GB300 capacity open for reservation — first clusters online January 2027 GB300 reservations open · online Jan 2027

Solutions

Four ways teams use the factory

The models differ, the loop does not. Each of these teams generates data, trains on it and validates the result — they just weight the three stages differently, which is exactly what capacity planning has to account for.

01 · Embodied AI & Humanoid Robotics

The demonstrations are the bottleneck. Not the GPUs.

Teams building general-purpose robot brains — end-to-end vision-language-action models that have to transfer across humanoids, arms and mobile bases.

RECORDED DEMONSTRATIONS hundreds DATA FACTORY randomisation millions of variations TRAINING one policy humanoid · arm mobile base failures return as new variations
A few hundred recorded episodes become millions of randomised variations, and one policy that survives a change of body.

The problem

Real demonstrations are the scarcest resource in robotics. A teleoperation session produces a few hundred episodes; a policy that generalises wants orders of magnitude more — across lighting, object poses, materials and failure cases no operator would think to stage. The largest open real-robot dataset ever assembled, pooled from 34 labs, holds roughly a million trajectories. Most teams have a fraction of a percent of that, and collecting more means a human in a room with a robot, one episode at a time.

How the loop runs here

Generate

Your demonstrations go in; domain-randomised variations come out. World foundation models expand each episode across lighting, backgrounds, object geometry and physical properties, and an evaluator scores which generated data is fit to train on.

Train

Pre-train a VLA model from scratch, or fine-tune an open robot foundation model such as NVIDIA Isaac GR00T on your own embodiment. Reinforcement and imitation learning run across thousands of simulated environments in parallel.

Validate

Dexterous manipulation trials in simulation before a single physical trial — grasp success, contact forces, recovery from disturbance — so the robot lab tests what already works, not what might.

Data in
Teleoperation logs, URDF / USD robot descriptions, RGB-D and tactile streams
Model families
Open VLA models, NVIDIA Isaac GR00T, Cosmos world models
Simulation
Isaac Sim / Isaac Lab, Newton physics for contact-rich manipulation
Deployment target
NVIDIA Jetson Thor and other on-robot computers

What we provide

  • Generation capacity that does not compete with your training jobs for GPUs
  • Multi-node clusters for foundation-model pre-training, reserved by contract
  • Parallel simulation for reinforcement and imitation learning at scale
  • Storage sized for episode libraries that grow every week

Industry context

Open X-Embodiment pooled 60 datasets from 34 robotics labs into over 1 million real robot trajectories across 22 embodiments — still small next to the text corpora behind today's language models.

Source: Open X-Embodiment (RT-X)

NVIDIA Isaac Lab sustains 4,096 parallel environments at over 135,000 frames per second on humanoid locomotion, turning policy training from days into minutes.

Source: NVIDIA Technical Blog

02 · Autonomous Driving & World Models

The scenarios that matter are the ones that never happen.

End-to-end driving stacks and world models trained on fleet video rather than hand-written rules.

FLEET VIDEO terabytes per day CURATE SCORE what matters weather occlusion night · glare GENERATED SCENARIO FAMILIES TRAINING
Curation decides what is worth keeping; generation supplies the cases the fleet never drove.

The problem

A single instrumented test vehicle can produce well over ten terabytes of sensor data in a day, and a fleet turns that into petabytes a week. Almost none of it is interesting. The frames that decide whether a release ships — the occluded pedestrian, the overturned load, the flooded intersection at dusk — appear a handful of times in a year of driving, if at all. Curation is where the compute goes, and generation is how you get the cases the fleet never happened to see.

How the loop runs here

Generate

Raw fleet video is filtered, captioned and scored at GPU speed, then world models extend the interesting fragments into scenario families: same geometry, different weather, different agent behaviour, different time of day.

Train

Large-scale video training for end-to-end policies and world models, on multi-node clusters where the interconnect keeps the GPUs fed rather than waiting on the network.

Validate

Rollouts against held-out scenarios and regression suites, so a model that improves the average case but breaks a known edge case is caught before it reaches a vehicle.

Data in
Multi-camera video, lidar and radar streams, CAN logs, HD map tiles
Model families
End-to-end driving policies, world models, open VLA architectures
Simulation
Scenario generation and sensor simulation on Omniverse-based stacks
Deployment target
In-vehicle compute; models are exported, not hosted by us

What we provide

  • Curation throughput that keeps up with what the fleet uploads
  • Scenario generation for the long tail you cannot collect
  • Training clusters with storage next to the compute, not across a network
  • Regression evaluation on every candidate build

Industry context

One instrumented autonomous test vehicle has been measured generating 11 to 152 TB of data per day; a 200-car fleet puts that in the petabytes per day range.

