Where to rent a GPU for AI: choose by workload and commitment
On-demand, community, marketplace, spot, reserved or serverless: which way of renting a GPU fits training, fine-tuning, batch or inference.
Most “where should I rent a GPU” questions are really two questions: how predictable is the workload, and how much commitment can you accept. Answer those first and the list of sensible providers gets short.
The six ways to rent
| Rental model | What you get | Billing seen in the index | Fits |
|---|---|---|---|
| On-demand VM or pod | A dedicated GPU until you stop it | Per second (Runpod, DigitalOcean with a 5-minute minimum) or per minute (Hyperstack) | Interactive work, fine-tuning, short training runs |
| Community / marketplace | A GPU on a third-party host, priced lower or by the host | Per second (Runpod Community, Vast.ai) | Cost-sensitive experiments where an occasional host problem is acceptable |
| Spot / preemptible | Spare capacity that can be reclaimed | CoreWeave publishes spot prices per region; DigitalOcean aims for at least 2 hours’ notice | Checkpointed batch jobs and training that can resume |
| Reserved / cluster | Guaranteed capacity for a term | DigitalOcean 12-month plans; Lambda clusters from 2 weeks and 16 GPUs; Hyperstack monthly-invoiced reservations | Steady workloads, large training runs |
| Serverless worker | A GPU that runs only while requests are processed, then scales to zero | Per second (Modal, Runpod) or per minute (Baseten) | Inference with uneven traffic, scheduled jobs |
| Contract (sales) | Custom terms, often for racks such as GB200 NVL72 | Not public | Frontier-scale training |
Pick by workload
Prototyping and fine-tuning on one GPU. Use on-demand with fine-grained billing so idle time costs little. Runpod Secure Cloud, Hyperstack and DigitalOcean publish single-GPU on-demand prices. To pay less and accept a third-party host, look at Runpod Community Cloud or a Vast.ai listing with a high reliability score.
Multi-GPU training in one node. You need 8 GPUs with NVLink (H100 SXM, H200, B200 in HGX servers). CoreWeave rents only whole 8-GPU instances for these, DigitalOcean offers 1 or 8, and Lambda 1, 2, 4 or 8. Price is per GPU, but you pay for all eight.
Multi-node training. Lambda’s 1-Click Clusters (16 GPUs and up, two-week minimum), Runpod Clusters, CoreWeave and Nebius are built for it. Committed pricing is mostly quoted through sales.
Batch jobs that can restart. Spot capacity is the cheapest listed way to get large GPUs, provided your job checkpoints. The index shows CoreWeave, DigitalOcean and Hyperstack spot rows.
Production inference. If traffic is steady, a reserved or on-demand GPU running all the time is easier to predict. If it is spiky or idle at night, a serverless platform bills only while a worker runs, plus cold-start time. See serverless GPU platforms compared.
Before you commit
- Compare the minimum you must rent (a whole 8-GPU node, 16 GPUs for two weeks, a 12-month term) as well as the hourly rate.
- Check what is billed besides the GPU: storage while stopped (DigitalOcean bills powered-off Droplets; Runpod charges more for idle volume disk), CPU and memory (Modal), and egress.
- Regions matter for data residency and latency. Several providers do not publish per-GPU region availability on their pricing pages, so the index marks those “unspecified”.
The full data is in the GPU Cloud Index. If price is the deciding factor, see the cheapest GPU cloud by GPU class.
