Cheap GPU cloud: 5 budget providers and what a GPU hour really costs
We read the pricing pages and billing docs of the GPU clouds that sell by the hour: the price per GPU, the billing unit, and the storage and idle costs under the rate card.
Short answer: to pay as little as possible per GPU hour, start with Vast.ai, a marketplace where an H100 SXM listed from $1.53 an hour and the median offer sat at $2.00 when we checked. If you want fixed prices and a simpler console, RunPod's Community Cloud rents an RTX 4090 at $0.34 an hour and an H100 SXM at $2.69, billed by the second. For an A100 80GB at a flat rate, Thunder Compute charges $1.09 an hour. For batch jobs that can be interrupted, Salad rents an RTX 4090 from $0.16 an hour, and an RTX 3090 from $0.09. And when the job is a multi-day training run that must not drop, Lambda costs more but rents data-center GPU servers at public prices.
The hourly rate is only half of the bill. Storage that keeps running after you stop an instance, a GPU left idle overnight and data transfer can double what you pay. Each card below says what makes you pay more.
The ranking is editorial: a sponsored placement is always labeled on its card, and no affiliate link changes a score.
The real price of a cheap GPU cloud
A GPU cloud rents you a graphics card in someone else's machine, by the hour. You pick a card, start an instance (a virtual machine or a container with the GPU attached), connect to it, run your job and stop it. The bill counts GPU-hours: one card for one hour. Eight cards for three hours is 24 GPU-hours. Most budget clouds bill that time by the second or the minute, so a 10-minute test costs a sixth of an hour, not a full hour.
The card decides the price more than the provider does. VRAM, the memory on the card, decides what fits: a model's weights in 16-bit take two bytes per parameter, so a 7-billion-parameter model needs about 14 GB before you count anything else. A 24 GB RTX 4090 runs it; training a much bigger model needs an 80 GB A100 or H100, which costs four to eight times more per hour (on RunPod Community, $1.39 for an A100 and $2.69 for an H100 against $0.34 for an RTX 4090). The H100 hourly price at nine clouds, with the monthly bill, is in our GPU cloud pricing index. Before you compare providers, check four things: the price of the exact card and its VRAM, whether the price is on-demand or interruptible, the billing unit, and what storage and data transfer cost when the GPU is off.
On-demand prices per GPU per hour, in US dollars, before tax, read on each provider's pricing page in 2026. Vast.ai prices are live marketplace offers and move every hour.
Vast.ai H100 SXM
Live marketplace offers
Median offer $2.00; RTX 4090 from $0.27, median $0.47
- Billed by the second
- Interruptible instances 50%+ cheaper
- What makes you pay more: storage billed while the instance is stopped
RunPod Community Cloud
Fixed prices, per second
H100 SXM $2.69; Secure Cloud $0.89 and $4.39
- No ingress or egress fees on Pods
- Volume disk $0.10/GB/mo running
- What makes you pay more: a stopped Pod's volume disk at $0.20/GB/mo
Thunder Compute A100 80GB
One fixed price on a big card
RTX A6000 $0.35; H100 SXM $3.20
- Billed per minute while running
- First 100 GB of disk per GPU included
- What makes you pay more: extra vCPUs at $0.04 per vCPU-hour
Salad RTX 4090
Consumer GPUs for batch jobs
$0.44 at High priority; RTX 5090 from $0.25
- vCPU and RAM included
- Billed per second while running
- What makes you pay more: High priority, 2.75 times the Lowest rate
Lambda 1x A6000
Data-center reliability
1x H100 PCIe $3.79; 8x H100 $4.59 per GPU
- Billed in one-minute increments
- No egress fees
- What makes you pay more: sales tax or VAT on top
One card, five prices: RTX 4090, A100 80GB, H100 SXM
The cards above are each provider's headline. Here is the same GPU across the five picks, so you compare like with like.
On-demand price per GPU per hour, in US dollars, before tax. Vast.ai: median of live offers when we checked, with the cheapest offer in brackets; it moves every hour. Salad: Lowest priority.
