Best GPU cloud for AI: where to rent a GPU for Stable Diffusion and LLM fine-tuning

The short answer

For most people renting a GPU by the hour, the choice is between two models:

• Pick Vast.ai if price is what matters most. It is a marketplace: independent hosts and datacenters list machines and set their own rates, so you usually find the lowest hourly price there, and interruptible bids go lower still. You trade predictability: prices, reliability and network quality vary from host to host.
• Pick RunPod if you want fixed prices and less hunting. It publishes one hourly rate per GPU, in a cheaper Community Cloud and a pricier Secure Cloud, adds Serverless endpoints that scale to zero, and charges no data transfer fees.

Paying in rubles? Timeweb Cloud and VDSina rent GPU servers on monthly plans. Prices below are as listed on October 29, 2026; GPU rates move weekly, so check the live page before you top up.

How to choose a GPU cloud

The hourly rate is only one line of the bill. What decides the real cost:

• VRAM first, then speed. The model has to fit in GPU memory. Image generation with Stable Diffusion or ComfyUI and LoRA fine-tuning of smaller models are usually done on 24 GB consumer cards (RTX 3090, 4090) or the 32 GB RTX 5090. Full fine-tuning or serving larger LLMs pushes you to 48 GB (L40S) or 80 GB and up (A100, H100, H200, B200).
• Price model. A fixed price list (RunPod) versus host-set marketplace prices (Vast.ai).
• On-demand vs interruptible. On-demand capacity runs until you stop it. Interruptible (spot) capacity is cheaper but can be taken away, so it suits jobs that save checkpoints.
• Storage and egress. Disks often keep billing while the GPU is stopped, and some providers charge for every byte in or out.
• Billing granularity and minimum deposit. Per-second billing makes short experiments cheap; a high minimum deposit or non-refundable credit makes trying a provider expensive.
• Serverless. If you are serving a model behind an API, paying only while requests run can beat keeping a GPU on all day.
• Trust and data. Who owns the machine, and can they see your data?

Marketplace vs Secure and Community Cloud

On Vast.ai's pricing page there is no single price per GPU. Hosts install Vast's software, list their machines and set the rate, which depends on GPU model, reliability, network and location. Vast lists 68+ GPU types. Vast's own FAQ notes that hosts can access data on their machines, so it recommends its Secure Cloud filter (vetted datacenters with ISO 27001 and/or Tier 3/4 certification) for sensitive or production work.

RunPod splits its capacity in two. Community Cloud is cheaper capacity from vetted hosts; Secure Cloud runs in RunPod's data-center partners. The same GPU costs more in Secure Cloud: an RTX 4090 is $0.34/hr in Community Cloud and $0.89/hr in Secure Cloud on the RunPod pricing page.

For a hobby session or a public-dataset fine-tune, the cheaper tier is fine; for client data or private weights, pay for the vetted tier.

GPU prices side by side

RunPod rates are its published on-demand pod prices. Vast.ai rates are the current median of live offers in Vast's public price feed; individual hosts list above and below the median. All per GPU-hour, as listed on October 29, 2026.

On the marketplace, the RTX 3090 and A100 came in below RunPod's fixed rates (and an RTX 3060 median was about $0.08/hr), while on the H200 and B200 RunPod's Community rates were lower that day. Neither is always cheaper: compare the exact GPU on the day you rent.

RunPod: best for predictable pods and serverless

RunPod rents NVIDIA GPUs from the RTX A5000 ($0.16/hr in Community Cloud) up to the H200, B200 and B300. Its products:

• Pods: GPU instances over SSH or Jupyter.
• Serverless: endpoints that scale to zero when idle, billed per second of worker time, e.g. $0.69/hr for 24 GB GPUs, $4.79/hr for an H100.
• Clusters: multi-GPU capacity for distributed training.

According to RunPod's billing docs, compute and storage are billed per second with no data transfer fees, and you can start with $10. Storage is where idle costs come from: a volume disk costs $0.10/GB per month while the pod runs and $0.20/GB per month while it is stopped; network volumes are $0.07/GB per month under 1 TB.

