Use Cases › Eliminating Idle Waste
A 30-person AI startup serving the autonomous-vehicle market replaced a 100% cloud-GPU footprint with eight on-prem A100s + opportunistic AWS burst. Cloud spend fell 70% in the first quarter. Training capacity doubled.
| Cloud spend | −70% |
| Training capacity | 2× |
| Time to first job | 14 days → 2 days |
| Architecture | Hybrid Cloud |
By Q3 2025, the company's monthly AWS GPU bill had reached $42,000 — about 18% of operating expenses. Engineers spent days waiting for spot capacity in their preferred region. Long training runs were either paused (when spots evicted) or expensively rebooked on On-Demand. The board wanted a 50% cost cut without slowing the model team.
CapEx was constrained: they could buy a small on-prem cluster, but not nearly enough to replace cloud entirely. The team needed a way to use both — without the manual scheduling toil that had killed previous "hybrid" attempts.
CNLab installed in two days against a freshly-procured 8×A100 server. The CLI, Jupyter, and VS Code integrations replaced their previous custom scripts. Hybrid Cloud burst-out was configured against their existing AWS account in five minutes; spot vs on-demand selection was tuned per workload class.
┌───────────────── ON-PREM (Seoul HQ) ─────────────────┐
│ 8×A100 80GB · 8 TB shared NVMe · 100 Gb │
│ CNLab Scheduler · MIG + 1% Block · Audit │
└────────────────────────────┬──────────────────────────┘
│ WireGuard
│
┌──────────────▼──────────────┐
│ AWS ap-northeast-2 │
│ EC2 G5/P5 Spot + On-Demand│
│ Auto-provision + return │
└─────────────────────────────┘
"We thought hybrid was going to be hard. We had stories about how the last team tried it. CNLab made it the default in three days, and our researchers never noticed the boundary. The cloud bill just stopped growing."
— Head of Infrastructure, computer-vision startup (Series B, autonomous mobility)