| Provider | GPU | Offering | USD / GPU-hour | Billing | Commitment | Regions | Source |
|---|---|---|---|---|---|---|---|
| Vast.ai | NVIDIA H100 SXM80 GB | MarketplaceOn-demand marketplace listing (1 GPU) | $2.28 lowest*listed: lowest listing $2.278/hrmedian $2.67 across 5 listings | per second | None stated | USEUAPACOTHER | Vast.ai pricing (live offers API)8 Nov 2026 |
| CoreWeave | NVIDIA H100 SXM80 GBmin 8× | Spot / interruptibleHGX H100 | $2.46listed: $19.71/hr per 8-GPU instance | not-stated | None stated | US | CoreWeave pricing8 Nov 2026 |
| Runpod | NVIDIA H100 SXM80 GB | CommunityCommunity Cloud pod | from $2.69listed: from $2.69/hr | per second | None stated | UNSPECIFIED | Runpod pricing8 Nov 2026 |
| Hyperstack | NVIDIA H100 SXM80 GB | ReservedH100 SXM reservation | from $2.72listed: starting from $2.72/hr | not-stated | Reservation contract (term via sales) | UNSPECIFIED | Hyperstack GPU pricing8 Nov 2026 |
| Hyperstack | NVIDIA H100 SXM80 GB | On-demandH100 SXM | $3.20listed: $3.20/hr | per minute | None stated | UNSPECIFIED | Hyperstack GPU pricing8 Nov 2026 |
| DigitalOcean | NVIDIA H100 SXM80 GB | ReservedHGX H100 12-month plan | $3.26listed: $3.26/GPU/hour | not-stated | 12-month contract | UNSPECIFIED | DigitalOcean GPU Droplet pricing8 Nov 2026 |
| Modal | NVIDIA H100 SXM80 GB | ServerlessServerless function (base rate) | $3.95listed: $0.001097/sec | per second | None stated | UNSPECIFIED | Modal pricing8 Nov 2026 |
| Runpod | NVIDIA H100 SXM80 GB | On-demandSecure Cloud pod | $3.99listed: $3.99/hr | per second | None stated | UNSPECIFIED | Runpod pricing8 Nov 2026 |
| DigitalOcean | NVIDIA H100 SXM80 GB | On-demandHGX H100 GPU Droplet | $4.41listed: $4.41/GPU/hour | per second | None; 5-minute minimum billing period | UNSPECIFIED | DigitalOcean GPU Droplet pricing8 Nov 2026 |
| Nebius | NVIDIA H100 SXM80 GB | On-demandHGX H100 | $4.50listed: $4.50 per GPU-hour | not-stated | None stated | UNSPECIFIED | Nebius prices8 Nov 2026 |
| CoreWeave | NVIDIA H100 SXM80 GBmin 8× | On-demandHGX H100 | $6.16listed: $49.24/hr per 8-GPU instance | not-stated | None stated; committed-use discounts via sales | US | CoreWeave pricing8 Nov 2026 |
| Lambda | NVIDIA H100 SXM80 GBmin 16× | Reserved1-Click Cluster, 16 GPUs | $6.16listed: $6.16 per GPU/hr | not-stated | 2 weeks to 1 year term, 16 GPUs | USEUAPACOTHER | Lambda pricing8 Nov 2026 |
| Lambda | NVIDIA H100 SXM80 GB | On-demand1x H100 SXM | Not publishedAnnounced $4.89 from 16 Nov 2026 (not yet in effect) | not-stated | None stated | USEUAPACOTHER | Lambda pricing8 Nov 2026 |
| Lambda | NVIDIA H100 SXM80 GBmin 8× | On-demand8x HGX H100 | Not publishedAnnounced $4.59 from 16 Nov 2026 (not yet in effect) | not-stated | None stated | USEUAPACOTHER | Lambda pricing8 Nov 2026 |
Normalised to US dollars per GPU per hour from each provider's published price; "listed" shows the wording as published. "Contact sales" and "Not published" mean no public price was found. We never estimate one. Where a provider prices a multi-GPU instance, the per-GPU figure is the instance price divided by its GPU count. * Marketplace prices are set by individual hosts and change constantly; the figure is the lowest live listing on the verification date, with the median of the listings sampled. Announced prices that have not yet taken effect are shown separately and never as the current price. Prices exclude storage, egress and taxes unless the row says otherwise.
