Live2,392 GPUs available across 3 regions
Cheap cloud GPUs at spot prices, billed by the minute.
Same hardware as on-demand, 30–53% cheaper. A 2-minute notice before an interruption, disks that survive it, an instance in under 60 seconds.
Every GPU, at least 20% below the cheapest rate we found.
Spot and on-demand run on the same nodes. Spot is capacity we may reclaim with a 2-minute notice; how often that happened over the last 30 days is in the table.
- SpotGPUs spot $0.55−20% below the lowest spot price elsewhere
- Lowest spot elsewhere $0.69
- SpotGPUs on-demand $0.89−23% below the lowest on-demand price elsewhere
- Lowest on-demand elsewhere $1.15
- Median on-demand, 87 providers $2.69
Same rule for all 47 models. How we check prices
| GPU | Spot | On-demand | Save | Reclaim rate 30 d | Available | Action |
|---|---|---|---|---|---|---|
| Data center· 8 of 21 | ||||||
| B200 SXMNew180 GB HBM3e · NVLink | $1.69/h | $2.49/h | −32% | 5–10% | 16 GPUs | Deploy B200 SXM |
| H200 SXM141 GB HBM3e · NVLink | $0.75/h | $1.09/h | −31% | 5–10% | 32 GPUs | Deploy H200 SXM |
| H200 NVL141 GB HBM3e · NVLink | $0.69/h | $0.99/h | −30% | 5–10% | 16 GPUs | Deploy H200 NVL |
| H100 SXMMost popular80 GB HBM3 · NVLink | $0.55/h | $0.89/h | −38% | <5% | 96 GPUs | Deploy H100 SXM |
| H100 NVL94 GB HBM3 · NVLink | $0.49/h | $0.79/h | −38% | <5% | 32 GPUs | Deploy H100 NVL |
| H100 PCIe80 GB HBM2e | $0.45/h | $0.69/h | −35% | <5% | 96 GPUs | Deploy H100 PCIe |
| GH20096 GB HBM3 · NVLink | $0.49/h | $0.79/h | −38% | 5–10% | 16 GPUs | Deploy GH200 |
| A100 SXM 80GB80 GB HBM2e · NVLink | $0.19/h | $0.29/h | −34% | <5% | 96 GPUs | Deploy A100 SXM 80GB |
| Workstation· 3 of 12 | ||||||
| RTX PRO 6000 BlackwellNew96 GB GDDR7 | $0.19/h | $0.39/h | −51% | 5–10% | 48 GPUs | Deploy RTX PRO 6000 Blackwell |
| RTX PRO 5000 BlackwellNew48 GB GDDR7 | $0.15/h | $0.29/h | −48% | 5–10% | 32 GPUs | Deploy RTX PRO 5000 Blackwell |
| RTX PRO 4500 Blackwell32 GB GDDR7 | $0.09/h | $0.19/h | −53% | 5–10% | 16 GPUs | Deploy RTX PRO 4500 Blackwell |
| Consumer· 3 of 14 | ||||||
| RTX 5090New32 GB GDDR7 | $0.09/h | $0.19/h | −53% | 10–15% | 128 GPUs | Deploy RTX 5090 |
| RTX 508016 GB GDDR7 | $0.06/h | $0.09/h | −33% | 10–15% | 64 GPUs | Deploy RTX 5080 |
| RTX 5070 Ti16 GB GDDR7 | $0.06/h | $0.09/h | −33% | 10–15% | 24 GPUs | Deploy RTX 5070 Ti |
- USD per GPU-hour, billed by the minute
- Multi-GPU nodes priced per GPU
- Persistent storage $0.08 per GB-month
- No egress fees, no charge for boot or reclaim
- Prepaid in crypto, no card, no identity check
Interruptions are a mechanism, not a surprise.
Spot capacity is priced lower because we can take it back. Here is exactly what happens when we do, and what stays yours.
- 2 minutes of notice, three waysThe reclaim notice is published on the instance metadata endpoint, sent to your webhook and shown in the console, 2 minutes before the stop. Poll it every few seconds and checkpoint.
- Stopped, never deletedThe persistent disk, its data, the template and the reserved IP stay attached to the stopped instance. Nothing on the disk is lost.
- Resume on spot, or switch to on-demandRelaunch the same disk on the next available spot capacity, or move it to an on-demand instance that is never reclaimed. One click or one API call, same image.
- Billed only while runningPer-minute billing stops at the reclaim event. Boot time, the notice window and stopped time are not billed for compute; the disk is billed per GB-month.
- Instance runningTraining job, checkpoint every 10 min to the persistent disk.T−12:00
- Reclaim notice publishedMetadata endpoint, webhook and console. Your handler flushes a final checkpoint.T−02:00
- Instance stoppedCompute billing ends. Disk, IP and template stay attached.T−00:00
- Auto-relaunch on next spot capacityOptional. Same disk, same image; the job resumes from the last checkpoint.T+04:10
- Or: switch to on-demandSame disk on a never-reclaimed instance, when the deadline matters more than the price.any time
Good fit for spot
Training and fine-tuning with checkpoints, batch and offline inference, embedding and evaluation runs, rendering and simulation, hyperparameter sweeps, data processing, CI and test pipelines.
Use on-demand instead
Latency-sensitive production APIs, interactive sessions you can't afford to lose, and anything with a hard deadline in the next few hours. Same GPUs, never reclaimed, 38% more on an H100.
Everything an instance needs. Nothing it doesn't.
A GPU cloud built around the workloads people actually run on spot capacity: fast to start, safe to lose, cheap to keep.
Running in under 60 seconds
Pick a GPU and a template, get an SSH prompt and a Jupyter URL. Images are pre-pulled on every node, so boot time is measured in seconds and not billed.
