Spot GPU prices
Spot capacity (also sold as preemptible or interruptible) is a lower rate for a machine the provider can stop when it needs the capacity back. Each card's cheapest spot rate sits beside the cheapest on-demand rate for the same card, then each provider's spot rate beside its own on-demand rate for the same machine.
In today's data: the cheapest spot H100 is $0.94/GPU-hour at Prime Intellect, 53% less than the cheapest on-demand H100 ($1.99, Voltage Park). Across 44 machines sold both ways, spot costs a median 47% less than the same provider's on-demand rate.
Cheapest spot rate for each GPU
Rates are per GPU-hour. The on-demand rate is the cheapest one a renter can book today for the same card at any provider; node sizes can differ.
| GPU | Spot from | Spot provider | On-demand from | Spot against on-demand | Spot offers | |
|---|---|---|---|---|---|---|
| B300 | $4.08 | Daytona and 3 more | $4.99 Prime Intellect | 18% less | 7 | Rent → |
| B200 | $3.59 | Daytona and 4 more | $3.49 Prime Intellect | 3% more | 5 | Rent → |
| H200 (SXM) | $2.61 | Daytona and 3 more | $3.14 Prime Intellect | 17% less | 4 | Rent → |
| H100 NVL | $1.29 | Microsoft Azure | $2.92 Massed Compute | 56% less | 2 | Rent → |
| H100 | $0.94 | Prime Intellect and 6 more | $1.99 Voltage Park | 53% less | 11 | Rent → |
| H100 SXM | $1.91 | TensorDock and 2 more | $2.25 TensorDock | 15% less | 4 | Rent → |
| A100 SXM 80GB | $1.06 | Microsoft Azure and 1 more | $1.38 Massed Compute | 24% less | 3 | Rent → |
| A100 PCIe 80GB | $0.68 | Microsoft Azure | $1.50 TensorDock | 55% less | 3 | Rent → |
| A100 | $1.19 | Spheron | $1.09 Thunder Compute | 9% more | 1 | Rent → |
| A100 SXM 40GB | $0.74 | Microsoft Azure and 2 more | $0.89 Jarvislabs | 17% less | 4 | Rent → |
| MI355X | $3.44 | Daytona and 1 more | $5.99 Daytona | 43% less | 2 | Rent → |
| MI325X | $2.00 | Vultr Out of stock | $3.80 DigitalOcean | 47% less | 1 | |
| L40S | $1.05 | Spheron | $0.96 Spheron | 9% more | 1 | Rent → |
| L4 | $0.10 | AWS and 1 more | $0.44 Jarvislabs | 76% less | 3 | Rent → |
| Tesla V100 | $1.40 | Google Cloud | $0.23 Verda (DataCrunch) | 6.1× as much | 1 | Rent → |
| Tesla T4 | $0.21 | Google Cloud | $0.35 Google Cloud | 40% less | 1 | Rent → |
| RTX 6000 Ada | $1.04 | Lium | $0.75 TensorDock | 39% more | 1 | Rent → |
| RTX PRO 6000 | $0.45 | Google Cloud and 2 more | $1.10 Google Cloud | 59% less | 3 | Rent → |
| RTX 5090 | $0.74 | Daytona | $0.35 HyperAI | 2.1× as much | 1 | Rent → |
| RTX 4090 | $0.57 | Daytona | $0.35 TensorDock | 63% more | 1 | Rent → |
Cite this data
GPU Rental Prices, "Spot GPU prices: interruptible rates against on-demand", prices checked 22 October 2026, https://gpurentalprices.com/spot
Quote it with a link to this page. Downloads of the whole dataset are on Data & API.
Same machine, spot against on-demand
Each row matches one provider's spot and on-demand rates for the same GPU variant, GPU count, instance and region where published. Largest saving first; a negative saving means spot costs more.
The saving is also the share of paid hours a job can lose to interruptions and restarts before spot costs more than on-demand: at 40% less, up to 40 of every 100 paid hours can go to work that is lost and run again. Checkpoints decide that share: the more often a job saves its state, the less work an interruption takes with it.
