New in Together GPU Clusters: Autoscaling, observability, and self-healing

Together AI is done pretending that GPU clusters are still a science experiment. The company's update to its GPU Clusters product (previously branded Instant Clusters) reads like an admission that customers stopped treating these things as one-off training runs a while ago and started running them like real infrastructure — the kind that needs uptime guarantees, permission systems, and someone to page when a node dies at 3am.

The headline addition is autoscaling, built on the Kubernetes Cluster Autoscaler. Instead of provisioning for peak load and eating the cost of idle GPUs during quiet periods, clusters now expand when pending pods pile up from a big training job or a traffic spike, then shrink back down once demand fades. It's not a novel idea in cloud computing generally, but for bare-metal-flavored GPU rental, it's a meaningful shift — you get elasticity without giving up the raw performance that made bare metal attractive in the first place.

Hardware failure is the other big theme, and for good reason. Anyone who's run a multi-hundred-GPU training job knows one flaky node can torch an entire run. Together's answer is a set of self-serve active health checks — everything from DCGM Diag 3 to multi-node NCCL and InfiniBand bandwidth tests — that users can trigger from the UI before committing to a large job. If something does fail mid-run, a three-click self-repair now cordons, drains, and rebuilds the node automatically, with acceptance tests gating whether a cluster is even marked