RTX 4090, 24 GB
Inference and small fine-tunes
- Salad: $0.16 (High priority $0.44)
- RunPod Community: $0.34 (Secure $0.89)
- Vast.ai: median $0.47 (from $0.27)
- Thunder Compute and Lambda: not offered
A100 80GB
Fine-tuning on one big card
- Vast.ai: median $0.93 (from $0.40), A100 SXM4
- Thunder Compute: $1.09 flat
- RunPod Community: $1.39, A100 SXM
- Lambda: $3.19 per GPU, 8x node only
- Salad: not offered
H100 SXM, 80 GB
Long training runs
- Vast.ai: median $2.00 (from $1.53)
- RunPod Community: $2.69 (Secure $4.39)
- Thunder Compute: $3.20
- Lambda: $4.89 (8x node $4.59 per GPU)
- Salad: not offered
Marketplace, GPU cloud or hyperscaler: where cheap comes from
The same H100 starts at $1.53 an hour on a marketplace and costs $12.29 per GPU on Azure. The gap is not a scam: you buy different things. Pick the kind of cloud first, then the provider.
The five picks below are ordered by use: the cheapest general-purpose GPU hour first, then the special cases (one big card, interruptible batch jobs, training runs that must finish). Each score weights price per GPU-hour at 30% and the four other criteria at 17.5% each, so a pick lower in the list can score higher.
The three kinds
Marketplaces: independent hosts set the price
CheapestVast.ai and TensorDock list machines that independent hosts rent out, from data centers to small operators. Each host sets its own rate, so the same card has many prices and the cheapest offer changes every hour. You save the most here, and you accept that a host can go offline mid-job. Filter on the host's reliability score and verified status before you rent, and keep checkpoints outside the machine. If the marketplace model is not for you, see our Vast.ai alternatives.
Salad goes one step further: its Community Cloud runs on GPUs in homes and small businesses, so you only get consumer RTX cards, running your container, and the owner can take the machine back.
- Examples
- Vast.ai, TensorDock, Salad
- Risk
- Host or node can drop
GPU clouds: fixed prices, their own data centers
MiddleRunPod, Thunder Compute, Lambda, Nebius and CoreWeave sell GPUs at a fixed price list, without the services of a big cloud. You get a machine with a GPU, storage and usually free data transfer. RunPod sells two tiers of the same cards: Secure Cloud, and a cheaper Community Cloud. Nebius lists an on-demand H100 at $4.50 an hour, or from $0.79 as a preemptible instance whose spot price can change as often as every 15 minutes; CoreWeave rents an 8-GPU H100 node at $49.24 an hour, $6.16 per GPU, priced for teams that rent whole clusters. For a GPU you keep for months, a dedicated GPU server or a cheap dedicated server can beat hourly rates.
- Examples
- RunPod, Thunder Compute, Lambda
- H100 range
- $2.69 to $6.16 per hour
Hyperscalers: the GPU comes with the whole cloud
Most expensiveAWS, Google Cloud and Azure bundle the GPU with CPUs, RAM, networking and every other cloud service. On demand in US East, AWS lists a single-H100 p5.4xlarge at $6.88 an hour, and Azure rents its ND H100 v5 as a whole 8-GPU machine at $98.32 an hour, $12.29 per GPU. They also bill data that leaves their cloud, which a budget cloud usually does not. They make sense when your data and your team already live there, not when the GPU hour is the line you want to cut.
- Examples
- AWS, Google Cloud, Azure
- H100
- $6.88 to $12.29 per hour
Vast.ai
Vast.ai is a GPU marketplace: hosts list their machines, set a price, and you rent the offer you want. That is why Vast.ai shows two numbers per card. When we checked, an H100 SXM started at $1.53 an hour with a median offer of $2.00; an A100 SXM4 80GB started at $0.40 with a median of $0.93; an RTX 4090 started at $0.27 with a median of $0.47. The cheapest single offer is often a machine with a weaker CPU, less disk or a poor reliability score, so budget on the median. Prices update every hour.
Three ways to rent change the bill. On-demand instances run until you stop them. Interruptible instances are 50% or more cheaper and can be reclaimed, which suits training that saves checkpoints. Reserved instances take up to 50% off on 1, 3 or 6-month terms. Everything is billed by the second from prepaid credits. A hundred hours of H100 at the median offer comes to $200; at the cheapest offer, $153.
What inflates the bill is everything around the GPU. Storage, bandwidth and the base price are set by each host and vary from machine to machine, and stopping an instance does not stop its storage bill: you have to destroy it. If your credit runs out and no card is saved, Vast.ai destroys the instance and its data. Renters wrote in 2026 that instances were turned offline without warning, leaving their data out of reach.
