OpenAI published benchmark results for its custom Jalapeño chip showing it outperforms Nvidia GB300 processors with 1.5 to 1.9 times more performance per watt and latency reduced by 1.7 to 3.6 times. This ASIC dedicated to AI inference, developed with Broadcom and manufactured by TSMC, operates at 550W with a roughly 50% cost advantage compared to Nvidia’s Blackwell architecture. OpenAI plans limited deployment by end of 2026 and full rollout in 2027, while continuing to purchase Nvidia and AMD GPUs for model training. This initiative is part of a broader trend where major AI players develop custom chips to reduce their dependence on Nvidia.
Source: Read the original article
