On September 10, 2026, RunPod officially notified all active users that Spot and Interruptible Pods are being removed from the platform and its API effective Friday, September 11, 2026. Any spot workloads still running on that day will be abruptly stopped.
The Immediate Economic Impact
For machine learning engineers, fine-tuning specialists, and ComfyUI production teams, this change is not merely an API tweak—it is a 2.5x to 3.2x price surge overnight.
| GPU Type | RunPod Spot (Deprecated) | RunPod On-Demand | Vast.ai Spot (Current) |
|---|---|---|---|
| NVIDIA RTX 4090 (24GB) | $0.34 / hr | $0.74 / hr | $0.22 – $0.32 / hr |
| NVIDIA A100 SXM (80GB) | $1.19 / hr | $2.49 / hr | $1.15 – $1.45 / hr |
| NVIDIA H100 PCIe (80GB) | $2.19 / hr | $4.69 / hr | $2.10 – $2.60 / hr |
Running a continuous 4x RTX 4090 cluster for diffusion LoRA training went from roughly $980/month to nearly $2,150/month on on-demand rates.
Why You Shouldn't Just Surrender to On-Demand
RunPod's recommended paths in their advisory are predictable: either pay full price for On-Demand or rewrite your stack into Serverless. But Serverless is ill-suited for persistent model weights, iterative ComfyUI pipelines, and 8-hour continuous LoRA epochs.
The real alternative is straightforward: Vast.ai still offers wholesale spot/interruptible GPU pricing at scale, with even larger global inventory than RunPod ever maintained. The only historical hurdle was eviction fragility.
The Solution: Vast.ai Spot + SpotWarp Failover
By pairing wholesale Vast.ai spot compute with SpotWarp, your training pipeline gains enterprise-grade resiliency without paying enterprise premiums:
- Continuous Background Delta-Sync: Changes to
/workspaceand active checkpoints are mirrored every 60 seconds with zero runtime GPU overhead. - Sub-Minute Eviction Failover: When a spot host reclaims capacity, SpotWarp's daemon automatically provisions a replacement node on Vast.ai (or RunPod on-demand as fallback) and restores weights in under 45 seconds.
- Hands-Off Training Resumption: Your training script resumes from the exact last saved step automatically.
Deploy Zero-Downtime Spot Compute
Never let host evictions kill your training runs again. Protect your RunPod and Vast.ai instances in under 5 minutes with automated delta-sync.