Descriptions:
At Nvidia GTC Taipei, CEO Jensen Huang unveiled four major updates that collectively push Nvidia deeper into open-source AI, agentic computing infrastructure, and physical AI simulation. TheAIGRID breaks down each announcement with direct keynote quotes and technical context.
The headliner is Nemo Tron 3 Ultra — a 550 billion total parameter open-source model built on a hybrid Mamba Transformer mixture-of-experts architecture, with 55 billion active parameters per token. Nvidia claims it is five times faster and 30 percent cheaper in total FLOPs than the best comparable open models, and is releasing not just the weights but the full training scripts and data. Alongside it, Nvidia introduced Vera, a CPU designed explicitly for the agentic era. With 88 custom Olympus cores, LPDDR5X memory delivering 40 percent lower peak latency versus x86, and a scalable coherency fabric unifying compute without chiplet splits, Vera is positioned as the orchestration layer for GPU-heavy AI factories running tool calls, sandbox execution, and retrieval workloads.
The third major release is Cosmos 3, an omni-model for physical AI trained on 20 trillion tokens of multimodal data — including nearly 4 billion images and 400 million videos — that unifies prediction, domain transfer, physical reasoning, and action generation in a single architecture. A nano variant targets edge deployment for robotics and autonomous vehicles. Together, these announcements signal Nvidia’s intent to own the full AI stack from silicon to world simulation.
📺 Source: TheAIGRID · Published June 02, 2026
🏷️ Format: News Analysis
