What is a spot GPU, and is it worth it?

Is spot cheaper than on-demand?

Yes, by about half. According to GPUwithCrypto's price check of 29 October 2026, the median spot discount on an H100 SXM across those five providers was 56%. Spot at Nebius and Jarvislabs, and the cheapest Vast.ai marketplace offer, were all below our on-demand H100 price. Our price rule compares only with on-demand prices: Each GPU is priced at least 15% under a named reference: the on-demand price of the same GPU at a self-serve provider, read on 29 October 2026 (H100: Massed Compute at $2.73 per GPU-hour; each GPU's reference is named on its price page). Spot, interruptible, marketplace and contact-sales prices are not used as references, and some of them are lower. We do not claim to be the cheapest for every GPU. Every on-demand H100 price we read is on H100 rental prices.

How is spot pricing set?

By spare capacity, so it moves. Nebius says the highest rate of a preemptible VM is always one cent below the current on-demand rate; its table shows only the "from" price. Google Cloud labels its spot column "Current Spot pricing". Salad sells priority tiers instead: on 29 October 2026 its lowest (batch) priority RTX 4090 cost $0.16 per hour against $0.33 for high priority, which is never preempted.

When is a spot GPU worth it?

When losing the GPU costs you minutes, not a deadline. At a 56% discount you pay 0.44 of the on-demand rate, so spot stays cheaper until the work you redo after interruptions exceeds 127% of the useful work (1 ÷ 0.44 = 2.27). The real risk is time, not money.

Good fits: batch inference over a dataset, hyperparameter sweeps, training that checkpoints every few minutes, rendering queues.

Poor fits: an inference API that must stay up, interactive notebooks and ComfyUI sessions, long jobs without checkpoints, and multi-GPU jobs, where losing one GPU stops all of them.

How do I make a training job survive a spot interruption?

Save checkpoints often, keep them off the instance, and resume from the last one. With Hugging Face's Trainer, set a save interval and a cap on how many checkpoints to keep, then call train(resume_from_checkpoint=True): it loads the last checkpoint in the output folder, with optimizer and scheduler states. Copy the output folder to storage that outlives the instance, and make each step safe to run twice.

from transformers import Trainer, TrainingArguments

args = TrainingArguments(
    output_dir="run",
    save_strategy="steps",
    save_steps=200,        # a checkpoint every 200 steps
    save_total_limit=3,    # keep the 3 most recent
)
trainer = Trainer(model=model, args=args, train_dataset=train_ds)

# after a restart: continue from the last checkpoint in ./run
trainer.train(resume_from_checkpoint=True)

Why are GPUwithCrypto instances on-demand only?

So a long job or a live endpoint is never cut by someone else. Instances are on-demand, not spot: nobody can take your GPU back while your balance covers it, and the instance runs until you stop it. When the GPU-hours of a GPU run out, its instances stop and their disks are kept; a new block of hours lets you start them again. Each GPU is priced at least 15% under a named on-demand reference read on 29 October 2026. For the H100, that reference is Massed Compute at $2.73, and our price is 44% below the median of 16 providers. All prices: pricing.