- H100 SXM from $1.53 an hour, median $2.00
- Interruptible instances 50%+ cheaper
- Billed by the second
- 68+ GPU types, from GTX consumer cards to the B300
- ✓Very low prices on most cards
- ✓Median price shown next to the cheapest offer
- ✓No long-term contract
- ✗Storage keeps billing until you destroy the instance
- ✗Reliability depends on the host
- ✗Data destroyed if credit hits zero with no card saved
RunPod
RunPod rents GPUs in two clouds at a fixed price list. Community Cloud is the budget tier: an RTX 4090 costs $0.34 an hour, an A100 SXM $1.39 and an H100 SXM $2.69. Secure Cloud charges $0.89, $2.19 and $4.39 for the same cards. On an RTX 4090, Community costs less than half of Secure: 100 hours is $34 against $89.
Pods are billed by the second for compute and storage, with no fees for data coming in or going out, which matters when you pull a 100 GB dataset or push a model to the Hub. Storage is where the bill hides: a volume disk costs $0.10 per GB per month while the Pod runs and $0.20 when it is stopped. A 500 GB checkpoint folder left on a stopped Pod costs $100 a month, twice what it costs while you use it. Network storage costs $0.07 per GB per month either way, so it costs less for data you keep. Savings plans can no longer be bought; reserved capacity goes through sales. Every card and storage tier is listed on our RunPod pricing page.
Watch your balance: when it reaches $0, Pods stop, and a Pod without a network volume is terminated with its data. Our RunPod review details reliability and support, and our RunPod alternatives cover the other options.
- RTX 4090 at $0.34 an hour on Community Cloud
- No ingress or egress fees on Pods
- Billed by the second
- Network storage at $0.07/GB/mo
- ✓Fixed, public prices for every card
- ✓Templates for notebooks and inference
- ✓Free data transfer
- ✗Stopped volume disk billed at twice the running rate
- ✗Pods without a network volume deleted at $0 balance
- ✗Deploying needs an hour of credit up front
Thunder Compute
Thunder Compute sells a short list of cards at flat prices: an RTX A6000 at $0.35 an hour, an L40 at $0.79, an A100 80GB at $1.09 and an H100 SXM at $3.20. Each instance includes a set of vCPUs and the first 100 GB of disk per GPU, so the price you see covers a usable machine. Extra vCPUs cost $0.04 per vCPU-hour and extra disk $0.219 per GB per month, billed only while the instance runs.
Instances are billed per minute while running, and data egress costs nothing. To keep your work between sessions, you take a snapshot of the disk at $0.05 per GB per month and restore it next time, instead of paying for a stopped machine. Thunder's own estimate: an A100 used 8 hours a day for 22 days, 176 hours, is $191.84 for the GPU. A hundred hours of A100 80GB costs $109, against $139 on RunPod Community and a median $93 on Vast.ai, where the price moves.
The catch is the range. There is no RTX 4090 or B200 here, the H100 at $3.20 costs more than RunPod Community's $2.69, and the price list covers single instances, not clusters. Reserved and volume pricing go through sales.
- A100 80GB at $1.09 an hour
- RTX A6000 at $0.35
- No egress fees
- Snapshots at $0.05/GB/mo
- ✓Flat prices, no marketplace guesswork
- ✓Disk included up to 100 GB per GPU
- ✓Billed only while running
- ✗Few GPU models
- ✗H100 dearer than RunPod Community
- ✗Not built for multi-node clusters
Salad
SaladCloud runs your containers on GPUs that already sit in homes and small businesses, whose owners are paid for idle time. That is how an RTX 4090 costs $0.160 an hour at the Lowest priority, with vCPU and RAM included, and an RTX 5090 $0.250. Billing is per second while the container runs; downloading the image and starting the container are not charged. A full month of one RTX 4090 at Lowest is $116.80.
The priority you pick is the price. High priority is never preempted by other workloads and costs $0.440 for the same RTX 4090, 2.75 times the Lowest rate. Medium and Low can be preempted by higher-priority groups after running for about 30 minutes, and Lowest suits retryable batch work and stateless inference behind a queue. Over 100 GPUs, long-running jobs or committed spend get a price below the list through sales.
The trade-off is the model itself: on Community Cloud you only get consumer cards (Salad Dedicated rents RTX PRO 6000 servers, on contract), you deploy containers through Salad's Container Engine, and node owners can take their machine back. There is no free trial. If you only need a model answer rather than a GPU, a cheap LLM API often costs less than renting the card.
