GPU cloud pricing in 2026: what an H100 hour really costs at 9 clouds
We read the pricing pages of nine GPU clouds, from a marketplace to AWS and Oracle: the hourly price per GPU, how it is billed, and the charges that land after the GPU stops.
Short answer: an H100 costs from $1.71 to $10.00 per GPU-hour across the nine clouds in this index, with a median of $4.50. The low end is a marketplace (Vast.ai) or a community tier (Runpod at $2.69); Thunder Compute, DigitalOcean, Nebius and Lambda sit between $3.20 and $4.59; an AWS capacity block, booked and paid in advance, works out to $5.970 per GPU, CoreWeave to $6.16 on an 8-GPU instance, and Oracle Cloud lists $10.00. Below, each price with its condition, the bill for a month, and the traps.
To pick a provider rather than a price, see our best GPU cloud ranking.
The index is editorial: a sponsored placement is always labeled on its card, and no affiliate link changes a price or an order.
The price of an H100 hour at 9 GPU clouds
A GPU cloud rents you a graphics processor in a data center by the hour, or by the second, so you can train or run AI models without buying the hardware. You start an instance, a virtual machine with one or more GPUs attached, and the meter runs until you stop it or, at some providers, until you delete it.
The unit that matters is the GPU-hour: one GPU for one hour. An 8-GPU instance billed $47.757 an hour costs $5.970 per GPU-hour. Then three ways to buy it: on-demand (start and stop when you want, full price), spot or interruptible (much cheaper, but the provider can reclaim the GPU), and reserved (a lower rate for a commitment of months). Each card shows the on-demand H100 price where the provider lists one (for AWS, its capacity-block rate), the condition attached, and what else you pay.
What to require from a GPU cloud: a published per-GPU price, the billing step (per second, per minute or per hour), the price of storage while the GPU is off, the egress price for moving your data out, and whether your instance can be preempted.
Before tax. Per GPU-hour unless stated. Read on each provider's pricing page in 2026.
Vast.ai
Marketplace, H100 SXM
On-demand or interruptible
- Prices set by supply and demand
- Interruptible 50%+ cheaper
- Billed per second
Runpod
H100 SXM, Community Cloud
Secure Cloud $4.39
- Billed per second, no commitment
- H100 PCIe from $1.99 on Community
- Idle volume disk $0.20/GB a month
Thunder Compute
H100 SXM instance
A100 80GB $1.09
- Billed per minute while running
- No egress fees
- 100 GB disk per GPU included
DigitalOcean
GPU Droplet H100
$3.26 on a 12-month reservation
- Billed per second, 5-minute minimum
- Still billed when powered off
- 15,000 GiB transfer included
Nebius
HGX H100
Preemptible from $0.79
- Spot price moves with capacity
- L40S from $1.55
- Network traffic free
Lambda
8x HGX H100 instance
A100 40GB $1.99
- 1x H100 SXM: $4.89 per GPU
- 16-GPU cluster: $6.16 per GPU
- Plus sales tax, VAT or GST
AWS
p5.48xlarge capacity block
Booked and paid in advance
- 8 x H100 or 1 x H100 block
- Same rate in every listed region
- B200 block $14.208 per GPU
CoreWeave
HGX H100, 8 GPUs
Spot $19.71 per instance
- Up to 60% off for committed usage
- No egress fees on storage
- Built for multi-GPU clusters
Oracle Cloud
BM.GPU.H100.8, 8 GPUs
$80.00 per 8-GPU server
- First 10 TB of egress free each month
- H200 server also $10.00 per GPU
- Highest H100 rate in this index
Nine H100 prices, from $1.71 to $10.00 per GPU-hour; median $4.50. All are on-demand rates except AWS, whose capacity block is booked and paid in advance. Vast.ai's price is the cheapest H100 SXM listing on its marketplace, which updates hourly, not a fixed rate.
The prices that really matter, cloud by cloud
For each cloud: what the entry price buys, how it is billed, and the condition that changes the bill.