Dedicated GPUs, KVM isolation
Each GPU is passed through to your own virtual machine with a dedicated share of CPU, RAM and local NVMe. No shared GPUs, no containers on someone else's kernel.
Persistent disks that follow you
Network disks survive stops, reclaims and instance type changes. Attach the same disk to spot today and on-demand tomorrow.
Per-minute billing with hard caps
Every started minute is billed at 1/60 of the hourly price, nothing else. Set a monthly spend cap and we stop instances before you cross it.
1 to 8 GPUs per node
Single GPUs for experiments, NVLink nodes with up to 8 GPUs for training. Same price per GPU whatever the node size.
Templates or your own container
PyTorch, CUDA, vLLM, ComfyUI and JupyterLab templates, or bring any OCI image from a public or private registry.
One CLI, one API, the same objects in the console.
Launch, watch the reclaim notice and resume from your own scripts. Everything the console does is a documented endpoint.
- CLI and Python SDKLaunch with a max spot price, attach disks, stream logs, stop. Scriptable in a Makefile or a scheduler.
- REST API with webhooksInstance lifecycle, reclaim notices and billing events as JSON. Token-scoped keys per project.
- Reclaim notice on the metadata endpointPoll
/v1/noticefrom inside the instance every few seconds and checkpoint when it flips.
$ spot launch --gpu h100 --tier spot --max-price 0.70 --disk train-data Instance i-7f3a2c H100 80GB us-east spot @ $0.55/h ssh ubuntu@i-7f3a2c.spotgpus.com · https://i-7f3a2c.spotgpus.com/lab ✓ reachable in 41s $ spot watch i-7f3a2c --on-notice ./checkpoint.sh 18:04:12 running $0.23 so far 18:12:40 notice reclaim in 2:00 → ./checkpoint.sh 18:14:40 stopped disk train-data kept, IP kept 18:18:50 relaunched i-9b10e4 spot @ $0.55/h resumed from ckpt-0420
3 regions, Tier III facilities, dedicated hardware.
Spot and on-demand run on the same nodes in the same facilities. Pick a region for data locality; prices are identical across regions.
Isolation by design
One tenant per virtual machine, GPU passthrough, no shared kernels. Instances live in a private network per project.
Encrypted at rest and in transit
Persistent disks are encrypted at rest; SSH and the API are the only entry points, both over TLS.
Public status page
Per-region health, reclaim rates and incident history, updated by the system that runs the platform, not by hand.
View statusEngineers on support
Tickets are answered by the people who operate the nodes, around the clock. No tiered queue, no bots.
Contact supportQuestions people ask before their first spot job.
How much does an H100 cost per hour?
On SpotGPUs, an H100 SXM is $0.55 per GPU-hour on spot and $0.89 on-demand, billed per minute. Across the 68 providers in our September 2026 index, published on-demand prices run from $1.15 to $11.06 with a median of $2.50 — <a href="/guides/h100-price-per-hour">the full comparison, provider by provider</a>.
How do I pay? Do you take cards?
Crypto only: USDT, USDT, Bitcoin, Ethereum, Solana, Litecoin, Monero, TRON. You add credit from $30 to a prepaid balance, the balance pays for compute by the minute, and nothing can be charged beyond it. No card, no bank transfer, no identity check. <a href="/pay-with-crypto">Currencies, networks and confirmation times</a>.
What is a spot instance, exactly?
A spot instance is the same GPU, the same virtual machine and the same network as our on-demand tier, sold from capacity that is not currently reserved. Because we can reclaim that capacity, it is priced 30–53% below on-demand. When we do reclaim it, you get a 2-minute notice, the instance is stopped, and your disk is kept.
What happens when my spot instance is interrupted?
You receive the notice 2 minutes ahead through the instance metadata endpoint, a webhook and the console. The instance is then stopped, not deleted: the persistent disk, its data and the IP reservation stay attached. You can relaunch on the next available spot capacity or switch the same disk to an on-demand instance in one click. Compute is billed only while the instance is running.
How often are spot instances actually reclaimed?
It depends on the GPU and the region. We publish the trailing 30-day reclaim rate for every model in the price table: most data-center models sit below 5%, meaning fewer than one interruption per twenty instance-days. Consumer cards in high demand run higher, and the table says so.
How does billing work?
Per minute, from the moment the instance is reachable until you stop it. Every started minute is billed at the hourly price divided by 60; there is no minimum duration and no rounding to the hour. Boot time and reclaim events are not billed. Persistent storage is billed per GB-month while the disk exists, including while the instance is stopped. There are no egress fees.
Which workloads are a good fit for spot?
Anything that can checkpoint or retry: training and fine-tuning with periodic checkpoints, batch inference, embedding jobs, rendering, hyperparameter sweeps, data processing and CI. Latency-sensitive production APIs should run on the on-demand tier, which is never reclaimed.
Can I mix spot and on-demand?
Yes. Both tiers use the same images, disks, templates and API. A common pattern is a small on-demand baseline for serving with spot capacity for everything that can wait a few minutes.
What do I get with an instance?
A dedicated GPU passed through to a KVM virtual machine, a dedicated share of CPU and RAM, local NVMe scratch, an optional persistent network disk, SSH and Jupyter access, and a choice of templates such as PyTorch, CUDA, vLLM and ComfyUI. You can also bring your own container image.
Is there a minimum spend or a contract?
No. You add credit (from $30), launch and stop whenever you want. Reserved capacity with a fixed price is available for teams that need guaranteed GPUs for a month or longer.
Your first H100 for $0.55 an hour, running in under a minute.
Prepaid in crypto, pay as you go: no card, no contract, no minimum commitment. Add credit, launch, stop whenever you want.
Need guaranteed capacity for a month or more? Reserved capacity has a fixed price and is never reclaimed.
