Nvidia Cosmos 3 + RTX Spark N1X: 20T Tokens for Physical AI at COMPUTEX

At COMPUTEX Taipei on June 1, 2026, Nvidia launched Cosmos 3, an open physical-AI world model trained on 20 trillion tokens of multimodal data, and RTX Spark laptops powered by the N1X processor — a direct push into Windows silicon alongside Microsoft Surface, Dell, HP, Lenovo, Asus, Acer, and MSI.

CEO Jensen Huang also confirmed Vera Rubin AI platforms in full production with fall 2026 shipments — the same week BIS tightened China export rules on Rubin-class chips.

What Is Cosmos 3?

Cosmos 3 is an open world model for robots, autonomous vehicles, and embodied systems — not a chatbot and not a generic video generator.

Nvidia says training data includes:

• ~1 billion images
• ~400 million real and synthetic videos
• Ambient audio, text, and human + robot action traces

The differentiator is action output: joint angles, gripper poses, and trajectories machines can train on — modeled by Ming-Yu Liu, VP of Nvidia's Cosmos Lab, as "autonomous actions are key."

Partners at launch include Agile Robots, Black Forest Labs, and Runway. Nvidia pitches Cosmos for rare or dangerous scenarios (collisions, edge-case road events) that are expensive to capture in the real world.

RTX Spark and N1X: Nvidia Enters the Windows Laptop Stack

RTX Spark brings N1X into consumer and prosumer laptops — historically Intel, AMD, and Apple Silicon territory.

For the full Vera CPU launch (88 Olympus cores, 1.8× vs x86 on agent sandboxes, NYSE/Anthropic/OpenAI adopters), see Nvidia Vera CPU for AI Agents. For Jensen Huang's N1X keynote (full CUDA on Windows, MediaTek, 128 GB unified memory), see Jensen Huang Nvidia N1X MediaTek. COMPUTEX also covered RTX Spark N1X laptops separately from datacenter Vera.

For developers, the message is platform consolidation: train in Cosmos, simulate on Nvidia GPUs, deploy agents on Vera, ship edge models on N1X — one vendor ladder from datacenter to laptop.

How This Intersects With Export Controls

The same week, US Commerce closed the China-HQ subsidiary loophole on Rubin/Blackwell exports. Cosmos is open and software — harder to embargo than silicon — but Rubin clusters training the largest Cosmos revisions remain subject to BIS licensing.

Physical-AI startups in China will accelerate Huawei/Ascend and domestic robot stacks; US and allied builders get the open weights + CUDA path.

Cross-read US BIS Closes Nvidia Blackwell Loophole and China Humanoid Robot 29-Digit IDs.

Key Takeaways

• June 1, 2026 (COMPUTEX): Cosmos 3 open world model — 20T tokens, super + nano shipping, edge soon
• Action-data focus: joint angles, grippers, trajectories — built for robots and AVs, not text-only AI
• RTX Spark / N1X: Nvidia enters Windows laptop market with Microsoft + major OEMs
• Vera Rubin: full production, fall 2026 shipments — same chip families under tighter export controls
• For developers: physical AI stack consolidating under Nvidia open models + silicon; plan compliance if training on Rubin in Asia

Sources

• Nvidia expands AI push with Cosmos 3 world model — Axios, June 1 2026
• Nvidia Windows PC chip and Cosmos 3 — trade press COMPUTEX coverage