An 8×H100 node on AWS costs $98.32/hour on demand. The same node on spot costs $57.76/hour. That's $40 an hour - about $29,000 a month - for accepting that AWS can take the machine back with two minutes' notice.
Spot pricing is the biggest discount in cloud GPUs and the most misunderstood. The usual advice ("use spot, it's 70% cheaper") is wrong in two directions: the discount varies enormously by provider, and on the cheapest providers it barely exists. CloudMart tracks on-demand and spot prices across 10 GPU providers, so here's what the spread actually looks like.
Real spot discounts, by provider
| Provider | GPU | On-demand /hr | Spot /hr | Discount |
|---|---|---|---|---|
| AWS | L4 ×1 (g6) | $0.80 | $0.10 | 88% |
| AWS | A100 40GB ×8 (p4d) | $32.77 | $13.07 | 60% |
| AWS | H100 80GB ×8 (p5) | $98.32 | $57.76 | 41% |
| CoreWeave | H100 80GB ×1 | $6.16 | $2.46 | 60% |
| CoreWeave | A100 80GB ×1 | $2.70 | $1.21 | 55% |
| RunPod | A100 80GB ×1 | $1.79 | $1.14 | 36% |
| RunPod | H100 80GB ×1 | $2.69 | $1.99 | 26% |
Prices as tracked on October 24, 2026. Spot rates move continuously - the live GPU marketplace has current numbers.
The pattern nobody mentions
Look down the discount column. The deepest cuts are at AWS and CoreWeave. The shallowest is RunPod, at 26%.
That's not a coincidence, and it's the single most useful thing to understand about spot pricing: the discount is a function of how inflated the on-demand price was in the first place. Spot is the market clearing price for idle capacity. Providers whose list price is already close to the market clearing price have little room to discount. Providers with a large enterprise margin have a lot.
So a big discount percentage tells you about the provider's pricing strategy, not about whether you're getting a good deal.
Which is why this comparison matters more
Forget percentages and compare the actual cost per GPU-hour. AWS's p5 spot price of $57.76/hr is for an 8-GPU node, which works out to $7.22 per GPU-hour.
| Option | $ per H100-hour |
|---|---|
| RunPod spot | $1.99 |
| Crusoe on-demand | $2.00 |
| CoreWeave spot | $2.46 |
| RunPod on-demand | $2.69 |
| CoreWeave on-demand | $6.16 |
| AWS spot (p5, per GPU) | $7.22 |
Two things fall out of that table. Crusoe's flat on-demand rate of $2.00/GPU-hr beats every spot price on the board except RunPod's - no reclaim risk, no checkpointing, no two-minute warnings. And AWS spot H100s cost more per GPU-hour than RunPod's on-demand price. You take on interruption risk, two-minute termination notices, and checkpointing complexity - and you still pay a 168% premium over simply renting the GPU outright somewhere else.
If you are already deep in AWS and your data has gravity there, that premium may be worth paying. If you are choosing freely, spot on a hyperscaler is close to the worst option on the board.
When spot is the right call
Spot earns its keep when all three of these are true:
- Your job checkpoints. If losing the machine means losing the run, the discount is an illusion. A job that restarts from scratch after 70% completion costs more than on-demand would have.
- Your job is long. Checkpointing has engineering cost. On a 30-minute job it isn't worth writing; on a four-day training run it absolutely is.
- Your deadline is soft. Spot capacity disappears exactly when demand spikes - which is often when you need it too.
When it isn't
- Inference serving production traffic. A reclaimed node is an outage. Use on-demand, or a provider with per-second billing so idle capacity costs little.
- Short interactive work. Notebook experiments, debugging, a quick fine-tune - the checkpointing overhead exceeds the savings.
- Anything on a deadline. "Cheaper, but it might not finish" is not cheaper if it has to finish.
- When the on-demand price is already low. At RunPod's 26% spread, you're accepting real operational risk for a quarter off. Often not worth it.
The checkpointing math
Rough rule: spot pays off when discount × job length exceeds expected restart cost. If spot is 50% off and your job has a 20% chance of being reclaimed once, you lose on average 10% of the run to redone work - still well ahead. At a 26% discount with a 40% reclaim rate, the margin is thin enough that the engineering time isn't worth it.
Checkpoint every 10-15 minutes of compute. More often and you waste time on I/O; less often and a reclaim costs real money.
Bottom line
- Cheapest H100 overall: RunPod spot at $1.99/GPU-hr.
- Best flat rate, no interruption risk: Crusoe at $2.00/GPU-hr on-demand undercuts most providers' spot prices outright - worth checking before you take on reclaim risk at all.
- Best discount-to-reliability trade: CoreWeave spot at $2.46 - 60% off with InfiniBand interconnect, which matters for multi-node training.
- Don't bother: hyperscaler spot, unless you're locked into the ecosystem. $7.22/GPU-hr with interruption risk is a bad trade.
- Skip spot entirely if your job is short, interactive, or serving users. RunPod or Lambda on-demand will cost less than your engineering time.
Not sure if your job should run on spot?
Tell the Planner your model size, run length, and deadline - it picks the provider and estimates the bill both ways.
Plan my training run →Related: H100 rental prices compared for on-demand rates across every provider we track, and the cheapest way to fine-tune an LLM if you're sizing a specific job.