Together AI began selling spare GPU capacity at half its standard cluster rate on September 10, 2026. Its new preemptible nodes use the same NVIDIA GPUs as regular nodes and join existing Kubernetes clusters in all supported regions. Together can reclaim that capacity when it needs the GPUs elsewhere.
The price comes with interruptions. When Together needs a preemptible node back, Kubernetes marks it unavailable and sends a termination signal to its workloads. A job gets up to five minutes to save a checkpoint and exit. The node is then removed. Together tries to refill the cluster toward its requested preemptible capacity as GPUs become available, but replacement is not immediate or guaranteed.
The feature suits retryable experiments, evaluations, batch inference and workloads that already save checkpoints. Coordinators and services with strict uptime requirements should stay on standard nodes. Teams can direct suitable tasks to discounted capacity using the together.ai/compute-class=preemptible node label.
Usage is metered every one to two minutes rather than billed in full-hour blocks. Together's published pricing lists an H100 at $1.99 per GPU-hour when preemptible, versus $3.99 on demand.