Launch HN: Expanse (YC P26) – Optimizing Idle GPU Capacity

Data center GPU clusters have a problem where users tend to over-request more resources than necessary to avoid job failures, resulting in actual efficiency staying at around 30-40%. This causes massive cost waste in large-scale computing environments.

Expanse is a solution that analyzes source code, job submission scripts, and hardware telemetry to predict the actual resources needed before job execution. It integrates with SLURM and Kubernetes, and provides the following key features.

• Resource Prediction: Predicts GPU VRAM, utilization, CPU, execution time, etc., and provides confidence intervals. It also warns in advance of potential failures such as OOM (Out of Memory) and suggests code-level optimization measures.
• Real-time Observability: Supports real-time monitoring by visualizing hardware status and code stack profiling on a dashboard with low overhead.
• Failure Diagnosis: When a job fails, it correlates hardware data with stack profiling to provide specific solutions at the code line level, beyond simple logs.

Expanse's model performs about 8x better at prediction than general-purpose LLMs (GPT, Claude, Gemini, etc.), and through cluster-specific models, accuracy improves further as usage increases.