OpenAI publishes early results for its Jalapeño inference chip
OpenAI's first custom inference accelerator points to a future where model serving is optimized across silicon, software, and product workloads together.
OpenAI has shared its first measured results for Jalapeño, a custom accelerator built specifically around language-model inference. The company tested the system with public models including GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.
The interesting part is not a single benchmark win. It is the decision to optimize the chip, memory path, networking, kernels, serving stack, and target workloads as one system. Interactive agents need both low response latency and enough aggregate throughput to serve many users, two objectives that often pull hardware configuration in different directions.
OpenAI reports higher performance per watt and lower end-to-end latency than the comparison systems in its chosen tests. These are vendor-published results and should be evaluated against independent measurements as the hardware reaches wider deployment.
Why it matters: Frontier AI companies are becoming infrastructure companies. Custom silicon can reduce dependence on general-purpose accelerators, but its real advantage will depend on software maturity, manufacturing scale, and how well workloads map to the architecture.
Source attribution
Original reporting and release information: OpenAI.