Rent a GPU for AI: cloud GPU prices from 13 providers, and the card your model needs

The cheapest cloud GPU for your model

Start from what you want to run. The tool works out the memory a model needs with the same formula as the rest of this site, keeps every card that holds it with room to spare, and prices each one at its cheapest listed provider. It opens on Llama 3.3 70B at 4-bit with 8k of context; change any of it.

Cloud GPU prices by card

The lowest on-demand price per GPU for each card, read from each provider on 19 October 2026. "Any host" includes marketplaces and home PCs; "run by the provider" leaves them out; the last price column is the cheapest of AWS, Google Cloud and Azure. Serverless prices, billed per second of work, are in the provider table below.

• The same card, very different prices. An H100 SXM is listed from $2.69 (RunPod Community, third-party host) to $12.29 (Azure, 8-GPU machines): 4.6 times as much for the same GPU.
• Vast.ai is named without prices. It is a marketplace where each host sets its own price, and its terms do not allow its prices in a comparison like this one without a written agreement. Check its live prices.
• Lium figures are the cheapest listing you could rent when we read its public node list: Secure machines only, not Spot ones that can be reclaimed, and one GPU where the machine rents them singly. Cards with nothing free at that moment are left out. Lium also publishes "reference" prices and what renters paid over 30 days; both are often below any listing, so we use neither.
• Prices move, and stock comes and goes. Nebius raised its H100 from $3.85 to $4.50 an hour on 1 October 2026. At RunPod and the marketplaces a listed card can be gone a minute later.
• Cheaper with strings attached. Interruptible machines cost less: Nebius's preemptible H100 from $0.79 an hour, Google's Flex-start H100 at $4.79. So do commitments: Hyperstack's reserved H100 SXM from $2.72, CoreWeave up to 60% off. This page compares on-demand prices only.
• Two providers publish no usable price list. TensorDock shows live prices only after you sign in; its public price list is dated July 2024. Voltage Park publishes only a starting price for a single H100, $1.99 an hour, and has merged with Lightning AI.

H100 rental: price per hour

Every on-demand H100 price here, cheapest first: what it costs to rent one by the hour with no contract. The SXM card is the fast one (3.35 TB/s); the PCIe card reads memory at 2 TB/s, and the H100 NVL has 94 GB at 3.9 TB/s. The cheapest rows are marketplaces, on machines third-party hosts run; from a provider that runs its own machines, a single H100 starts at $2.50 (PCIe, Hyperstack) and $3.49 (SXM, Jarvislabs, in India).

Cheaper per hour, dearer per token

For one person chatting with a model, generation speed is set by memory bandwidth: each new token reads the model's weights once. A card that costs twice as much an hour but reads memory three times faster is cheaper per token. Here is Llama 3.3 70B at 4-bit, about 45.8 GB with 8k of context, on every card that holds it alone, at the cheapest single GPU from a provider that runs the machine:

On paper the MI300X is the cheapest per token, at $7.94 per million: it costs 1.9 times the A100 80 GB's hourly price ($1.35, the cheapest hour here) but reads memory 2.7 times faster, so each token costs 30% less. It is AMD's card and runs vLLM and llama.cpp through ROCm rather than CUDA, where nobody has measured the formula's efficiency; check your software first.

Among NVIDIA cards, the H200 ($4.47 an hour on DigitalOcean) costs 3.3 times the A100 80 GB's hourly price but only 1.3 times as much per token, and generates 2.5 times faster. When you are the one waiting for the answers, that is often worth paying.

These are ceilings for one user, worked out from memory bandwidth, not benchmarks. A star marks a speed past 100 tokens a second, where the formula overstates; its cost per token counts 100, a best case. A server batching many users is limited by compute as well as bandwidth, which this comparison does not measure.

Every provider, every price

Each provider's on-demand price per GPU-hour for the cards here, as published on 19 October 2026. "×8" means the card comes only in 8-GPU machines.

What the hourly price leaves out

The hourly rate is the start of the bill. What a stopped machine costs, where your models live between sessions and what it costs to download your results differ more than the rates do.

• Stopping is not always stopping. DigitalOcean bills a powered-off GPU Droplet in full, Hyperstack bills a stopped VM in full until you hibernate it, and Lambda has no stop, only terminate; nor does Lium, where you delete a pod to stop paying. At RunPod the GPU stops billing, but the pod's volume disk costs double while it waits.
• An empty balance can delete your work. RunPod, Vast.ai, Lium, Hyperstack and SaladCloud bill from prepaid credit. When it runs out, RunPod terminates pods that have no network volume, and their data cannot be recovered; Lium emails warnings, then removes pods 2 minutes after a charge empties the balance.
• Only in machines of eight: Lium's RTX PRO 6000; Lambda's A100 80 GB; Hyperstack's H100 SXM, H200 and B200; CoreWeave's L40S, A100 80 GB, RTX PRO 6000, H100 SXM, H200 and B200; AWS's A100 80 GB, H200 and B200; Google Cloud's H100 SXM; Azure's H100 SXM, H200 and MI300X. CoreWeave also approves your organisation before you can rent at all.
• Quotas before GPUs. AWS: new accounts have a GPU quota of 0, so request an increase first. Google Cloud: may need a GPU quota increase first; no GPUs on free trial accounts. Azure: request vCPU quota first.
• Data out. Free at RunPod, SaladCloud, Lambda, Hyperstack, Nebius and CoreWeave. AWS charges $0.09 per GB after 100 GB a month, Google Cloud $0.12 per GB (Premium tier), and Vast.ai each host's own rate, uploads included.

