Token factory for Physical AI · Indonesia
The Token Factoryfor Physical AI
Language models read text. Physical AI has to take in the world: video from multiple cameras, depth, motion and action, all arriving continuously. NexInfra runs the GPUs that generate, train and validate those models.
Illustrative estimate by NexInfra, based on the published compression ratios of NVIDIA Cosmos Tokenizer (8×16×16 and 4×8×8) at 1280×720, 24 fps.
What we produce
One factory, three production lines
Physical AI development runs in a loop: generate data, train, validate, repeat. We have dedicated GPU capacity for each stage, so a surge in one stage never stalls the others.
Data Factory
Synthetic data and world-model generation.
- Open world foundation models such as NVIDIA Cosmos
- A few real demonstrations into thousands of variations
- Curate → Augment → Evaluate
Training Clusters
Pre-training and post-training for robot and driving models.
- Multi-node clusters, low-latency fabric
- Fine-tune Isaac GR00T or your own VLA models
- High-throughput storage for video-scale datasets
Inference & Validation
Simulation-based testing before real-world deployment.
- Thousands of scenarios before a policy reaches a robot
- Closed-loop evaluation in digital twins
- Token-based inference endpoints for open models
Who we build for
Built for the teams putting AI into the physical world
Embodied AI & Humanoid Robotics
Teams building general-purpose robot brains: end-to-end VLA models that transfer across humanoids, robot arms and mobile platforms.
foundation-model pre-training · teleoperation data · dexterous manipulation
Autonomous Driving & World Models
End-to-end driving stacks and world models that predict how a scene will unfold.
large-scale video training · scenario generation · long-tail simulation
Robotaxi & Autonomous Mobility
Driverless fleets whose software has to be validated again and again before each release.
log replay · sensor simulation · closed-loop safety validation
Industrial, Logistics & Smart Factory
Manufacturers, warehouses and ports automating flexible, high-mix tasks.
factory digital twins · pick-and-place training · AMR fleet simulation
The stack
Built on the NVIDIA full stack
Physical AI needs three kinds of computers: one to train, one to simulate, and one on the robot to act. NexInfra provides the first two as a service, so you can focus on the third.
| Compute | NVIDIA GB300 and B300 for training and large-scale inference; RTX PRO 6000 Server Edition for simulation, synthetic data and rendering |
|---|---|
| Networking | NVLink within nodes, InfiniBand / Spectrum-X RDMA fabrics between nodes |
| Acceleration | CUDA-X libraries, NCCL for distributed training, TensorRT for optimized inference |
| Physical AI software | NVIDIA Cosmos, Isaac GR00T, Isaac Sim / Isaac Lab, Omniverse |
| Orchestration | Kubernetes with GPU Operator, MIG partitioning, cluster monitoring |
| Deployment target | Models are exported for on-robot computers such as NVIDIA Jetson Thor |
First capacity online in Indonesia: GB300 and B300 training clusters, with RTX PRO 6000 Server Edition nodes for simulation and synthetic data.
Expansion planned on the NVIDIA Vera Rubin platform. Reservations for the first wave of capacity are open now.