Empty or with a model
Provision a bare slice for interactive work, or deploy a model onto it in a single step.
Auto or manual
Let the platform pick a GPU and carve the slice, or target a specific GPU yourself.
Self-driving agents
AI agents request capacity when they need it and release it when idle — no tickets, no waiting.
Features
Provisioning that fits the work
Auto-provisioning
The platform finds an online GPU with enough free VRAM, carves a right-sized slice, and launches the pod.
Manual placement
Pin a workload to a specific GPU and pod when you need precise control over placement.
Empty pods
Spin up a ready slice without a model for bring-your-own or interactive use, then release it when done.
AI agent capacity
Agents provision and free pods programmatically, so capacity flows with demand instead of sitting idle.
Use cases
Capacity that moves with demand
Bursty agentic workloads
Agents scale capacity up under load and hand it back when the burst passes.
Interactive experiments
Grab an empty pod, do the work, and release the slice — no long-lived reservation.
One-step model serving
Provision a slice and deploy a model from the directory together in a single flow.
Targeted placement
Pin a latency-sensitive workload to a specific GPU for predictable performance.
FAQ
Common questions
What is a pod?
A workload that occupies one VRAM slice on a GPU. A pod can be empty (a ready slice) or run an AI model deployed onto it.
Auto-provisioning vs manual?
Auto-provisioning chooses an online GPU with enough free VRAM and carves the slice for you. Manual lets you target a specific GPU and pod.
How do AI agents get capacity?
Agents request a pod through the platform when they need to run work, and release it when idle — the fleet allocates and reclaims capacity automatically.
Where do the models come from?
Models are deployed from the managed model directory, so a pod-with-model can be launched without per-deploy downloads.