Expanse

Expanse predicts what a compute job will need before it runs. One command installs it onto an existing SLURM, Kubernetes or Nomad cluster, where it collects telemetry alongside workloads without changing submit scripts or pipelines. Before a job enters the queue it returns predicted GPU allocation, memory use, runtime and completion probability, each presented as a confidence distribution rather than a single number, so infrastructure decisions rest on measured evidence instead of an engineer's estimate. It is aimed at high performance computing centers, AI infrastructure teams and anyone running inference fleets on GPUs they already own. The platform has two halves that feed each other. Analyse predicts memory, runtime and failure risk ahead of execution. Diagnose explains what actually happened afterward from logs, metrics and exit state, and every execution sharpens the next prediction. The models arrive pretrained on real scientific and AI workloads and then adapt to the local environment. On serving fleets the same loop runs with nothing for engineers to invoke: each model is sized to what it genuinely uses, so more models pack onto a GPU without interfering, and the full prediction distribution is exposed through an API for custom schedulers and simulators to consume. No pricing is published. The site carries no pricing page and routes buyers to a discovery call, so cost is quote based. The company is early stage, founded by four people out of Edinburgh and Imperial with prior roles in quantitative finance and trading firms. Its headline evidence is a national supercomputing center where Expanse says it uncovered more than $8 million in monthly compute waste.