Provider facts
Runpod
GPU pods on Secure Cloud (Runpod-operated data centres) or Community Cloud (vetted third-party hosts), plus Serverless endpoints and multi-node Clusters.
- Billing
- per second
- Self-serve
- Yes
- Regions
- UNSPECIFIED
- Commitment
- None for Pods ("billed per second with no commitment"); reserved capacity via sales
- Storage
- Volume disk $0.10/GB/mo running, $0.20/GB/mo idle; network storage from $0.05/GB/mo
- Footprint
- "30+ regions" (no region list on the pricing page)
Verified 8 Nov 2026 · Runpod pricing
Vast.ai
A marketplace: independent hosts set prices for on-demand and interruptible GPU rentals; Vast.ai also sells reserved terms through sales.
- Billing
- per second
- Self-serve
- Yes
- Regions
- US, EU, APAC, OTHER
- Pricing
- "Prices are set by the market, not by Vast"; per-second billing, no minimum
- Reliability
- Varies by host; listings show a reliability score and verification status
- Reserved
- 1, 3 or 6-month terms via sales ("up to 50% off")
Verified 8 Nov 2026 · Vast.ai pricing (live offers API)
Hyperstack
GPU cloud with on-demand VMs billed per minute, spot VMs on selected GPUs and reserved contracts invoiced monthly.
- Billing
- per minute
- Self-serve
- Yes
- Regions
- UNSPECIFIED
- Billing
- On-demand "billed per minute"; reservations invoiced monthly
Verified 8 Nov 2026 · Hyperstack GPU pricing
DigitalOcean
Developer cloud selling GPU Droplets with 1 or 8 GPUs (NVIDIA and AMD), 12-month reserved plans and interruptible spot Droplets.
- Billing
- per second
- Self-serve
- Yes
- Regions
- UNSPECIFIED
- Billing
- Per second with a 5-minute minimum; powered-off Droplets still billed
- Spot
- Reclaimable; aim of at least 2 hours’ notice; price locked at creation
- Price list
- Effective 1 Aug 2026
Verified 8 Nov 2026 · DigitalOcean GPU Droplet pricing
CoreWeave
Large-scale GPU cloud selling mostly 8-GPU HGX instances and NVL72 racks, with on-demand, spot and committed pricing.
- Billing
- not-stated
- Self-serve
- Not stated
- Regions
- US, EU
- Unit
- Mostly 8-GPU instances; GB200 instances are 4 GPUs (2 superchips)
- Commitment
- "Up to 60% discounts" for committed usage via sales
- Regions
- Separate North America and Europe price tables
Verified 8 Nov 2026 · CoreWeave pricing
Modal
Serverless platform: Python functions and containers scale to zero and bill per second for GPU, CPU and memory.
- Billing
- per second
- Self-serve
- Yes
- Regions
- UNSPECIFIED
- Rate basis
- Listed GPU rates are base prices; non-preemptible execution is 3× base and region selection 1.15–1.75×
- Extras
- CPU and memory billed separately; idle containers billed until the scale-down window (60 s default)
- Free usage
- $30/month included compute on the Starter plan
Verified 8 Nov 2026 · Modal pricing
Sources
- Runpod pricing, checked 8 Nov 2026
- Runpod docs: Serverless pricing, checked 8 Nov 2026
- Lambda pricing, checked 8 Nov 2026
- Vast.ai pricing, checked 8 Nov 2026
- CoreWeave pricing, checked 8 Nov 2026
- Nebius prices, checked 8 Nov 2026
- Hyperstack GPU pricing, checked 8 Nov 2026
- DigitalOcean GPU Droplet pricing, checked 8 Nov 2026
- Modal pricing, checked 8 Nov 2026
- Baseten docs: instance types, checked 8 Nov 2026
Index rows carry their own source link and verification date. This page is revisited when: a provider changes its billing granularity, minimum term, spot or reservation terms, or adds or removes a rental model