- RTX 4090 from $0.16 an hour
- vCPU and RAM included
- Billed per second, only while running
- Thousands of distributed GPUs on one API
- ✓Very low prices on consumer GPUs
- ✓No charge for image download or startup
- ✓Scales to many GPUs for batch work
- ✗Community Cloud: consumer cards only, no A100 or H100
- ✗Nodes can be reclaimed by their owners
- ✗Containers only, and no free trial
Lambda
Lambda is not the cheapest per hour, and it is here for a reason: it rents data-center machines at public prices, which is what a multi-day training run needs. A single A6000 costs $1.09 an hour, an A10 $1.29 and an H100 PCIe $3.79. Instances are billed in one-minute increments, and there are no egress fees.
One detail on the price list saves money: an 8x HGX H100 node costs $4.59 per GPU, less than the $4.89 of a single H100 SXM. If your job uses eight cards, renting them as one node is $2.40 an hour cheaper than eight single instances, and the cards are linked for multi-GPU training. A hundred hours of a single H100 PCIe comes to $379, against $199 for the same card on RunPod Community, where it starts at $1.99 an hour.
Mind the idle time: a running instance bills whether you use it or not, until you terminate it, and filesystems bill even when they are not attached to an instance. Prices exclude sales tax, VAT or GST. Reserved clusters start at 16 GPUs, through sales.
- 8x H100 at $4.59 per GPU, below the 1x price
- No egress fees
- Billed per minute
- A6000 at $1.09 an hour
- ✓Data-center machines at public prices
- ✓Multi-GPU nodes cheaper per GPU
- ✓Free data transfer
- ✗Higher hourly rates than marketplaces
- ✗Filesystems bill even when unattached
- ✗Tax added on top of list prices
8 ways to cut your GPU cloud bill
These come from the providers' billing docs, read down to the footnotes, and from what renters report once the first invoices arrive.
The tips
Budget on the median, not the cheapest offer
BudgetA marketplace's headline price is its single cheapest offer. On Vast.ai the H100 SXM started at $1.53 when the median sat at $2.00, 31% more. The cheapest machine often has a weak CPU, a slow disk or a low reliability score. Plan your budget at the median and treat anything below it as a bonus.
- Vast.ai H100 SXM
- $1.53 from, $2.00 median
Destroy what you stop
Big winStopping a GPU instance does not always stop the bill. On Vast.ai, storage keeps billing until you destroy the instance. On RunPod, a volume disk costs $0.20 per GB per month when the Pod is stopped, twice the running rate; on Lambda, filesystems bill even when unattached. Keep data you reuse in cheap network storage or snapshots ($0.07 per GB on RunPod, $0.05 on Thunder Compute) and delete the rest.
- RunPod stopped disk
- $0.20/GB/mo
- Thunder snapshot
- $0.05/GB/mo
Run restartable jobs on interruptible capacity
Big winInterruptible instances on Vast.ai are 50% or more cheaper, on Salad, High priority costs 2.75 times the Lowest rate on the same RTX 4090, and a preemptible H100 on Nebius starts at $0.79 against $4.50 on demand. All three can stop your job. Save a checkpoint every few minutes to storage outside the machine, so a reclaimed node costs you minutes, not hours.
- Vast.ai
- 50%+ off
- Salad RTX 4090
- $0.16 vs $0.44
Use per-second billing for short jobs
Short jobsVast.ai, RunPod and Salad bill by the second, Thunder Compute instances and Lambda by the minute (Thunder's sandboxes, for short isolated jobs, bill per second). For a 10-minute test, that is the difference between paying ten minutes and paying a full hour on a provider that bills hourly. Group many short experiments in one session rather than starting a fresh instance each time.
- Per second
- Vast.ai, RunPod, Salad
- Per minute
- Thunder, Lambda
Rent the smallest card that fits the job
Big winInference on a 7B model fits in 24 GB. On RunPod Community, 100 hours of RTX 4090 cost $34; the same 100 hours on an H100 SXM cost $269. Move to an 80 GB card only when the model or the batch size needs it, and to an H100 only when the job is long enough for its speed to pay back the price.
- RTX 4090, 100 h
- $34
- H100 SXM, 100 h
- $269
Rent eight GPUs as one node
TrainingPer-GPU prices are not always lower for single cards. On Lambda, an 8x H100 node costs $4.59 per GPU against $4.89 for one H100 SXM: $2.40 an hour saved on eight cards, with the fast links between them that multi-GPU training needs.
- Lambda 8x H100
- $4.59 per GPU
- Lambda 1x H100 SXM
- $4.89
Check what moving data costs
Hidden costRunPod Pods charge nothing for data in or out, and Thunder Compute and Lambda charge no egress. On Vast.ai, each host sets its bandwidth price per TB, and it applies to all data sent to or from the instance. Check that line before you pull a large dataset onto a marketplace machine.
