According to Jeff Dean on X, a new GTC 2026 video features his discussion with NVIDIA’s Bill Dally covering computer architecture, model training pipelines, specialized inference hardware, and custom interconnects. As reported by Jeff Dean’s post, the conversation examines compute–memory balance in modern architectures, the scaling demands of model training, and how custom interconnects improve cluster efficiency for large language models. According to Jeff Dean’s announcement, the session also highlights opportunities for domain-specific accelerators to cut inference latency and cost, offering practical guidance for enterprises deploying generative AI at scale.