Pick RunPod if you want a known hourly rate, scale-to-zero serving, or no egress bill.

The downsides: credits are non-refundable. At zero balance, pods stop and those without a network volume are terminated with their data. The pricing page shows no spot rates. Details and the sign-up credit are in our RunPod review.

Vast.ai: best for the lowest hourly price

Vast.ai's strength is choice: filter offers by GPU, VRAM, reliability, location and price, and rent the machine as a Linux Docker instance over SSH, Jupyter or a web portal. An API, CLI and Python SDK make launches scriptable.

Three rental types, from Vast's rental-type docs:

• On-demand: the host's fixed rate; cannot be interrupted.
• Interruptible: you set a bid and a higher bid can stop your instance, killing running processes. Vast puts the saving at 50–80% off on-demand. Save checkpoints often.
• Reserved: prepay for a discount of up to 50% based on commitment length, with on-demand priority.

Active rentals are billed per second. The catch is in Vast's billing FAQ: storage is charged for stopped instances too, and bandwidth is charged for every byte sent or received. When your credit hits zero, instances are stopped but not destroyed. Vast's docs don't state a minimum deposit.

Pick Vast.ai if you want the cheapest GPU hour and your jobs checkpoint well enough for interruptible bids.

The downsides: prices and host quality vary; outside Secure Cloud a host can access data on its machine; no Windows or remote desktop. More in our Vast.ai review.

RunPod vs Vast.ai: which for your job

• Stable Diffusion or ComfyUI for a few hours: both work. Vast.ai's RTX 3090 and 4090 offers are usually the cheapest; RunPod's one-click templates and fixed RTX 4090 rate are simpler. On either, delete the disk when you are done or it keeps billing.
• LLM fine-tuning over a weekend: Vast.ai interruptible bids if your script saves checkpoints; RunPod or Vast on-demand if a restart would cost you more than the saving.
• Serving a model as an API: RunPod Serverless is the most direct fit, since endpoints scale to zero and there are no egress fees on responses.
• Large datasets: RunPod charges no data transfer fees; on Vast.ai, check each host's bandwidth price.
• Sensitive data: RunPod Secure Cloud or Vast.ai's Secure Cloud filter.
• Either way: credit is prepaid (RunPod's is non-refundable; on Vast.ai spent credit is final), so top up for the week, not the year, and destroy instances you are done with, because stopped disks keep billing.

Paying in rubles: Timeweb Cloud and VDSina

Timeweb Cloud lists cloud and dedicated GPU servers from the GTX 1080 and RTX 3090/4090 up to A100 80 GB and H100 80 GB, with the H200 on pre-order, on its GPU servers page. Prices are not published: you leave a request and a manager quotes, with a minimum advance payment of 5,000 RUB. Plans are set per month but charged hourly. See the Timeweb Cloud review.

VDSina sells cloud servers with a slice of an NVIDIA L4 or L40S on its GPU pricing page: from 4,770 RUB per month (159 RUB per day) for 8 cores, 8 GB RAM and 2 GB of GPU memory, up to 25,500 RUB per month for 32 cores and 8 GB of GPU memory, in Moscow or Amsterdam. Billing is daily and there is no trial. With 2–8 GB of GPU memory these suit Windows apps, rendering and light inference, not 24 GB-class Stable Diffusion workflows or LLM fine-tuning. See the VDSina review.

What about Lambda, Paperspace, AWS and Google Cloud?

Real options, rarely the cheapest hour. Lambda lists a 1x H100 SXM at $4.29/hr on its pricing page (October 29, 2026), with reserved capacity through sales. Paperspace, now part of DigitalOcean, offers notebook-first GPU machines. AWS and Google Cloud bring the widest regions and compliance coverage, but quotas and egress charges make them a better fit for teams already running there.

For VPS, storage and other infrastructure, browse hosting and cloud services.