OpenAI is taking a major step beyond AI models with Jalapeño, its first custom AI inference chip, developed with Broadcom. The chip is designed specifically for running already-trained models and generating responses, rather than training new models. OpenAI says the move gives it greater control over the infrastructure behind products such as ChatGPT while potentially lowering the cost and energy required to serve rapidly growing AI demand.
The most attention-grabbing part is the performance claim against Nvidia hardware. OpenAI says Jalapeño delivered 1.5–1.9 times more AI work per watt and 1.7–3.6 times lower end-to-end latency than the Nvidia GB200 and GB300 systems used for comparison across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. For highly interactive workloads, OpenAI reports 2.1–4.1 times higher performance.
However, this does not mean Nvidia has suddenly been displaced. Jalapeño is optimized for inference, while Nvidia's accelerators support both training and inference, and OpenAI explicitly says it will continue deploying Nvidia and other accelerators widely. Jalapeño is also not being sold commercially; OpenAI plans limited deployment toward the end of 2026 and broader scaling in 2027.
The broader takeaway is that AI companies are increasingly becoming semiconductor companies themselves. OpenAI says its models helped accelerate Jalapeño's development, while the resulting chip is designed around its own models, software and serving workloads. This points toward a full-stack AI race in which models, software, chips, memory and data-center systems are designed together. For Nvidia, the threat is therefore less about losing OpenAI overnight and more about major AI customers gradually developing specialized alternatives that reduce their dependence on general-purpose GPUs.
