Together AI Cuts GPU Cluster Pricing in Half for Interruptible Workloads

Together AI has opened a public preview of preemptible compute for its GPU Clusters, offering NVIDIA GPU nodes at a flat half of the on-demand price. The trade is direct: teams can cut costs for restartable work, but Together can reclaim that capacity when it needs it elsewhere.

The preview is available on Kubernetes clusters in all regions and can be added to new or existing Together GPU Clusters. Rather than creating a separate cluster, the discounted nodes join an existing one alongside standard nodes, which Together says are fulfilled synchronously and are never preempted.

A discount built on spare capacity

Preemptible nodes draw from unused capacity on the same NVIDIA accelerated-computing infrastructure. They are billed sub-hourly, with usage measured about every one to two minutes, so a short-lived worker is charged for roughly the time it runs rather than a full hour.

The interruption is the product constraint

When Together reclaims a node, it marks the node unavailable, sends affected pods a termination signal and emits a Kubernetes event. The workload then has a drain period of up to five minutes to checkpoint—save enough state to resume later—and shut down before the node disappears.

That makes recovery behavior, not GPU type, the deciding factor. Together recommends the capacity for short experiments, evaluations, fine-tuning, distillation, batch inference and other work that can resume from a checkpoint, retry or requeue. It says multi-day jobs without checkpoints and strict service-level serving without a fallback are not suitable.

How customers separate cheap capacity from critical services

• Schedule interruption-tolerant workers onto nodes carrying Together’s preemptible-compute label.
• Keep coordinators, login pods and user-facing serving replicas on standard nodes.
• Set workloads to handle termination signals, save progress and exit within the drain window.

Capacity is cheaper, not guaranteed

Together will automatically refill toward a cluster’s requested preemptible target as unused capacity becomes available. But the live allocation can remain below that target when capacity is tight, and the preview has no minimum node lifetime. Customers therefore cannot treat the configured target as a reservation.

The setup also preserves a standard-capacity floor: every cluster must have at least one standard node, and nodes cannot be converted between standard and preemptible types in place during the preview. Together says support for Slurm, more regions and in-place conversion are planned next. For now, the value proposition rests on a practical question for teams: whether their software can make an abrupt GPU departure routine rather than disruptive.