US Semiconductors: AI Infra Summit Day 1: AI Safety & Alignment Will Require MORE Compute, NOT Less

We attended Day 1 of the AI Infra Summit in Santa Clara, CA and focused on compute-specific takeaways. Frontier AI model developers see AI safety and alignment as a compute-intensive process, implying safety concerns may drive more, rather than less, demand for compute — counter to investor concerns that tighter safety requirements could slow frontier model development and, by extension, compute demand. Developers do not see compute scaling slowing; the bottleneck is shifting from model architecture innovation alone toward securing massive amounts of compute, power and infrastructure, while improving efficiency through platforms such as Astra and next-generation hardware like Vera Rubin. Next-generation facilities could accommodate 30-40% more GPUs. Vera Rubin offers 3x token throughput improvement out of the box and 5x post optimization. Nvidia noted agentic AI is 100x more demanding than Chat in 2023, requiring new benchmarks;