< Back to all clusters
[TECHNOLOGY] · United States · 8 sources

started · updated

OpenAI unveils Jalapeno custom AI chip to rival Nvidia

OpenAI has unveiled performance data for its first custom AI inference chip, named Jalapeno, developed in collaboration with Broadcom and Celestica. Designed specifically to run large language models rather than train them, the chip aims to reduce operational costs and reliance on Nvidia hardware.

Benchmark testing conducted using SemiAnalysis' InferenceX tool indicates that Jalapeno outperforms Nvidia’s GB200 and GB300 systems in several key metrics. Specifically, the chip reportedly delivers between 1.5 and 1.9 times more AI work per watt and achieves 1.7 to 3.6 times lower latency. For highly interactive workloads, such as AI agents, performance advantages can reach up to 4.1 times.

OpenAI utilized its own advanced AI models to assist in the chip's design and circuit verification, which helped shorten the development cycle from concept to tapeout to just nine months. The chip is rated at 700 watts, though testing showed actual power draw remained at or below 550 watts. OpenAI expects to begin limited deployment of the Jalapeno chip by the end of 2026, with broader rollout following in 2027.

Entities

Broadcom · Celestica · Jalapeno · Nvidia · OpenAI · TSMC

Claims

What the coverage asserts, and how many sources carry each claim.