Vera Rubin entered full production, announced at GTC Berlin on October 21. Alongside it, NVIDIA unveiled the DSX POD, a modular containerized AI factory design that reduces on-site deployment time from six months to as little as 14 days, with partner deliveries starting in early 2027. Jensen Huang's keynote introduced open frontier models across six domains: Nemotron (language), Cosmos (world foundation), GR00T (humanoid robotics), Alpamayo (autonomous vehicles), Ising (quantum-classical), and Earth-2 (physics AI). NVIDIA stated it has been building these model families for a decade and intends to continue.
New Nemotron 3.5 releases include Super VL, Bolt, and retriever models for enterprise post-training. The Nemotron Coalition includes European partners Black Forest Labs, H-Company, and Mistral. Companies including Dassault Systèmes, Salesforce, SAP, Shopify, and WPP are building custom models on Nemotron.
The DSX platform is a full-stack AI factory reference architecture integrating compute, networking, storage, software, power, and cooling. European AI clouds and infrastructure providers are using DSX to build AI infrastructure for enterprises, developers, and governments. Power infrastructure is described as the primary bottleneck on AI factory growth, with over $300 billion in equipment backlog and more than 200 gigawatts waiting in U.S. interconnection queues. GE Vernova, Hitachi Energy, Schneider Electric, and Siemens Energy are building products and reference architectures validated against DSX requirements.
Ahead of GTC Berlin, NVIDIA indicated plans to announce approximately $2 billion in new European investments; combined with roughly $8 billion invested over the prior year, total European commitments could reach nearly $10 billion.
Post-training results demonstrated at the event include Shopify's Nemotron 3.5 Lightning fine-tune for its Sidekick merchant assistant, which surpasses a leading frontier model on Shopify's internal benchmark in accuracy while cutting serving costs by 90% and delivering responses 7x faster. You.com, NVIDIA, and CoreWeave post-trained Nemotron 3.5 Lightning using supervised fine-tuning and RL rollouts, lifting BrowseComp accuracy from 36.97% to 45.45% while reducing tool calls by 30% per task, completed in three weeks.
The Vera Rubin agentic system comprises four components: NVL72 for models, MGX for secure agent processing, CMX for knowledge serving, and RTX for CUDA-X tools. NVIDIA also introduced OpenShell, a secure AI agent reference architecture for enterprise use.