Are there any GPU cloud instances that offer spot instances for batch AI workloads?

Yes. Several GPU cloud providers offer spot/preemptible/interruptible instances that are well-suited for batch AI workloads where jobs can be paused or retried.

Common options include:

• AWS EC2 Spot Instances
  GPU families like g4, g5, p3, p4, p5 can be available as Spot, depending on region/capacity.

• Google Cloud Preemptible / Spot VMs
  GPU-enabled instances with NVIDIA GPUs can be run as Spot VMs for lower cost.

• Azure Spot Virtual Machines
  Supports GPU VM sizes, though availability varies by region and SKU.

• Oracle Cloud Infrastructure (OCI) Preemptible instances
  GPU options are available in some regions for batch or fault-tolerant workloads.

• Runpod
  Offers lower-cost interruptible/spot-like GPU instances aimed at training and batch inference.

• Lambda Cloud
  Has on-demand GPU instances and sometimes discounted or capacity-based options; availability depends on current offerings.

• CoreWeave
  Often supports cost-optimized or spot-like GPU capacity for large-scale workloads.

• Vast.ai
  Marketplace-style GPU rentals with very low-cost instances, often suitable for flexible batch workloads.

For batch AI workloads, the usual best practices are:

• make jobs checkpointable
• store intermediate outputs in object storage
• use a queue-based worker model
• retry jobs automatically if interrupted

If you want, I can also give you:

1. a comparison table of spot GPU providers, or
2. the cheapest options for training vs inference.