NVIDIA GR00T 2 paper reveals a frontier robotics stack where only 50% of a 2,000-GPU cluster goes to pretraining
skypilot_org · x · 2026-11-10
The SkyPilot team, working with NVIDIA Robotics on GR00T 2 (a world-action model driving parallel grippers, dexterous hands, and humanoids), shares a rare look at a frontier robotics lab's infrastructure:
- Highly mixed workloads: pretraining, RL post-training, simulation, eval, inference, plus data processing on 10k vCPUs.
- Heterogeneous fleet: GB300 NVL72 + H100 for training; RTX PRO 6000 + CPUs for data processing; L40 for sim and eval; Jetson Thor for edge inference. Stack: PyTorch + Ray + SkyPilot + Kubernetes (NVSentinel).
- Under-discussed point: only 50% of the 2,000-GPU cluster is used for pretraining — the other half serves researchers (and their agents) firing off many small jobs to test ideas.
The author argues frontier AI stacks must optimize for research efficiency, not just hero runs: the more jobs/ideas/agents a cluster can support per unit of compute, the better.
Related event: NVIDIA Open-Sources GR00T 2, a World-Action Model for Robots(20 posts)→
More from Embodied
- Wired tries the OpenAI-powered Microduck, a tiny robot duck that chases cats — Wired AI · 2026-11-12
- RoboCap-10k released: 10,000 hours of open egocentric robot data with metric actions — databoydg · 2026-11-12
- Claude Fable 5.2 Solves Rubik's Cube on Trossen WidowX Robot Arms in Anthropic Launch Demo — kscottz · 2026-11-12
- Scoble: AI glasses will get smaller and cheaper, and China's market appetite dwarfs the US — Scobleizer · 2026-11-12
- Blackmagic launches 108-mic 7th-order ambisonic microphone priced at $1,909 — Scobleizer · 2026-11-12
- NVIDIA unveils GR00T 2 humanoid open world action model, benchmarked on AI2's MolmoSpaces — Kyle_L_Wiggers · 2026-11-12