OpenAI’s Jalapeño Chip Tops Blackwell

Jalapeño looks like a surprisingly strong first-generation inference ASIC, but the headline numbers need qualification: the chip is still in limited deployment, and the comparisons are not yet independently reproducible at scale.

• –Its strongest advantage is serving efficiency, directly targeting OpenAI’s enormous inference costs.
• –Generalized performance across OpenAI, DeepSeek, and Moonshot models makes this more credible than a narrowly optimized accelerator.
• –Lower latency matters especially for agentic workloads, where delays compound across many sequential model calls.
• –Jalapeño is inference-only, so OpenAI remains dependent on Nvidia and other partners for training capacity.
• –The bigger strategic shift is full-stack control: OpenAI can co-design models, kernels, serving software, memory, networking, and silicon.