NVIDIA AI Updates: June 1, 2026

1. Cosmos 3 Combines Reasoning and Generation in One Physical AI Foundation Model

NVIDIA. NVIDIA introduced Cosmos 3, an open world foundation model that combines vision reasoning and multimodal generation across text, video, images, ambient sound, and action in a single model. It uses a mixture-of-transformers architecture with a reasoning block that interprets a scene and a generation block that produces physically grounded outputs, including native action data such as joint angles, gripper positions, and trajectory points for robots. NVIDIA is distributing the model through its build platform, Hugging Face, and GitHub, with NIM microservice deployment, positioning it for robot task training, autonomous vehicle scenario generation, and synthesis of rare edge cases for safety testing. Source

2. NVIDIA Vera CPU Targets the CPU-Bound Parts of Agentic AI Loops

NVIDIA. NVIDIA detailed the Vera CPU, a processor built around 88 custom Olympus cores and optimized for the CPU-intensive stages of agentic workloads such as sandboxed code and tool execution, data retrieval, scheduling, and orchestration. NVIDIA reports up to 1.2 TB/s of LPDDR5X memory bandwidth, roughly 40% lower peak memory latency than x86 CPUs, and a configurable 250W to 450W TDP, with a neural branch predictor that sustains two taken branches per cycle on branch-heavy agent code. The company claims more than 1.8x higher sandbox performance under full load versus competing CPUs, framing the chip around tokens-per-dollar output for AI factories rather than traditional cloud metrics. Source

3. DSX OS Packages Open Software for Running Multi-Tenant AI Factories

NVIDIA. NVIDIA released DSX OS, a set of open-source, modular software components for operating and scaling multi-tenant AI factories. The stack includes DSX Exchange, an MQTT-based hub linking compute, power, cooling, and facilities systems, alongside power tools (DSX MaxLPS and DSX Flex), the NVIDIA Infra Controller for bare-metal lifecycle management, NVSentinel and Fleet Intelligence for GPU health and fleet visibility, and the KAI Scheduler, Dynamo, and Grove for workload scheduling and inference. NVIDIA says power optimization can let facilities run up to 40% more GPUs at peak energy efficiency within a fixed power budget, while automation shifts operations toward preventive remediation. Source

4. Alpamayo Adds Closed-Loop Post-Training for Autonomous Vehicle Models

NVIDIA. NVIDIA described Alpamayo, an open portfolio of AV models, simulation frameworks, and physical AI datasets that includes the AlpaSim simulator and the AlpaGym closed-loop training framework. Closed-loop post-training lets a driving policy learn from the consequences of its own steering, braking, and navigation decisions inside simulation, addressing how small errors compound over time in real deployment rather than comparing outputs to static ground truth. AlpaGym pairs AlpaSim microservices and NVIDIA Physical AI datasets with distributed Cosmos-RL training, uses GRPO as its default reinforcement learning algorithm, and combines progress rewards with collision and offroad penalties, scaling from a single GPU to multi-node clusters. Source