Beta
PricingReading · ~3 min · 61 words deep

Spot Pricing

Spot pricing sells preemptible GPU capacity at 60-90% off on-demand · you can be evicted with 2 minutes notice but prices are dramatically lower.

TL;DR

Spot pricing sells preemptible GPU capacity at 60-90% off on-demand · you can be evicted with 2 minutes notice but prices are dramatically lower.

Level 1

AWS EC2 Spot, GCP Preemptible/Spot, Azure Spot VMs sell excess capacity cheaply. A $32/hour A100 on-demand might be $4/hour on spot. Workloads that can checkpoint and resume (training, batch inference) use spot aggressively; interactive workloads cannot. H100 spot pricing varies wildly · $6-40/hour depending on region and demand.

Level 2

Spot availability fluctuates · when a cloud provider's on-demand demand spikes, spot instances are reclaimed. Eviction notice is typically 30 seconds to 2 minutes. Training frameworks (PyTorch Lightning, DeepSpeed, HF Accelerate) support spot-friendly checkpointing. For inference, spot is used for batch endpoints that can tolerate partial completion. Runpod, Vast.ai, and Lambda Cloud run spot-style markets · often cheaper than hyperscaler spot.

Level 3

Spot pricing follows market dynamics · real-time auction with provider-set floor. H100 spot went from $80/hour (2023 crunch) to $2/hour (2025 glut) in some regions. Workload architecture for spot: job queues with automatic reschedule, small-step checkpoints (every 5-15 min), and spot-diversity (bid across multiple instance types + zones). Enterprises running LLM training burn 60-80% of budget on spot.

The takeaway for you
If you are a
Researcher
  • ·60-90% discount vs on-demand
  • ·30 sec to 2 min eviction notice
  • ·Market-priced · fluctuates with demand
If you are a
Builder
  • ·Best for training · requires frequent checkpointing
  • ·Bad for serving · eviction hurts latency
  • ·Use spot across zones + instance types to diversify
If you are a
Investor
  • ·Spot market reflects real compute scarcity signals
  • ·GPU glut (2025) dropped H100 spot <$2/hour · now recovering
  • ·Key input for AI unit economics · training costs 70%+ on spot
If you are a
Curious · Normie
  • ·Cheap AI computer time but you can get kicked off
  • ·Like standby airline tickets · discount for uncertainty
  • ·Used by AI companies for training, not for running apps
Gecko's take

Spot is how AI actually gets trained in 2026. Cap-table-level cost advantage for teams that can checkpoint.

60-90% vs on-demand. H100 spot ranges $2-40/hour depending on region and cloud.