Platform
Three production lines, one factory
Every Physical AI team runs the same loop: make data, train on it, prove the result in simulation. Each stage wants different hardware and a different balance of storage, network and GPU. We keep capacity for all three, so work never queues behind itself.
Data Factory
Real robot data is slow and expensive to collect. Synthetic worlds close the gap.
We run open world foundation models on GPU capacity reserved for generation, so a data push never eats into training time.
- What runs here
- Open world foundation models such as NVIDIA Cosmos, plus the curate–augment–evaluate workflow: filter and caption raw footage, expand it across lighting, weather, materials and rare cases, then score what is fit to train on.
- What you send in
- Recorded demonstrations, teleoperation logs, fleet footage, CAD or USD scenes of the environment.
- What comes out
- Curated, labelled datasets and generated scenes, ready to train on and stored beside the training clusters.
- Hardware
- NVIDIA RTX PRO 6000 Server Edition nodes for generation and rendering, with high-throughput storage.
Training Clusters
Pre-training and post-training for the models that move machines.
Multi-node clusters with the fabric and storage that video-scale training actually needs, not a general-purpose VM pool.
- What runs here
- Pre-training a vision-language-action model from scratch, fine-tuning an open robot foundation model such as NVIDIA Isaac GR00T, or reinforcement and imitation learning across thousands of parallel simulated environments.
- Interconnect
- NVLink inside the node and InfiniBand or Spectrum-X RDMA between nodes, because on a job this size the network, not the GPU, is usually what stalls.
- Storage
- Datasets stay next to the compute. No moving terabytes of video between stages for every experiment.
- Hardware
- NVIDIA GB300 and B300 clusters, reserved by contract rather than shared with unrelated tenants.
Inference & Validation
A policy that works in the lab still has to survive the long tail.
Closed-loop evaluation in digital twins, run at scale, so a regression is caught in simulation rather than on a factory floor or a public road.
- What runs here
- Scenario sweeps in NVIDIA Isaac Sim / Isaac Lab and Omniverse, log replay against recorded fleet data, and benchmark suites re-run on every model release.
- Inference endpoints
- Token-based endpoints for open world and reasoning models, for teams that want inference without managing a cluster.
- What comes out
- Pass/fail results per scenario, regression reports between releases, and the traces behind each failure.
- Then what
- Validated policies are exported for the robot's own computer, such as NVIDIA Jetson Thor. That third computer is yours; the first two are ours.
Division of labour
What you bring, what we run
You bring
- Your models, policies and checkpoints
- Robot descriptions (URDF / USD) and recorded demonstrations
- The scenarios your system has to survive
- Your team's own tooling and experiment tracking
We run
- GPU capacity sized per stage, reserved for you
- Cluster setup, drivers, orchestration and monitoring
- Storage and networking built for video-scale datasets
- The Physical AI toolchain, kept current
How it works
From first call to first run
Scope
Tell us about your models, data volume and timeline.
Design
We propose the GPU mix, cluster size and pipeline setup.
Pilot
Run a proof of concept on real capacity.
Scale
Move to reserved capacity as your models grow.
Engagements are quoted to the workload — on-demand tokens, GPU-hours, or a dedicated cluster with its own SLA. Tell us what you are running and we will come back with a proposal.