NVIDIA

Overview

A GPU manufacturer and the de facto standard for AI compute infrastructure. The infrastructure core of the AI era. Also active in AI research — particularly meaningful in-house work in the directions of embodied AI and agentic AI.

Key People

• Jensen Huang — CEO
• Jim Fan — Senior Research Scientist, AI Agents

Models

• Cosmos 3 Super — Cosmos 3 Super (64B, MIT open weights). Released 2026-06-01. The world's first open omnimodal physical AI foundation model. R-Bench #1 open model.
• Cosmos 3 Nano — small, real-time inference
• Cosmos 3 Edge (coming soon) — edge on-device

Research Streams

• Cosmos 3 (2026-06-01) — physical AI omni-model (unified world generation + reasoning + action). Trained on 20T tokens. MoT architecture. → Cosmos 3 Super
• Project GR00T — humanoid robot foundation model
• CaP-X (2026-04, open-source) — vibe agents alive in physical world (Jim Fan team)
• DrEureka — LLM writes robot skill training code
• Voyager — lifelong learning agent in the Minecraft environment

Business Significance

• Every frontier lab depends on its compute — Anthropic, OpenAI, DeepMind, Meta, and others
• The OpenAI Broadcom partnership (2026~) is an attempt to diversify away from NVIDIA dependence
• Aggressively recruits talent under the "warm GPUs" motto

Recent Activity

• 2026-06-17: ENPIRE — agentic robot self-improvement on real hardware — Jim Fan's NVIDIA GEAR Lab (with CMU, UC Berkeley) published ENPIRE: a system where a fleet of 8 real robots autonomously runs its own research loop — reading papers, proposing hypotheses, resetting physical scenes, running trials, verifying results, and rewriting control code — with zero human researchers in the loop. Results: 99% pass@8 on contact-rich tasks (GPU seating, zip-tie tying). New finding: a physical scaling law — 8 parallel robots improve policies superlinearly faster than fewer robots. The first empirical demonstration of a data-parallel scaling law for real-world physical manipulation. → ENPIRE: Agentic Robot Policy Self-Improvement in the Real World, Jim Fan (source) (arXiv)

• 2026-06-07: EgoScale — humanoid dexterous manipulation from egocentric human video — Jim Fan's team trained a humanoid with 22-DoF dexterous hands (Sharpa Wave tactile, on a Unitree H2 Plus chassis) to assemble model cars, operate syringes, sort poker cards, and fold shirts. Training data: 20,000+ hours of egocentric human video, zero robot teleoperation. Key finding: log-linear scaling law (R² = 0.998) between human video volume and action prediction loss, which directly predicts real-robot success rate. Also demonstrated live VR teleoperation inside a virtual environment on Unitree G1 (PICO headset). Open-sourced: weights, code, post-training dataset, eval set, whitepaper. → Jim Fan (source)

• 2026-06-01: NVIDIA Cosmos 3 released — the world's first fully open omnimodal physical AI foundation model. Cosmos 3 Super (64B, MIT license), Nano, Edge (coming soon). R-Bench #1 open model. Trained on 20T tokens (1B images, 400M videos). Cosmos Coalition launched: global partnerships including Agile Robots, Black Forest Labs, Runway, Skild AI. → Cosmos 3 Super (source)

• 2026-04: CaP-X open-sourced (Jim Fan) (source)

Strategic Position

• Compute supplier + AI researcher dual role
• Particular emphasis on embodied AI — Project GR00T is the cornerstone
• Compared to AI labs, emphasizes "infrastructure for agents to operate in environments" over the models themselves

Related

• Jim Fan
• Embodied Agents
• Cosmos 3 Super
• Project GR00T page TBD