The clouds
Vast.ai: the lowest listing, set by the market
Vast.ai is a marketplace: prices are set by supply and demand across its data centers, so you compare offers rather than read one rate. The cheapest H100 SXM listing starts at $1.71 an hour, with a median listing at $1.73: the lowest headline H100 price in this index; only spot and preemptible offers, such as Nebius's from $0.79, go lower. Interruptible instances are 50% or more cheaper again, but they can be reclaimed: use them for batch training with checkpoints, not for an endpoint your users hit.
Billing is per second, with no minimum hours. For steady work, reserved terms of 1, 3 or 6 months take up to 50% off and guarantee the capacity. The catch is variance: two listings for the same GPU can differ in network speed, disk and reliability, so run a short job before a long one.
- Cheapest H100 SXM
- $1.71 an hour
- Reserved, 1 to 6 months
- up to 50% off
Runpod: per-second pods, Community or Secure Cloud
Runpod rents GPU pods billed per second with no commitment, on two tiers. An H100 SXM costs $4.39 an hour on Secure Cloud, or from $2.69 on Community Cloud; an H100 PCIe is $2.89 on Secure Cloud and from $1.99 on Community. On Secure Cloud, an A100 SXM costs $2.19 and an L40S $1.09.
What blows up a Runpod bill is the disk: a volume disk costs $0.10 per GB a month while the pod runs and $0.20 while it is idle. A 100 GB volume left on a stopped pod costs $20 a month for nothing. Our Runpod pricing breakdown and Runpod review go further, and Runpod alternatives lists the other pod clouds.
- H100 SXM Community
- from $2.69 an hour
- H100 SXM Secure
- $4.39 an hour
Thunder Compute: a low fixed price and no egress
Thunder Compute lists an H100 SXM instance at $3.20 an hour and an A100 80GB at $1.09, billed per minute while the instance runs. The first 100 GB of disk per GPU is included, extra disk costs $0.219 per GB a month, snapshots $0.05 per GB a month, and there are no egress fees. Instances are never preempted.
A useful sanity check: one A100 80GB for 176 hours a month, 8 hours a day on 22 working days, costs $191.84 for the GPU. Run the same A100 around the clock for 720 hours and it is $784.80. Teams can also reserve dedicated GPUs for a fixed term at a lower rate, by quote.
- H100 SXM
- $3.20 an hour
- A100 80GB, 176 h
- $191.84
DigitalOcean: simple GPU Droplets, billed even when off
DigitalOcean's GPU Droplets put an H100 at $4.41 per GPU-hour on-demand, or $3.26 on a 12-month reservation. An L40S costs $1.57 and an RTX 4000 Ada $0.76. Billing is per second with a 5-minute minimum, and the H100 Droplet comes with 15,000 GiB of transfer. Its spot plans cover the B300 and AMD Instinct GPUs, not the H100.
The trap: when you power off a GPU Droplet, you are still billed for it, because the disk, CPU, RAM and IP stay reserved. You stop paying only when you destroy it. The 12-month rate saves $828 a month per H100 kept busy around the clock (($4.41 − $3.26) × 720 hours), but only if you would have run it all year anyway. Leaving DigitalOcean altogether? See switch from DigitalOcean.
- H100 on-demand
- $4.41
- 12-month reserved
- $3.26
Nebius: on-demand or preemptible at a moving price
Nebius lists an HGX H100 at $4.50 per GPU-hour on-demand and from $0.79 as a preemptible instance. The preemptible rate is a spot price: it moves with capacity and demand, as often as every 15 minutes while your job runs, so a job budgeted at $0.79 can cost more by the time it ends. An L40S starts at $1.55 on-demand with an Intel CPU, and its preemptible version is billed at a flat rate rather than a spot price.
Networking ingress and egress are listed as free, which matters when you move large datasets; object storage on the Standard tier charges $0.015 per GiB of egress. Prices are shown before taxes, VAT among them. Commitments of several months lower the rate, through the sales team.
- H100 on-demand
- $4.50
- Preemptible
- from $0.79
Lambda: the per-GPU price falls with instance size
Lambda prices its 8x HGX H100 instance at $4.59 per GPU-hour, so about $36.72 an hour for the whole machine. Smaller instances cost more per GPU: an H100 SXM is $4.69 in a 4x instance, $4.79 in a 2x and $4.89 alone, while a single H100 PCIe is $3.79. An A100 40GB costs $1.99. Prices are before sales tax, VAT or GST.
