital for modern AI applications, especially as autonomous AI agents require fast, continuous multi-step reasoning. Because inference drives the vast majority of perpetual cloud operating costs, owning custom silicon allows OpenAI to optimize memory access, reduce energy consumption, and control expenses.

Despite this hardware milestone, OpenAI will not abandon Nvidia. The company plans to maintain a collaborative partnership for heavy training workloads while gradually deploying Jalapeno internally starting late 2026, with full scaling expected in 2027. Jalapeno marks the first step in OpenAI's long-term hardware roadmap, proving that future AI competition relies heavily on infrastructure efficiency.