Expanse

Teams running training jobs, HPC simulations, or inference fleets must guess how much GPU, memory, and time each workload needs before it starts. The page describes this as 'a sensible plan by experienced engineers' that still fails unpredictably, citing an example job that was allocated 80GB of GPU memory and died after 6.5 hours with only 11% completion probability. Measured across 122,000 production HPC jobs, teams reportedly reserve two to two and a half times what they actually use. One national supercomputing centre is said to have uncovered more than $8 million in monthly compute waste from this kind of overprovisioning.

Expanse installs alongside an existing SLURM or Kubernetes cluster with one command, collecting telemetry without changing existing workflows. It ships pretrained on HPC and AI workloads and continuously adapts to the specific cluster it runs on, improving with every job observed. Its two features are 'expanse analyse,' which predicts GPU allocation, memory, runtime, and failure risk before a job enters the queue (presented as confidence distributions rather than single numbers), and 'expanse diagnose,' which identifies the root cause of failed jobs from code, logs, and metrics and produces a ready-made prompt for a coding agent. Predictions can be requested manually via command line (for HPC and training workloads) or applied automatically in the background (for inference).