NVIDIA details Rubin GPU and Vera CPU architectures for agentic AI

In a cluster of developer-blog posts, NVIDIA framed its next platform around agentic workloads. The Rubin GPU architecture is positioned for continuous 'AI factories' that produce intelligence at scale, marking a shift from discrete model training to always-on agent inference. The Vera CPU, built with Olympus cores for maximum single-thread performance, reflects NVIDIA's argument that as agents run code, invoke tools and retrieve context, more of the latency-critical path lands on the CPU rather than the GPU.

On the manufacturing side, contract maker Wistron opened its first US plant — a 324,000-square-foot greenfield facility in Fort Worth, Texas — to build NVIDIA superchips domestically, reinforcing US-based AI infrastructure amid supply-chain and policy pressure. NVIDIA also reported a mixture-of-experts pre-training world record on its GB300 NVL72 system, arguing MoE has become the dominant frontier-training approach and reshaping what now limits large-scale training.

The competitive backdrop is intensifying: the CUDA software moat is under attack from startups like Infinity, DeepSeek is building its own inference chip, and Google is developing 'Frozen v2.' NVIDIA's answer is a full-stack agentic story — GPU, CPU, domestic manufacturing and training records.

Watch whether Rubin/Vera adoption and the agentic-CPU thesis hold up against these efficiency-focused challengers.