We host AI agents on the Kubernetes we already run, each agent in its own virtual-machine-backed sandbox, on GPU servers we build in house. Vendor agnostic: NVIDIA, AMD, Intel, Ascend, Tenstorrent.
Agents run on the Kubernetes we already operate, each in its own sandbox, so the isolation boundary is a virtual machine rather than a namespace.
Every hosted agent gets its own sandbox on the cluster we already operate. The workload looks like a container to the orchestrator and behaves like a virtual machine to the host, which is what you want when the code running is not yours.
We do not build around one GPU vendor. The stack is chosen for the workload, not the brand.
We work across NVIDIA, AMD, Intel, Ascend and Tenstorrent. Vendor choice follows the model, the toolchain and the budget — not a partnership agreement.
We build our own GPU servers: four to six GPUs per server, liquid cooled, in a chassis we design rather than buy. Building them ourselves is what lets us mix vendors in the same rack.