Three kinds of GPU cloud

What differs between providers is who runs the machine, how you pay, and what you have to ask for first.

Serverless skips the machine. Modal and RunPod Serverless bill per second of work, model loading included, and almost nothing while idle. Per hour of work they cost more: RunPod Serverless charges $1.10 an hour for an RTX 4090, against $0.34 to $0.89 for a pod. They win when requests come in bursts, and lose when the GPU works all day.

GPU marketplaces compared

A GPU marketplace rents you other people's machines: independent hosts list them, and the platform takes the booking and the payment. That is why marketplaces hold the cheapest hour for most cards on this page, and why the question to ask is about the host, not the price. These four publish prices anyone can check:

• Each is quoted the way it sells. RunPod Community and SaladCloud publish a price per card. On Vast.ai and Lium each host sets its own price, so there is no fixed one. We quote the cheapest Lium listing free when we read it, which can be gone a minute later. Vast.ai publishes a median for each card, but its terms do not allow its prices in a comparison without a written agreement, so check its live prices.
• Lium is a marketplace on Bittensor, a crypto network that also pays its hosts in tokens. By its own stats feed it lists about 1,600 GPUs from 81 hosts in 23 countries, and billed $1.6 million in the last 30 days. You pay by card, stablecoins or TAO, from $10. There is no stop: you delete a pod to stop paying, so keep what you want on a volume.
• All four bill from prepaid credit, and any host's machine can go away. Keep models and results somewhere you control, not only on the rented disk.

How a provider gets listed. The same rules for every provider: on-demand prices per card that anyone can read without signing in, and rental by the hour or second with no contract. A marketplace with no fixed price is quoted at the median it publishes, where at least 20 GPUs of the card are listed and its terms allow it, or else at the cheapest listing you can rent when we read it; never at a reference or promotional price. Vast.ai is named without prices because its terms do not allow them in a comparison. What it takes to start (a card, a quota request, an approval) is shown with each one. TensorDock and Voltage Park are missing because their prices are not public in that way. A listing is free and editorial: no provider pays to be listed or placed, and no link on this page earns us anything. Prices are re-read at least once a month; a provider whose price has changed can write to info@nodegrove.io, and we check it at the source.

Who runs the machine you rent

A cloud GPU is someone else's computer, and the kind of provider decides whose. On Vast.ai, Lium and RunPod's Community Cloud the machine belongs to an independent host the platform lists or vets; RunPod's terms say it is not responsible for what hosts do, and Vast.ai's security FAQ says protection from hosts "varies significantly". Lium's FAQ says the host of an ordinary machine can read your pod's files, memory and environment variables; its confidential machines, which use Intel TDX, keep the host out. SaladCloud runs containers on PCs in homes and small businesses. RunPod's Secure Cloud runs in Tier 3 and 4 data centres, and the GPU clouds and big clouds run or contract their own.

For public models and test data, a third-party host is fine and the cheapest hour going. For client data, private documents or anything under a contract, use a machine the provider runs, and delete its disks when you are done.

Rent or buy

NVIDIA lists the RTX 5090 from $1,999; what a shop charges may differ. Rented, it is $1.19 an hour in a data centre on RunPod, or $0.50 on SaladCloud's home PCs. Your own card also draws power: at 575 W and $0.20 a kWh, about $0.12 an hour. Take that off, and the card has paid for itself after about 1,860 hours against RunPod, or 5,190 against SaladCloud.

Over three years, that makes renting cheaper below about 12 hours a week of use against the data-centre price, or 33 hours against the home-PC price; above it, buying wins. That leaves out the rest of the PC, what the card sells for later, and the cards you cannot buy at all: an 80 GB data-centre GPU is not a personal purchase. Put your own figures into the build vs rent calculator; the local AI server guide covers the three build tiers.

