On-Demand vs Spot vs Reserved GPU Pricing Explained
Cloud GPUs are available in three pricing tiers: on-demand (pay full price, guaranteed availability), spot/preemptible (60-90% cheaper but can be interrupted), and reserved/committed use (20-40% cheaper than on-demand with a 1-3 year commitment). Signwl tracks all three tiers and presents a blended average that reflects the true cost of cloud GPU compute.
On-Demand Pricing
On-demand instances are the standard pricing tier — you pay the listed hourly rate and the GPU is yours for as long as you need it. There's no commitment and no risk of interruption.
On-demand is the most expensive tier but offers maximum flexibility. It's best for production workloads that need guaranteed availability, development environments, and short-term projects where the overhead of managing spot instances isn't worthwhile.
Spot / Preemptible Pricing
Spot instances (called preemptible on some platforms) offer the same GPU hardware at dramatically reduced prices — typically 60-90% below on-demand rates. The trade-off is that the cloud provider can reclaim your instance with short notice (usually 30 seconds to 2 minutes) when demand is high.
Spot pricing is ideal for fault-tolerant workloads like training with checkpointing, batch inference, and experimentation. The savings are substantial — an H100 that costs $8-12/hr on-demand may be available for $2-4/hr as a spot instance.
Signwl tracks spot discounts as a proxy for GPU utilisation and capacity. High spot discounts indicate excess capacity; low discounts indicate tight supply.
Reserved / Committed Use Pricing
Reserved instances offer a fixed discount (typically 20-40% below on-demand) in exchange for a commitment — usually 1 or 3 years. Some providers offer flexible commitments where you commit to a spend level rather than specific instances.
Reserved pricing is best for stable, long-running production workloads where GPU demand is predictable. The commitment reduces cost compared to on-demand while guaranteeing availability — something spot instances can't offer.
Why Signwl Uses Blended Pricing
Signwl presents a blended average across all three pricing tiers because real-world GPU costs are a mix. Most organisations use a combination — reserved for baseline production workloads, on-demand for burst capacity, and spot for training and batch jobs.
The blended price gives a more accurate picture of what organisations actually pay for cloud GPU compute than any single tier alone.
Frequently Asked Questions
What is spot pricing for GPUs?
Spot pricing offers cloud GPUs at 60-90% below on-demand rates in exchange for the risk of interruption. The cloud provider can reclaim spot instances when demand is high. It's ideal for training workloads with checkpointing and batch inference.
Should I use spot or on-demand for AI training?
Spot for training if you can checkpoint regularly (most modern frameworks support this). The 60-90% savings are substantial. Use on-demand for production inference that can't tolerate interruptions.
How much can I save with reserved GPU instances?
Reserved commitments typically save 20-40% vs on-demand pricing. A 3-year commitment offers the largest discount. This works best for stable, predictable GPU workloads.
Related Accelerators
Continue Learning
Explore Signwl's GPU Data
Live pricing, regional analysis, and comparisons for 39 GPU and AI accelerator types.