Every offer on Aquanode is tagged on-demand or preemptible. On-demand capacity runs in operator-managed facilities with consistent uptime expectations; preemptible compute is real, often cheaper, hardware contributed by independent hosts. The cards above price only the on-demand tier, since that is the number most buyers researching GPU cloud pricing are actually comparing.

Every price on this page is an on-demand, list price. Aquanode does not currently offer spot/interruptible pricing or a separate committed-use/reserved tier: what you see is what a new deployment pays right now, with no bidding and no capacity auction. That also means a price here can move between page loads: it is a live minimum across current on-demand supply for that model, not a fixed rate card, and it is refreshed every 6 hours rather than daily.

Want the cheaper preemptible rate too? See the GPU Availability Index for lowest, median and p90 pricing across both tiers.

What's included in the price

The $/GPU/hr figure above already includes the vCPU, RAM and local storage that ship with that instance. Aquanode computes it by dividing the whole node's price by its GPU count, so there is no separate compute-only line item to add on top: what you see on each card is what an hour of that instance costs, in full.

Network egress and ingress policy is set by each on-demand provider individually and is not uniform across the marketplace. If a workload moves a lot of data in or out, check that specific offer's terms in the console before committing to it, rather than assuming a blanket policy.

How much does an AI server cost?

Renting an 8x H100 server, the standard shape for training and serving large models, costs $22.22/hr on Aquanode right now, or about $16,221 for a month of continuous use. Per-second billing means you pay only for the hours the server actually runs.

Aquanode prices a multi-GPU node per GPU, not as a separate bundled tier, so a larger node costs the same rate per card. The figures below are that same on-demand rate multiplied by node size, for planning a multi-GPU job, not a distinct volume-discounted price.

GPU	1x	2x	4x	8x
H100	$2.78	$5.56	$11.11	$22.22
H200	$5.04	$10.08	$20.16	$40.32
B200	$9.35	$18.70	$37.40	$74.80
A100	$1.50	$3.00	$6.00	$12.00

Which GPU for which workload

H100

The default choice for training and fine-tuning models up to the ~30-40B range on a single card, and for high-throughput FP8 inference.

H200

Same compute as H100 but 76% more VRAM and 43% more memory bandwidth, so a 70B model fits on one card in FP8 with room for a long KV cache, and it suits longer-context inference where KV-cache size, not compute, is the bottleneck.

B200

The highest-throughput card on this marketplace: more than double H100's FP8 throughput and over double the memory bandwidth, plus native FP4 tensor-core support for the newest FP4-quantized model formats.

A100

Still one of the most cost-effective cards for BF16 training and fine-tuning at the 7-30B scale, and its Multi-Instance GPU (MIG) support lets one 80GB card be partitioned into up to seven isolated instances for many small inference workloads at once.

AMD MI300X

192GB on a single card (more than double an H100's 80GB), so it fits a 70B model in FP16 with room to spare and holds larger models without splitting across GPUs, at a memory bandwidth (5.3 TB/s) ahead of H100's 3.35 TB/s.

RTX 5090

The newest and highest-VRAM consumer flagship on this marketplace (32GB, Blackwell generation).

RTX 4090

A consumer flagship that's genuinely useful for renting by the hour: strong FP16/INT8 throughput and 1,008 GB/s bandwidth make it capable for fine-tuning and inference on models that fit in 24GB, at a lower hourly rate than data-center cards.

L40S

A PCIe, air-cooled inference and fine-tuning card that doesn't need SXM/NVLink server infrastructure.

RTX A6000

A workstation-class card with 48GB VRAM at a lower TDP (300W) than the data-center accelerators, making it a reasonable single-card fit for 7-13B fine-tuning and inference without needing a rack-scale power/cooling budget.

How billing works

Per-second billing

Usage is metered per second while an instance is running and billed against your account's credit balance. There is no minimum term and no contract: stop an instance and it stops costing you.

Stop instead of terminate

Stopping a pod captures its environment and volume and closes the provider lease, so there is nothing running and nothing to bill for while it's stopped. Auto-stop does this on its own after a period of low GPU use, off by default.

Volumes

Attach a volume so datasets and checkpoints persist independently of any one GPU instance, on any provider. A detached volume keeps costing while it exists, the same as any storage you don't delete.

Held environment and volume storage

Capturing your environment and volume is free. Holding them in Aquanode's managed storage bills at $0.02 per GiB per 30-day month, metered per second. A stopped pod keeps costing at that rate until you delete it, the same as any other stored data.

Storage that survives a provider switch

A pod's environment and volume are not tied to the provider they were captured on: both can be restored onto a different provider entirely, which matters when the cheapest current offer for a model moves.

Rent a GPU now Talk to us

FAQ

How much does it cost to rent an AI server?

On-demand GPU rates on Aquanode currently range from $0.231/hr to $9.35/hr per GPU, depending on the model. Preemptible rates on many of the same models run lower; see the full live range on the GPU Availability Index.

What's the difference between on-demand and preemptible pricing?

On-demand offers run in operator-managed facilities; preemptible compute is contributed by independent hosts and is often cheaper for the same GPU model. This page prices the on-demand tier only. The GPU Availability Index shows both.

How much does an H100 cost per hour?

The current lowest on-demand rate for H100 is $2.78/hr per GPU, across 21 live on-demand offers. Rates are refreshed every 6 hours; see the live prices above for the current price.

Is Aquanode billed hourly or by the second?

By the second. Usage is metered per second while an instance is running and billed against your account's credit balance, so a partial hour only costs for the seconds actually used.

Are there contracts or minimum commitments?

No. Every listing is on-demand with no signed contract and no minimum term. Stop an instance whenever you want and billing for it stops immediately.

What's the cheapest on-demand GPU rental right now?

RTX 3060 is currently the cheapest on-demand listing, at $0.231/hr per GPU. Rates are refreshed every 6 hours, so check the live prices above for the current cheapest option.

Does Aquanode offer spot or interruptible pricing?

Not currently. Every price on this page is an on-demand list price; spot/interruptible pricing and a separate committed-use/reserved tier are not offered today.

What happens to billing when I stop a pod?

Stopping captures the pod's environment and volume and closes the provider lease, so billing for that instance stops the moment it's stopped. It starts again only once you start the pod, which can land on a different offer if the original one is no longer the cheapest.

How do I actually pay for GPU usage?

Usage is billed against your account's credit balance as it accrues, per second, rather than as a separate invoice per deployment. A team workspace shares one credit balance across everyone in it.

Is there a minimum GPU rental time?

No. There is no minimum term on any listing. A deployment that runs for ten seconds is billed for ten seconds, never a full hour.

Can I rent more than one GPU in the same instance?

Yes. Multi-GPU nodes are listed with their GPU count, and Aquanode prices them per GPU rather than as a separate bundled tier, so an 8x node costs eight times the single-GPU rate for that model. See the multi-GPU table above for current flagship examples.

Does the price include vCPU, RAM and storage, or are those billed separately?

They're included. The $/GPU/hr figure is the whole node's price divided by its GPU count, so the vCPU, RAM and local storage that ship with that instance are already in the number you see, with no separate compute line item.