“Broadcom's customized ASICs are more cost-effective and efficient for specific, repetitive AI inference tasks than the general-purpose GPUs used for training large language models.”