From a GPU per robot
to one pool per fleet.
A server GPU has about 29× the memory bandwidth of a robot chip from the same generation.
On-site compute for your robot deployments
We put a shared GPU pool at your customer's site, one hop from every robot, and run it for you.
Box
Box
Backed by & building with
Why fleet compute
Smarter robots need bigger brains.
A GPU on every robot works for today's models. It won't for the next ones: onboard chips are capped by the robot's power budget, and robot models keep growing, with no sign of stopping.
From a GPU per robot
From one modelto the whole pipeline.
Act, plan, check safety, track progress: each with a deadline.
From the fastest answerto the most finished work.
Up to 12× the output of a GPU per robot.
Sharing adds a network and a fleet to run. That's the part we run.
NVIDIA Research + Stanford, ROSA (2026). Their numbers, not ours yet.
Read the long page
What we run for you
Bring your models. Connect your fleet. We run the pool.
Inside the node
at the site · on the site's LAN
Hardware
The Box
The GPUs at the site. Sized in robots, not chips, with headroom to grow.
Software · operator console
The Console
The screen you run it from. Declare your fleet, check every model's deadline, see what runs where. Try it in the sandbox.
Software · reliability loop
The Brain
The loop that keeps the pool dependable. Fuses GPU, power, cooling, network and workload signals; holds tail latency in band.
No field team
We install, monitor and maintain the hardware at every site. Your team builds robots.
One compute tier
The same standard pool at every deployment, sized in robots.
No upfront cost
Billed monthly per deployment. Bundle it into your robots-as-a-service price, or pass it through.
How it starts
Try it, run your model, then one site.
Three steps, each one non-binding until the pilot.
01
The sandbox
Twenty minutes in the operator console, the one that runs on the box, on seeded telemetry. By invite.
02
A hosted trial
Your model on GPUs we operate, for about a week at no cost, to show it runs on our pool and give you a sizing range.
03
An on-site pilot
One Box at one of your customer sites, billed monthly from day one.
FAQ
Questions robot teams ask
On-site compute for robot deployments, run as a service. We put a shared GPU pool at the site, one hop from every robot, and operate it: install, monitoring, 24/7 operations and hardware refresh. We call it fleet compute: one pool serves the whole fleet's real-time inference, close enough for the control loop and big enough for the model, and trains between shifts.
Tight control loops run 30-200 Hz - a 5-33 ms budget. Practical cloud round-trips measure 50-150 ms: outside the loop, before egress cost. The Box keeps inference and data on-site.
Neither. The Box serves the inference your stack calls; your agent control and orchestration stay yours, with a safe fallback if the Box is unavailable. Reflexes stay on the robot - ~100 Hz local control and safety fallback - while the fleet's heavy models run from the pool.
More questions - deployment, hardware, data, the sandbox - are answered on the long page. Still weighing it? The sandbox is non-binding - start there.
Ask for a link
Try it out.
Start with the sandbox, or a hosted trial on your own model. Alex reads every request and replies within a business day.
- The console that runs the box.
- Non-binding.
- Workload data stays on the site in production.
Request received.
Alex will reply within one business day.