Fixed 50% pricing
Preemptible nodes are billed at a flat 50% of the on-demand rate. The rate remains fixed rather than moving with a spot market.
Usage is metered every one to two minutes, with sub-hourly billing.
On-Demand GPU compute when the workload needs it.
Launch managed NVIDIA GPU clusters in minutes. Pay by the hour, scale to the run, and terminate when the work is done.
Managed infrastructure without the infrastructure backlog.
Self-serve NVIDIA capacity with the networking, orchestration, storage, and operating controls needed for real workloads.
Clusters ready in minutes
Move from request to run without a long procurement cycle or manual setup.
Multi-node fabric
Run distributed workloads over InfiniBand and managed orchestration.
Kubernetes or Slurm
Choose the workload manager that fits the team and the run.
Persistent storage
Keep long-lived, resizable shared storage attached to the work.
Health and recovery
Monitor node health and act on hardware or software faults.
Scale on demand
Scale up to 256 GPUs, subject to real-time availability.
Preemptible compute at half the price
Now in public preview for Together GPU Clusters: interruptible GPU compute, with a five-minute drain window and automatic refill toward your target.
For interruption-tolerant work
Short experiments, batch inference, and temporary bursts are good candidates when interrupted work can resume or retry.
The common thread: work that is chunked, checkpointed, or retryable. Multi-day runs without checkpointing and strict-SLO serving with no fallback are not a fit.
Up to five minutes to checkpoint
When capacity is reclaimed, the node is cordoned and your pods receive SIGTERM. Your workload has up to five minutes to checkpoint and exit before the node is removed.
Automatic capacity refill
The cluster retains its preemptible target after the node is removed and automatically refills toward that target as capacity becomes available. You do not need to request replacement capacity.
Before you enable it
- Treat preemptible nodes as ephemeral. There is no minimum node lifetime during the preview, so checkpoint frequently and plan for capacity to disappear at any time.
- Clusters require at least one standard node , and nodes can't be converted between compute types in place.
- Public preview supports Kubernetes clusters. Slurm support is planned.
The operating model should fit the work.
Use on-demand capacity to learn quickly. Keep control of the environment. Commit only when the workload is stable.
Pay by the hour
Use capacity for the run in front of you. No upfront commitment. Terminate when the work is done.
Provision it yourself
Create and manage clusters through the console, CLI, or API—with Kubernetes or Slurm orchestration.
Reserve when ready
Start on demand for flexibility, then move predictable workloads to reserved capacity without changing platforms.
Capacity for uncertainty.
Reserve after the workload is understood—not before the first run.
Test a hypothesis
Spin up capacity for short experiments, exploratory training, or a new distributed configuration. Shut it down when the answer is clear.
Absorb a temporary spike
Add capacity for a training sprint or increased inference traffic without turning a burst into a long-term commitment.
Establish the baseline
Use on-demand runs to understand utilization, then reserve the capacity that becomes predictable.
How AI-native compute runs in practice.
Product updates and technical guides for provisioning, multi-node training, autoscaling, reliability, and shared cluster operations.
Start with the capacity you need now.
Create a cluster for the next run. Reserve capacity when the workload becomes predictable.