Its 1-Click Clusters, booked for two weeks to a year through the sales team, cost more per GPU than the 8x instance: $6.16 for 16 H100s, $5.85 for 64 and $5.54 for 256. Eight H100s for a 720-hour month come to $26,438.40 before tax: price your training run in GPU-hours before you choose the instance size. To compare dedicated hardware instead, see dedicated GPU servers.
- 8x H100
- $4.59 per GPU-hour
- 1x H100 SXM
- $4.89 per GPU-hour
AWS: H100s by the capacity block, paid up front
AWS publishes its H100 rate as a capacity block: capacity you book in advance for a set period. A p5.48xlarge with eight H100s costs $47.757 an hour, or $5.970 per GPU-hour, the same in every region the grid lists, and a p5.4xlarge with one H100 costs the same $5.970. A p6-b200.48xlarge with eight B200s is $113.666 an hour, $14.208 per GPU.
The block is paid up front when you schedule it, at the rate of the day, and AWS updates those rates with supply and demand. On top, data you move out of AWS is billed per GB beyond the 100 GB a month that are free across all AWS services. If you are already leaving AWS, our switch from AWS guide covers the move.
- p5.48xlarge block
- $47.757 an hour
- Per H100
- $5.970
CoreWeave: cluster-grade H100s, no egress fees
CoreWeave rents an 8-GPU HGX H100 instance at $49.24 an hour on-demand, about $6.16 per GPU, the same figure it lists for a single-GPU inference instance. The same instance on spot capacity is listed at $19.71 an hour in North America, 60% less, and committed usage takes up to 60% off the on-demand price.
Storage carries no ingress, egress or transfer fees: AI Object Storage costs $0.06 per GB a month on the Hot tier and $0.015 on Cold. CoreWeave sells GPUs at cluster scale; for one developer renting one GPU for an afternoon, the pod clouds above cost less. Other options are in CoreWeave alternatives.
- 8x HGX H100
- $49.24 an hour
- Spot, same instance
- $19.71 an hour
Oracle Cloud: the highest list price, 10 TB of free egress
Oracle Cloud lists its BM.GPU.H100.8 shape, eight H100s in one server, at $10.00 per GPU-hour, so $80.00 an hour for the machine: the highest H100 price in this index. Its H200 server, BM.GPU.H200.8, is listed at the same $10.00 per GPU-hour, so the H200 costs no more than the H100 there.
What Oracle gives back is the network: the first 10 TB of outbound data transfer each month are free, from North America, Europe and the UK as from its other regions. Against AWS's 100 GB, that is a real line on a bill for a team that ships large model checkpoints out every week.
- BM.GPU.H100.8
- $10.00 per GPU-hour
- Free egress
- 10 TB a month
Seven traps and one tip to pay less for a GPU hour
The comparison guides list hourly prices. These are the lines on the pricing pages that change what you really pay, each with its number.
The traps
Trap 1: a stopped GPU that still bills
At DigitalOcean, a powered-off GPU Droplet is still billed at its full rate: $4.41 an hour for an H100 until you destroy it. At Runpod, the GPU stops billing but an idle volume disk costs $0.20 per GB a month. Snapshot, then delete.
- DigitalOcean H100 off
- $4.41 an hour
Trap 2: a spot price is not a quote
Nebius's preemptible H100 starts at $0.79, but spot prices move with capacity. Vast.ai's interruptible instances are 50% or more cheaper, and can be reclaimed. CoreWeave's spot H100 instance is listed at $19.71 an hour against $49.24 on-demand. Budget spot jobs at the on-demand price and treat the saving as a bonus.
- Nebius
- $0.79 vs $4.50
Trap 3: billing steps on short jobs
Per-second billing is not always per second: DigitalOcean bills at least 5 minutes per Droplet. Thunder Compute bills per minute. For dozens of short test runs, the minimum is part of the price.