- Free transfer
- RunPod, Thunder, Lambda
- Vast.ai
- Set by each host, per TB
Keep a card on file, or a buffer of credit
Avoid data lossBudget clouds run on prepaid credit. On RunPod, Pods stop at a $0 balance and a Pod without a network volume is terminated with its data. On Vast.ai, an account with no saved card has its instances and stored data destroyed after a short grace period. A forgotten top-up can cost you the checkpoint of a week-long run.
- RunPod
- Stops at $0
- Vast.ai
- Data destroyed without a card
Do you make a tool in this category? Email us: hello@hackceleration.com
Stay cheap or pay more: the signals
RunPod Community or Vast.ai, a 24 GB card billed by the second.
Salad at Lowest priority, or interruptible instances on Vast.ai.
Thunder Compute's A100 80GB at $1.09 an hour, disk included.
Pay for data-center nodes on Lambda or RunPod Secure Cloud: a dropped instance costs more than the price gap.
Avoid marketplace hosts; compare the data-center providers in our GPU cloud ranking.
Cheap GPU cloud: your questions
Should I rent a cloud GPU or buy one?
Rent, unless the card would run most hours of the month. A full month of an RTX 4090 at Salad's Lowest priority is $116.80, and $343.10 at Vast.ai's median offer of $0.47 an hour. Those rates include electricity and the rest of the machine. Buying only wins if the card runs most hours of the month, for many months, and you have somewhere to run it.
What is an interruptible or spot GPU instance?
Spare capacity sold at a discount that the provider can take back. On Vast.ai, interruptible instances are 50% or more cheaper than on-demand. On Salad, Medium and Low priority run for about 30 minutes before a higher-priority workload can preempt them. Use them for jobs that save checkpoints often, never for an unsaved notebook.
Why is the same GPU priced so differently from one cloud to another?
Because you buy different things. On a marketplace, prices are set by supply and demand, host by host, so Vast.ai shows a cheapest offer and a median for each card. Hyperscalers bundle the GPU with CPUs, networking and services, and Azure's ND H100 v5 comes only as a whole 8-GPU machine. Even within one provider, the card count changes the rate: Lambda's 8x H100 node is $0.30 cheaper per GPU than its single H100 SXM.
How much VRAM do I need to run a model?
Count two bytes per parameter for a 16-bit model, then add room for activations and the batch. A 7B model needs about 14 GB for its weights, so a 24 GB RTX 4090 or RTX 3090 runs it. Quantized to 4 bits, the same weights take about a quarter of that. Fine-tuning needs much more than inference, which is where 48 GB and 80 GB cards come in.
What is the cheapest cloud GPU right now?
A consumer card at Lowest priority on Salad: an RTX 3090 costs $0.09 an hour with vCPU and RAM included, and the portal lists other classes, such as the RTX 3060 and RTX 3070, from $0.015. That price buys a node its owner can take back. Among the five providers here, the lowest rate for an RTX 4090 that other workloads cannot preempt is RunPod Community's $0.34, and for 80 GB of VRAM at a fixed price, Thunder Compute's A100 at $1.09.
Are serverless GPUs cheaper than renting a Pod?
Not per hour. On RunPod, a serverless worker on a 24 GB RTX 4090 costs $1.34 an hour, against $0.34 for the same card as a Community Cloud Pod. Serverless wins on uneven traffic: flex workers scale to zero when no requests come in, a scaled-down worker is not billed, and the default idle timeout is 5 seconds. If the GPU would be busy less than about a quarter of the time, serverless costs less than a Community Pod left running; above that, rent the Pod.
Can I train on a free GPU before paying for one?
Yes, for learning and small tests. Google Colab gives free access to GPUs in its notebooks, but its free tier heavily restricts them and its usage limits fluctuate. Kaggle notebooks come with a weekly GPU quota that resets every week and is 30 hours, sometimes more depending on demand. That is enough to follow a course or fine-tune a small model, not to serve customers. When you outgrow it, an RTX 4090 at $0.34 an hour on RunPod Community costs $10.20 for those same 30 hours.
Is a marketplace GPU safe for private data?
Be careful. On a marketplace, your job runs on a machine owned by an independent host or, on Salad, on a PC in someone's home. That is fine for public datasets and open models. For customer data, use data-center capacity such as RunPod Secure Cloud or Lambda, and read the provider's data processing terms before you upload anything.
GPU prices move every month. Get the changes in your inbox
One email when a new comparison goes live: prices read at the source, the costs others miss, and a straight verdict.
