Challenge
Spot GPU prices made large simulation sweeps unpredictable and often unaffordable.
Solution
1,024 dedicated H200s on N-01 at a fixed hydro rate for three years.
Result
3.1× the simulation throughput at 29% lower cost per sim-hour.
Simulation is the training set
Vektor builds picking robots for warehouses. Before a new grasping policy reaches a real shelf, it runs through millions of simulated hours: different boxes, different lighting, different ways for things to fall over.
Until 2026 those sweeps ran on spot GPUs across three clouds. When prices rose, experiments were cut, and the robots learned less than the team wanted them to.
“The question used to be what we could afford to simulate. Now it is what we want to learn.”
A fixed rate for a variable workload
Vektor reserved 1,024 H200s on N-01 on a three-year term. Power and cooling are part of a fixed GPU-hour rate, so a week of heavy sweeps costs the same per hour as a quiet one.
How the cluster is used
Simulation workers fill the pod around the clock, scheduled by Slurm.
Policy training takes priority twice a day and pre-empts sim jobs cleanly.
Results stream to Vektor's lab in Gothenburg over a private link.
The result
Simulation throughput is 3.1 times what Vektor ran in the cloud, at 29% lower cost per simulated hour. The team now runs about ten million robot-hours a week, and two new grasping policies have reached production since May.