- DigitalOcean
- 5-minute minimum
Trap 4: a per-GPU price on an 8-GPU machine
Lambda's $4.59 and CoreWeave's $6.16 are per GPU on 8-GPU instances: $36.72 and $49.24 an hour for the machine. If your job fits on one GPU, rent one: Lambda's single H100 SXM costs $4.89 and its H100 PCIe $3.79, and AWS sells a one-H100 capacity block at the same $5.970 as the 8-GPU one.
- Lambda 1x H100 SXM
- $4.89 an hour
Trap 5: moving data out
Thunder Compute and Nebius list egress as free, and CoreWeave charges none on storage; DigitalOcean's H100 Droplet includes 15,000 GiB of transfer; Oracle Cloud gives the first 10 TB a month. AWS gives 100 GB a month across all its services, then bills per GB. With a training set of several terabytes, the egress line can decide the cheapest cloud more than the GPU rate.
- AWS free egress
- 100 GB a month
Trap 6: a capacity block is paid in full, used or not
An AWS capacity block is charged up front when you schedule it. Book four p5.48xlarge for 48 hours and you reserve 192 instance-hours; stop them one hour early and you used 188, but the reservation charge still covers all 192. Book the hours your run needs, not a comfortable margin.
- Reserved vs used
- 192 vs 188 instance-hours
Trap 7: no monthly cap on a GPU Droplet
DigitalOcean's billing mentions a standard 672-hour monthly limit, but a GPU Droplet used beyond it is still charged for every second. A 720-hour month on an H100 bills $3,175.20, not $2,963.52 (672 × $4.41).
- H100, 720 hours
- $3,175.20
Tip: reserve only what runs all year
DigitalOcean's 12-month H100 rate, $3.26 instead of $4.41, saves $828 a month on a GPU busy 720 hours. On a GPU busy a third of the time, on-demand at $4.41 for 240 hours costs $1,058.40, less than the $2,347.20 a reserved GPU costs for the full month. Shorter terms exist: Vast.ai reserves for 1, 3 or 6 months at up to 50% off, and CoreWeave takes up to 60% off for committed usage.
- Saving at full use
- $828 a month
What a GPU cloud really costs your project
Which GPU for which job, the cost of a month, and where hyperscalers fit.
The budget
Pick the GPU before the cloud
An H100 is built for training and serving large models; an A100 does most fine-tuning for much less ($1.09 at Thunder Compute, $1.99 for a 40GB at Lambda); an L40S ($1.09 at Runpod, $1.57 at DigitalOcean) handles inference and image generation. Paying H100 prices for a job that fits on an L40S multiplies the bill for nothing.
- L40S
- $1.09 to $1.57
- H100
- $1.71 to $10.00
H100 or H200
An H200 carries 141 GB of memory against 80 GB for an H100, for a gap that depends on the cloud: $4.47 against $4.41 at DigitalOcean, $5.40 against $4.50 at Nebius, $5.29 against $4.39 on Runpod Secure Cloud, $6.866 against $5.970 per GPU on an AWS capacity block, and the same $10.00 at Oracle Cloud. Pay for the H200 when the model or its batch does not fit in 80 GB; otherwise the H100 is the cheaper hour.
- DigitalOcean H200
- $4.47
- Nebius H200
- $5.40
One H100 for a month, or a year
A month around the clock is 720 hours. One H100 then costs $1,231.20 at Vast.ai's cheapest listing, $1,936.80 on Runpod Community, $2,304 at Thunder Compute, $3,175.20 at DigitalOcean, $3,240 at Nebius, $3,304.80 at Lambda (per GPU, 8x instance), $4,298.40 on an AWS capacity block, $4,435.20 at CoreWeave and $7,200 at Oracle Cloud, before storage and egress.
Over a year (8,760 hours), the same GPU costs $14,979.60 at Vast.ai's floor, $28,557.60 on DigitalOcean's 12-month rate and $87,600 at Oracle Cloud. That yearly figure is what to set against a quote for buying the hardware, plus its power and hosting, before you decide to own rather than rent.