| Provider / GPU | Region | On-demand $/GPU-hour | Spot $/GPU-hour | Saving |
|---|---|---|---|---|
| AWS · L4 · 1 GPU | us-east-1 | $0.80 | $0.10 | 87.0% |
| AWS · L4 · 8 GPUs | us-east-1 | $1.67 | $0.22 | 86.8% |
| Microsoft Azure · A100 PCIe 80GB · 1 GPU | eastus | $3.67 | $0.68 | 81.5% |
| Microsoft Azure · H100 SXM · 8 GPUs | eastus | $11.06 | $2.04 | 81.5% |
| Microsoft Azure · H100 NVL · 1 GPU | eastus | $6.98 | $1.29 | 81.5% |
| Microsoft Azure · H100 NVL · 2 GPUs | eastus | $6.98 | $1.29 | 81.5% |
| Microsoft Azure · H100 SXM · 8 GPUs | eastus | $12.29 | $2.27 | 81.5% |
| Microsoft Azure · A100 PCIe 80GB · 4 GPUs | eastus | $3.67 | $0.68 | 81.5% |
| Microsoft Azure · A100 PCIe 80GB · 2 GPUs | eastus | $3.67 | $0.68 | 81.5% |
| Microsoft Azure · A100 SXM 40GB · 8 GPUs | eastus | $3.40 | $0.74 | 78.3% |
Show the other 34 matched machines
| Provider / GPU | Region | On-demand $/GPU-hour | Spot $/GPU-hour | Saving |
|---|---|---|---|---|
| Microsoft Azure · A100 SXM 40GB · 8 GPUs | eastus | $3.40 | $0.75 | 78.0% |
| Microsoft Azure · A100 SXM 80GB · 8 GPUs | eastus | $4.10 | $1.06 | 74.2% |
| Microsoft Azure · A100 SXM 80GB · 8 GPUs | eastus | $4.10 | $1.06 | 74.2% |
| Google Cloud · B200 | us-central1 | $16.11 | $4.95 | 69.2% |
| Prime Intellect · H100 | Not stated | $2.43 | $0.94 | 61.3% |
| Google Cloud · RTX PRO 6000 | us-central1 | $1.10 | $0.45 | 58.9% |
| Spheron · RTX PRO 6000 | Not stated | $2.35 | $1.11 | 52.8% |
| Together AI · H100 | Not stated | $3.99 | $1.99 | 50.1% |
| Together AI · H200 (SXM) | Not stated | $5.99 | $2.99 | 50.1% |
| Together AI · B200 | Not stated | $8.19 | $4.09 | 50.1% |
| Together AI · B300 | Not stated | $9.99 | $4.99 | 50.1% |
| Spheron · B200 | Not stated | $11.12 | $5.71 | 48.7% |
| Google Cloud · H100 | us-central1 | $9.80 | $5.31 | 45.8% |
| Google Cloud · Tesla V100 | us-central1 | $2.48 | $1.40 | 43.7% |
| Daytona · RTX 5090 | Not stated | $1.29 | $0.74 | 42.6% |
| Daytona · RTX PRO 6000 | Not stated | $3.03 | $1.74 | 42.6% |
| Daytona · MI355X | Not stated | $5.99 | $3.44 | 42.6% |
| Daytona · B200 | Not stated | $6.25 | $3.59 | 42.6% |
| Daytona · B300 | Not stated | $7.10 | $4.08 | 42.5% |
| Daytona · H100 | Not stated | $3.95 | $2.27 | 42.5% |
| Daytona · H200 (SXM) | Not stated | $4.54 | $2.61 | 42.5% |
| Daytona · RTX 4090 | Not stated | $0.99 | $0.57 | 42.4% |
| AWS · A100 SXM 40GB · 8 GPUs | us-east-1 | $2.74 | $1.63 | 40.5% |
| Google Cloud · Tesla T4 | us-central1 | $0.35 | $0.21 | 40.4% |
| Google Cloud · L4 | us-central1 | $0.56 | $0.34 | 40.1% |
| Google Cloud · H200 (SXM) | us-central1 | $9.31 | $5.59 | 40.0% |
| Google Cloud · A100 SXM 40GB | us-central1 | $2.93 | $1.76 | 40.0% |
| Google Cloud · A100 SXM 80GB | us-central1 | $3.93 | $2.36 | 40.0% |
| Spheron · H200 (SXM) | Not stated | $5.51 | $3.35 | 39.2% |
| Spheron · A100 | Not stated | $1.48 | $1.19 | 19.6% |
| Spheron · H100 | Not stated | $2.65 | $2.20 | 17.0% |
| TensorDock · H100 SXM | Not stated | $2.25 | $1.91 | 15.1% |
| AWS · H100 SXM · 8 GPUs | us-east-1 | $6.88 | $7.22 | -4.9% |
| Spheron · L40S | Not stated | $0.96 | $1.05 | -9.4% |
Cite this data
GPU Rental Prices, "Spot GPU prices: interruptible rates against on-demand", prices checked 22 October 2026, https://gpurentalprices.com/spot
Quote it with a link to this page. Downloads of the whole dataset are on Data & API.
Before you rent spot
A spot rate can change between checks; each one here is the rate read at the day's check. How much warning a provider gives before it stops a spot machine decides how often a job should save its state, and whether the disk survives decides what a restart costs. Each provider's own documentation, read 22 October 2026:
Notice before a spot machine stops, by provider
| Provider | Notice before it stops | The disk after an interruption | Source |
|---|---|---|---|
| AWS | 2 minutes, through instance metadata and EventBridge, best effort; none when the machine hibernates | Kept when the interruption behaviour is Stop: the EBS volumes are preserved and billed | AWS 1 ↗ · AWS 2 ↗ |
| Daytona | None: a spot GPU sandbox can be stopped at any time, with no warning or webhook | Destroyed with the sandbox; Daytona keeps the sandbox's record, with the eviction time, for 24 hours | Daytona ↗ |
| Google Cloud | None by default, or 120 seconds if set when the machine is created; then a best-effort shutdown of up to 30 seconds | Kept when the machine is stopped (the default) and billed; deleted with it if the termination action is Delete | Google Cloud ↗ |
| Lium | Not published | Not published | Lium ↗ |
| Microsoft Azure | Up to 30 seconds, best effort, if you opt in to Scheduled Events | Kept under the default Deallocate policy and billed; deleted with the machine under Delete | Microsoft Azure ↗ |
| Prime Intellect | Not published | Not published | Prime Intellect ↗ |
| Spheron | None: no grace period | Attached volumes survive the machine's termination by default | Spheron 1 ↗ · Spheron 2 ↗ |
| TensorDock | Not published | Storage is still billed while the machine is outbid | TensorDock ↗ |
| Together AI | Up to 5 minutes while the node drains: a maximum, not a guarantee | Not published | Together AI ↗ |
| Vultr | Not published | Not published | Vultr ↗ |
When a job cannot be stopped, compare reserved rates or the on-demand offers on each GPU page.
Spot, on-demand or serverless: which fits the job · Spot against on-demand pricing · Every matched discount