Solutions › Hybrid Cloud Bursting
My GPU First — utilize on-prem 100% before reaching for cloud. When local capacity fills, auto-burst to AWS, GCP, Azure, Naver, or KT Cloud. Return automatically when the job completes. Spend cloud money on demand, never by default.
A simple operating principle. Run jobs on the GPUs you own — they're already paid for. Reach for cloud only when local utilization crosses a threshold you set. The moment a local slot opens up, the cloud workload returns.
Every job tries on-prem first. Only when no slot fits within the SLO does the scheduler reach for cloud.
For cloud, prefer Spot. Fall back to On-Demand only when spot eviction risk exceeds tolerance for that workload.
A cloud-running workload migrates back the instant a matching local slice frees up. The cloud bill stops the moment your GPU does.
CNLab's scheduler doesn't just react — it predicts. By learning your workload patterns, it pre-provisions cloud capacity 90 seconds before local saturation hits, avoiding any user-visible queue time.
Jobs queue on local clusters. 100% utilization before cloud is touched.
When utilization > 90% sustained, the scheduler arms burst-out.
Spot or On-Demand instance auto-provisioned on the configured cloud.
Pending or running job placed on the cloud node. Live migration if running.
Workload migrates back when local frees up. Cloud node auto-terminated.
Live migration applies between on-prem and cloud the same way it works between two on-prem clusters. The CUDA context, memory pages, and process state move via RDMA-direct transfer (on-prem → on-prem) or encrypted TCP (on-prem ↔ cloud). The user sees nothing.
CNLab tracks per-AZ spot interruption rates. Below 5% interruption: prefer Spot. Above: On-Demand for SLO-bound jobs, Spot for resilient ones.
Cloud nodes terminate the moment training completes. No hourly waste, no orphaned instances, no Slack post-mortems.
For data-heavy training, scheduler picks the cloud region closest to your S3/GCS bucket. Cuts egress costs by 60% vs naive placement.
If you have AWS Reserved or GCP Committed-Use discounts, CNLab routes burst-out to those reservations first. Effective rate beats Spot.
The complete picture: on-prem clusters as the primary pool, predictive burst-out, encrypted egress to AWS / GCP / Azure / Naver / KT, and auto-return when local capacity frees up.
Click the diagram to view full resolution.
Estimate your annual savings from My-GPU-First. Assumes 70% cloud-cost reduction on the workload that today runs cloud-only.
Based on 70% reduction × monthly spend × 12. Real savings depend on workload mix; we'll model your actual numbers in a 30-min consult.
Mark a project, dataset, or model as on-prem-only. Scheduler never bursts those workloads to cloud, regardless of capacity.
All on-prem ↔ cloud traffic goes over TLS 1.3 + WireGuard VPN, with optional cloud-provider PrivateLink/Interconnect.
User-level access control applies in both planes. A user blocked from cloud bursts simply queues on-prem; no policy bypass.
Audit trail (who provisioned what, when, where) lives on the on-prem control plane — never in the cloud, even for cloud-running jobs.
EC2 G5/P5, Spot & On-Demand
A3/G2 Compute, Spot VMs
ND/NC series, Spot VMs
G2/G3 GPU servers
G-Class GPU instances
Free 30-min consult. Tell us your monthly cloud bill and we'll model the on-prem-first save in your inbox within one business day.