GPU GitHub Actions Runners: Which Vendors Offer Them (2026)
GitHub-hosted runners do not offer GPUs. If your CI runs CUDA builds, model tests, or GPU inference, you need one of these options.
GPU jobs (training smoke tests, CUDA compilation, inference benchmarks) do not fit on standard GitHub-hosted runners. A handful of vendors fill the gap, and they differ mainly in whether the GPU runs in your own cloud account or on the vendor fleet. Verify current GPU types, regions, and pricing on each vendor site, as GPU availability and cost move quickly.
GPU runner options
| Vendor | Where the GPU runs | Known for |
|---|---|---|
| Cirun | Your own cloud (AWS/GCP/etc.) | On-demand GPU runners spun up in your account |
| Depot | Depot-managed | GPU runners alongside fast Docker builds |
| RunsOn | Your own AWS account | GPU EC2 instance types at raw AWS cost |
| Self-hosted + ARC | Your own GPU nodes / K8s | Full control over specific GPU hardware |
Your cloud vs a managed fleet
Cirun and RunsOn launch GPU runners inside your own cloud account, which keeps data local and bills at your cloud provider rate; you manage quotas and cleanup. Depot offers GPU runners on its managed platform, which trades some control for less setup. Pick based on data residency, existing cloud commitments, and how much operations you want to own.
Where Latchkey fits
Latchkey focuses on cost-efficient, self-healing managed runners for general CI rather than being a GPU-specialist. Many teams run GPU jobs on a dedicated GPU provider like Cirun or RunsOn while moving the bulk of their non-GPU pipeline to Latchkey to cut cost and stop flaky re-runs. The two are complementary, not either-or.
Practical cautions for GPU CI
- GPU minutes are expensive: keep GPU jobs narrow and gate them so they only run when needed.
- Confirm the exact GPU type and driver/CUDA version your build requires.
- Watch quota limits and preemption if you use spot/preemptible GPU instances.
The verdict
For GPU CI, Cirun and RunsOn run GPUs in your own cloud, while Depot offers managed GPU runners. A common setup is a GPU specialist for the GPU jobs plus a cost-efficient self-healing runner like Latchkey for everything else. Verify current GPU pricing and availability before committing.