Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
59% of compute wasted. Expanse just launched to recover $8.5M monthly per cluster—by predicting what jobs actually need before they run.
Why it matters
A Y Combinator-backed startup is shipping a deep learning tool that solves a massive infrastructure problem: GPU clusters running at 30-40% utilization because researchers over-request resources to avoid job failure. Expanse uses multimodal AI (source code + hardware telemetry) to predict actual resource needs and outperforms frontier LLMs by 8x—directly applicable to AI labs, quant funds, and datacenters burning millions monthly on wasted compute.
The key facts
12 to knowYC P26 launch (Expanse)
Founded by Ismaeel Bashir, Eren Bilge, Yafet Mekonnen, Nikodem Dyzma
59% of compute wasted on measured national-scale HPC cluster (122k jobs)
$8.5M monthly compute waste on single cluster at on-demand rates
Custom model outperformed Gemini 3.5 Pro, Claude Opus, GPT-5.5, Codex 5.3 by 8x
34% better than baseline predictors on EPCC datasets
Hooks into SLURM and Kubernetes schedulers
Predicts GPU VRAM, utilization, memory, CPUs, walltime with confidence intervals
Live observability dashboard with hardware telemetry and stack profiling
Targets clusters with 100+ GPUs
Paid pilot model: two-week measurement window followed by fixed monthly fee per cluster
Founder experience: HPC/GPU workloads at major quant funds and HPC facilities
Go to the source
Hacker Newsnews.ycombinator.com
Publisher excerpt: Hey HN, we’re Ismaeel, Eren, Yafet and Nikodem. We built Expanse (https://expanse.sh/) to increase the effective capacity of your HPC/GPU clusters running schedulers/orchestrators like Kubernetes and SLURM. We read the source code, job submission script, and the hardware a workload is about to run…