RunPod alternatives

If you are leaving RunPod, the right alternative depends on why:

• Cheaper consumer cards: SaladCloud ($0.33 for an RTX 4090, home PCs, containers only). Vast.ai is a marketplace where each host sets its own price; compare its live prices.
• A cheaper single H100: Hyperstack ($2.50, PCIe) or Jarvislabs, in India ($3.49, SXM).
• Serverless with less to manage: Modal, per second of work, with free credit every month.
• Large clusters and contracts: Lambda, Nebius and CoreWeave.
• ComfyUI without the setup: the ComfyUI cloud guide compares managed services; for chat models, the Ollama Cloud guide covers paying per token instead.

Questions

How much does it cost to rent a GPU?

From $0.17 an hour for an RTX 3090 (SaladCloud, home PCs) to $14.24 per GPU for a B200 on AWS, sold as 8 for $114 an hour. The cards most people rent: an RTX 4090 from $0.33 ($0.89 in a data centre), an A100 80 GB from $1.35, an H100 from $1.99 ($2.50 in a data centre) and an H200 from $3.59 (RunPod Community, third-party host). On demand, before tax, read 19 October 2026.

What is the cheapest cloud GPU?

By the hour, an RTX 3090 for $0.17 (SaladCloud, home PCs). From a provider that runs the machine, an L4 for $0.44 on Jarvislabs, in India, or a 48 GB RTX A6000 for $0.50 on Hyperstack. The cheapest card that runs your model, and the cheapest per token, are often different cards: the picker above works out both.

How much is an H100 per hour?

From $2.69 for an H100 SXM (RunPod Community, third-party host). From a provider that runs the machine, $2.50 for the PCIe card and $3.49 for the faster SXM, up to $12.29 on Azure. The H100 section above lists all 20 prices, with which ones come only as 8-GPU machines.

Where can I rent an H100?

One H100 by the hour, no contract: on a marketplace from $1.99 (PCIe, RunPod Community, third-party host), from a provider that runs the machine at $2.50 (PCIe, Hyperstack) or $3.49 (SXM, Jarvislabs, in India), or on AWS as p5.4xlarge for $6.88. Google Cloud and Azure rent their H100 SXM machines only as eight GPUs. Stock comes and goes by the minute at every provider, so check the console before you plan around one.

What are the best Vast.ai alternatives?

Another marketplace, or a machine the provider runs. RunPod Community Cloud lists an RTX 4090 at $0.34 from vetted hosts; SaladCloud rents one on a home PC for $0.33 with CPU and memory included; Lium, a newer marketplace, had one from $0.70 when we read it. For private data, pick a provider that runs the machine, such as RunPod, Hyperstack or Jarvislabs. The marketplaces section above compares what happens to your data on each.

What GPU do I need to run a 70B model?

Llama 3.3 70B at 4-bit with 8k of context needs about 45.8 GB, just over the 45.6 GB a 48 GB card holds with the site's 5% headroom. So one 80 GB card, from $1.35 an hour for an A100 80 GB on Hyperstack, or two 48 GB cards. At 8-bit it needs about 81 GB, an RTX PRO 6000 from $1.69; at 16-bit about 150 GB, an MI300X from $2.59, or two A100 80 GB cards from $2.70.

Is renting a GPU cheaper than buying one?

Below about 12 hours a week, yes: over three years, renting an RTX 5090 in a data centre costs less than buying one at NVIDIA's listed price and paying for its power. Against the cheapest home-PC rental the line moves to about 33 hours a week.

What are the best RunPod alternatives?

SaladCloud for cheaper consumer cards, a marketplace such as Vast.ai or Lium for any host's machine, Hyperstack and Jarvislabs for a cheaper single H100, Modal for serverless, and Lambda, Nebius or CoreWeave for clusters.

Which is cheaper, RunPod or Vast.ai?

RunPod lists an RTX 4090 at $0.34 an hour on its Community Cloud and $0.89 on Secure Cloud. Vast.ai's hosts set their own prices, so check its live prices for the card you want. Vast also charges for data transfer, where RunPod does not.

Is Vast.ai safe?

It is a marketplace: your instance runs on a host's machine, and Vast's own security FAQ says protection from hosts "varies significantly". For sensitive work it points you to Secure Cloud, its hosts in certified data centres. Fine for public models and test data; for anything confidential, use Secure Cloud or a provider that runs its own machines.

Do cloud GPUs charge when stopped?

Storage almost always does. Some providers bill the whole stopped machine: DigitalOcean until you destroy it, Hyperstack until you hibernate it.

Can I rent a single H100 on AWS?

Yes: p5.4xlarge is one H100 SXM for $6.88 an hour on demand in US East. Google Cloud ($11.06 per GPU) and Azure ($12.29) sell their H100 SXM machines only as eight GPUs; Azure rents a single H100 NVL for $6.98.

What is GPU as a service?

Renting GPU time instead of buying hardware: by the second or hour on a machine of your own (RunPod, Lambda, Hyperstack), by the second of work with no machine to manage (Modal, RunPod Serverless), or inside a big cloud next to the rest of your systems (AWS, Google Cloud, Azure).