GPU capacity at owned-hardware economics.
AI teams are racking GPUs faster than their ops tooling can absorb, across providers, sites, and countries. At the same time, models and training data are exactly the assets many companies refuse to put on rented clouds.
No separate pipeline for GPU hardware
A GPU server enrolls, installs from a versioned image, and joins the private mesh by the same zero-touch path as everything else: no GPU-specific tooling to build or maintain. Its inventory (GPUs, memory, disks, link quality) appears in the fleet view.
Capacity wherever you can buy it
Build training and inference capacity across your own sites and providers, and run clusters on top the same way. One cluster can pull nodes from a colocation, a rented GPU host, and your own rack at once.
Models and data never leave your control
Everything runs on hardware you own or rent, and traffic flows machine to machine rather than through our control plane. Models, weights, and training data stay inside your perimeter, including when the work spans several sites.