Source: Tuxera

NVIDIA's pipeline curated 20 million hours of video in 14 days on Blackwell hardware — the same job runs over three years on a CPU-only pipeline.

Source: NVIDIA Newsroom

03 · Robotaxi & Autonomous Mobility

Validation is the release gate.

Driverless fleets whose software is re-validated before every release — for regulators as much as for the road.

SCENARIO LIBRARY · ONE CELL = ONE SIMULATED RUN release gate PASS fail scenario family post-training RE-ENTERS THE LIBRARY · RE-RUN ON THE NEXT BUILD
Each candidate build is replayed across the whole library; a failure becomes a scenario family and re-enters the next run.

The problem

A robotaxi stack does not ship because it passed a benchmark. It ships because every candidate build has been replayed against the whole scenario library, and because someone can show a regulator what happened in each case. That library only grows: every incident, every disengagement, every near-miss becomes a permanent test. Validation stops being an experiment and becomes a recurring industrial workload with a deadline attached.

How the loop runs here

Validate — this stage leads

Log replay against recorded drives, sensor simulation, and closed-loop safety runs across thousands of scenarios in parallel, with pass/fail per scenario and the trace behind every failure.

Generate

Each new failure becomes a scenario family: the same situation under variations the fleet has not driven, so the fix is proven against the class of problem rather than the single recorded instance.

Train

Targeted post-training on the cases that failed, without disturbing what already passes.

Data in
Drive logs, disengagement reports, scenario libraries, sensor recordings
Model families
Planning and perception stacks, end-to-end policies
Simulation
Closed-loop simulation with sensor models and traffic agents
What comes out
Per-scenario results, regression reports between releases, failure traces

What we provide

  • Burst capacity sized to your release cadence, reserved so the deadline holds
  • Parallel scenario execution rather than an overnight queue
  • Reproducible runs: the same build, the same seeds, the same result
  • Artefacts you can hand to a safety case, not just a dashboard number

Industry context

NVIDIA's position at GTC 2026: "Physical AI has arrived — every industrial company will become a robotics company." The validation burden arrives with it.

Source: NVIDIA Newsroom, 16 Mar 2026

04 · Industrial, Logistics & Smart Factory

The line cannot stop for data collection.

Manufacturers, warehouses and ports automating the flexible, high-mix work that fixed automation was never able to handle.

site CAD USD · SKUs DIGITAL TWIN AMR TRAIN policy the real line line changes · re-train against the twin
The twin is built from the CAD you already have, and every line change re-enters it instead of stopping production.

The problem

Traditional automation pays off when the part, the pose and the path never change. The work that is left over — mixed-SKU depalletising, cable routing, screw fastening on a non-standard fixture, picking something soft or transparent — changes constantly, and every site is laid out differently. You cannot halt a line for months of data collection, and a policy trained at one facility rarely survives the move to the next.

How the loop runs here

Generate

A digital twin of the actual site — CAD of the cell, the real SKUs, the real lighting — becomes the source of training data, built without touching the production line.

Train

Manipulation and navigation policies trained against that twin, then hardened with domain randomisation so they survive the gap between the model of the site and the site.

Validate

Full-shift simulated runs and AMR fleet behaviour before a robot is installed, including the failure modes that only appear when several robots share a floor.

Data in
Facility CAD / USD, SKU geometry, cycle-time targets, existing cell layouts
Model families
Manipulation policies, navigation stacks, open robot foundation models
Simulation
Site digital twins, Newton physics for contact-rich assembly, AMR fleet simulation
Deployment target
On-robot and edge compute inside the facility

What we provide

  • Project-shaped capacity: a burst per site or product line, then a steady trickle
  • Digital-twin generation from the CAD you already have
  • Re-training runs when the line changes, without rebuilding from scratch
  • Results that transfer between sites rather than starting over at each one

Industry context

The humanoid robot market is projected to reach US$38 billion by 2035 — more than a sixfold increase over previous forecasts — and the industrial floor is where most of those units are aimed.

Source: NVIDIA Blog

Not on the list?

If your models touch the physical world, the loop is the same

Agriculture, mining, construction, inspection, surgical robotics, defence logistics — the workloads look different on the surface and identical underneath. Tell us what you are training and we will size it.

Contact

Ready to scale your Physical AI?

Tell us what you're building. Our team will reply within 3 business days.

Email

sales@nexinfra.ai

Office

Wisma Tamara, 7th Floor, Suite 701, Jalan Jenderal Sudirman Kav. 24, Karet, Setiabudi, South Jakarta 12920, Indonesia