Introducing preemptible compute: the same compute, half the price
Together AI
Together AI is selling GPU capacity that can be yanked away, for half price. The catch is you get five minutes to save your work before the node disappears.
Based on reporting by Together AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Together AI has put preemptible compute into public preview for Together GPU Clusters, and it’s a straightforward trade: lower prices in exchange for the chance your GPU gets reclaimed. The discount is flat at 50% of the on-demand rate, billing is sub-hourly, and the feature is available on Kubernetes clusters in all regions.
This is not a separate product so much as a second class of node inside the same cluster. Standard nodes stay synchronous and never get preempted. Preemptible nodes tap unused NVIDIA accelerated capacity, can be taken back when that capacity is needed elsewhere, and are labeled together.ai/compute-class=preemptible so teams can keep coordinators, login pods, and serving replicas on standard capacity.
The interesting part is the shutdown behavior. When a node is reclaimed, Together cordons it, sends a TogetherPreempted event, and delivers SIGTERM to the pods. Workloads then get up to five minutes, via terminationGracePeriodSeconds, to checkpoint and exit before the node is removed. The cluster keeps its preemptible target and refills toward it automatically as capacity opens up, so nobody has to keep asking for replacement nodes.
Together is aiming this squarely at jobs that can take a hit and keep going: short experiments, ablations, config sweeps, quick fine-tunes, evals, distillation, batch inference, and temporary inference bursts. The company says the preview is already being used internally for research, dev/test, and some non-critical production work. Multi-day runs without checkpointing and strict-SLO serving are explicitly not the fit.
Setup is simple enough. Add a preemptible target to a new or existing cluster through the console, CLI, or API, then schedule eligible workloads onto the labeled nodes. The API exposes both the requested target, desired_preemptible_gpus, and the capacity currently live, allocated_preemptible_gpus. Together says Slurm support, more regions, and in-place conversion between compute types are next.
My take — AI-written commentary, not fact-checked reporting
This is the kind of pricing that makes sense only if the system is built to lose on purpose. Half-price GPU time is tempting, but the real product here is operational discipline: checkpoint or get kicked in the teeth. The industry keeps rediscovering that cheap compute is never really cheap if the workload can’t survive a gentle shove.
Read more about this at: Together AI