- Cheapest
- $1,231.20
- Oracle Cloud
- $7,200
Hyperscalers or specialist clouds
AWS and Oracle Cloud price their H100s inside 8-GPU machines (AWS also sells a one-GPU block): $5.970 and $10.00 per GPU. Google Cloud and Azure sell them inside their own virtual machines, priced by region. GPU clouds price per GPU-hour, from $1.71 to $6.16 in this index. Stay on a hyperscaler when your data and services already live there and egress would eat the saving; otherwise the GPU clouds set the price. Moving off Google Cloud? See switch from Google Cloud.
- AWS per H100
- $5.970
- Oracle per H100
- $10.00
Do you make a tool in this category? Email us: hello@hackceleration.com
How we collected the prices
For each cloud, we opened the provider's official pricing page in 2026 and recorded the on-demand price of an H100 per GPU-hour, the billing step, the reserved or spot option, and the charges around the GPU: disk, idle state, egress. Where a provider prices whole instances, we divided by the number of GPUs. Vast.ai is a marketplace: we took its cheapest H100 SXM listing. For AWS we took the capacity-block rate, a fixed grid by region, so its price is a block booked in advance, and we say so wherever it appears. Google Cloud and Azure sell H100s as whole virtual machines priced by region, with no per-GPU rate we could show with its condition, so they are not in the median. We measure price, not performance: to choose a provider, read our best GPU cloud ranking.
- SourceOnly the provider's pricing pageOfficial page
- GPUOn-demand, per GPU-hourH100
- Instances8-GPU prices divided by 8Divided
- MarketplaceVast.ai's lowest H100 SXMCheapest listing
- ExtrasDisk, idle, egressPriced
- MedianFrom $1.71 to $10.009 prices
The index is editorial: a sponsored placement is always labeled on its card.
GPU cloud pricing: your questions
How much notice do you get before a spot GPU is reclaimed?
Often little or none. DigitalOcean aims to give at least 2 hours' notice on a spot GPU Droplet, but may reclaim it with less or no notice in some cases; its spot price is locked when you create the Droplet. Nebius works the other way: the spot price itself can change as often as every 15 minutes while the job runs. Save checkpoints and script your restarts.
Are eight single-GPU instances the same as one 8-GPU node?
No. An 8-GPU node links its GPUs inside one machine: AWS's p5.48xlarge has 900 GB/s NVSwitch between its eight H100s and 3,200 Gbps of network, while its one-GPU p5.4xlarge has 100 Gbps and, with one GPU, no peer-to-peer link. Eight separate instances are fine for independent jobs and slow for one model split across GPUs.
Do GPU cloud prices include tax?
Usually not. Lambda's prices are plus applicable sales tax, VAT or GST; Nebius shows all prices without any applicable taxes, VAT among them; DigitalOcean applies tax by law in some countries. A European team should add its VAT rate to every hourly price before comparing a quote.
Why is the same H100 priced so differently?
Because "H100" covers several cards and several offers. The SXM version costs more than the PCIe card: $4.39 against $2.89 on Runpod Secure Cloud, $4.89 against $3.79 for a single GPU at Lambda. Add a marketplace host, a community tier and a hyperscaler, and the same class of GPU goes from $1.71 to $10.00 an hour in this index.
Do I pay extra for the operating system on a GPU instance?
Yes on AWS, for licensed systems. On a capacity block, the operating system is billed on top of the reservation while the instances run: Red Hat Enterprise Linux on a p5.48xlarge adds $1.8432 per instance-hour, billed per second, so four instances that run 188 instance-hours in all (47 hours each) add $346.52. SUSE Linux Enterprise Server is billed per hour, with a one-hour minimum.
Can I reserve GPUs for a short training run?
Yes, from a few days. An AWS capacity block covers a set period you choose, such as 48 hours; Lambda's 1-Click Clusters start at two weeks; Vast.ai reserves for 1, 3 or 6 months. The trade-off is the bill: the AWS block is paid up front for every reserved hour, and AWS updates block rates with supply and demand.
What if the GPU I want is sold out?
You wait, switch type or reserve. Lambda's instances are self-serve, first-come access; at Thunder Compute, if creation fails because a type is sold out, you check again and pick another configuration. Guaranteed capacity comes with reservation: Vast.ai's reserved terms and Runpod's Reserved Clusters both promise it, against a commitment.
Get the next pricing index in your inbox
GPU prices move every few weeks. We'll write when one changes.
