Cheapest GPU Cloud in 2026 — Live Price Comparison

The GPU cloud market is fragmented — prices vary 4-5× for the same hardware depending on provider, region, and pricing tier. GPUHawk tracks 108 live listings across 11 providers so you don't have to check each one manually.

Cheapest GPU Cloud Options by Model

These are the lowest on-demand prices we're currently tracking for each GPU model. Prices update every 6 hours from 11 providers. Click any GPU model to see all providers and pricing tiers.

GPU ModelCheapest PriceEst. MonthlyProviderVRAMAction
RTX A4000 $0.0700/hr $51/mo TensorDock 16 GB Rent →
RTX 3090 $0.1000/hr $73/mo Salad Cloud 24 GB Rent →
RTX 4090 $0.1600/hr $117/mo Salad Cloud 24 GB Rent →
RTX A5000 $0.2100/hr $153/mo TensorDock 24 GB Rent →
RTX A6000 $0.4022/hr $294/mo Vast.ai 48 GB Rent →
L40S $0.6000/hr $438/mo Salad Cloud 48 GB Rent →
A100 SXM $0.6563/hr $479/mo Vast.ai 39 GB Rent →
A100 PCIe $0.6951/hr $507/mo Vast.ai 80 GB Rent →
L40 $0.8200/hr $599/mo RunPod 48 GB Rent →
Gaudi 2 $1.2100/hr $883/mo Sesterce 96 GB Rent →
GH200 $1.6400/hr $1197/mo Sesterce 96 GB Rent →
H100 PCIe $1.8000/hr $1314/mo FluidStack 80 GB Rent →
H100 SXM $1.8300/hr $1336/mo Sesterce 80 GB Rent →
H100 NVL $1.9500/hr $1424/mo Sesterce 94 GB Rent →
H200 SXM $2.0900/hr $1526/mo Sesterce 141 GB Rent →
B200 SXM $4.1688/hr $3043/mo Vast.ai 179 GB Rent →
HGX B300 $7.3900/hr $5395/mo RunPod 288 GB Rent →

Prices shown are on-demand (guaranteed). See all 108 listings →

GPU Cloud Provider Comparison

All 11 providers tracked by GPUHawk, ranked by cheapest available GPU. Each provider has different strengths — some focus on consumer GPUs at low prices, others specialize in enterprise H100/A100 clusters.

ProviderGPU ModelsCheapest GPUTotal ListingsAvailable GPUs
TensorDock 7 $0.0700/hr 7 RTX A4000, RTX A5000
Salad Cloud 5 $0.1000/hr 5 RTX 3090, RTX 4090, L40S
Vast.ai 12 $0.1222/hr 12 RTX A6000, A100 SXM, A100 PCIe, B200 SXM
RunPod 15 $0.2500/hr 17 L40, HGX B300
FluidStack 5 $0.3200/hr 5 H100 PCIe
Hyperstack 5 $0.5000/hr 5
Verda 4 $0.5500/hr 4
Sesterce 9 $0.7200/hr 9 Gaudi 2, GH200, H100 SXM, H100 NVL, H200 SXM
CoreWeave 7 $0.7700/hr 7
Lambda Labs 5 $0.9900/hr 5
Vultr 2 $2.0100/hr 2

How Prices Change: Spot vs On-Demand

On-demand instances are guaranteed — you keep them as long as you're paying. Spot (interruptible) instances are cheaper because the provider can reclaim them when demand spikes. The savings can be massive:

GPU ModelOn-DemandSpotSavings
A100 PCIe $0.6951/hr $0.5618/hr 19% cheaper
A100 SXM $0.6563/hr $0.1535/hr 77% cheaper
B200 SXM $4.1688/hr $3.8483/hr 8% cheaper
H100 NVL $1.9500/hr $2.5900/hr -33% cheaper
H100 PCIe $1.8000/hr $1.9900/hr -11% cheaper
H100 SXM $1.8300/hr $1.4422/hr 21% cheaper
H200 SXM $2.0900/hr $0.5000/hr 76% cheaper
HGX B300 $7.3900/hr $2.0044/hr 73% cheaper
L40 $0.8200/hr $0.6900/hr 16% cheaper
L40S $0.6000/hr $0.4704/hr 22% cheaper
RTX 3090 $0.1000/hr $0.1342/hr -34% cheaper
RTX 4090 $0.1600/hr $0.1950/hr -22% cheaper
RTX A4000 $0.0700/hr $0.1089/hr -56% cheaper
RTX A5000 $0.2100/hr $0.1600/hr 24% cheaper
RTX A6000 $0.4022/hr $0.1356/hr 66% cheaper

When to use spot instances

  • Training runs with checkpointing — save progress every N steps, resume if interrupted
  • Batch inference — process a large dataset where partial completion is OK
  • Hyperparameter sweeps — run many short experiments in parallel

When to use on-demand

  • Production serving — API endpoints that need to stay up
  • Interactive development — Jupyter notebooks and debugging sessions
  • Critical deadlines — can't afford an interruption mid-job

Why checking regularly matters

GPU cloud pricing is not static. Providers adjust prices based on supply and demand, new hardware launches shift the market, and promotional pricing comes and goes. A provider that was cheapest last month may not be cheapest today. GPUHawk tracks all 11 providers automatically so you always see the current best deal.

How to Choose the Right Cloud GPU

Budget Tier

RTX 3090 / RTX 4090

24 GB VRAM. Best price-to-performance for inference, Stable Diffusion, and fine-tuning models up to ~13B parameters. The RTX 4090 is the default choice for cost-conscious ML work.

RTX 4090 pricing →
Mid Tier

A100 40GB / L40S

40-48 GB VRAM. Handles medium models, multi-GPU training, and production inference. The A100 is the industry standard; the L40S is newer with better efficiency.

A100 40GB pricing →
High Tier

A100 80GB / H100

80 GB VRAM. Required for training models above 30B parameters and running large models at full precision. The H100 is 2-3× faster than A100 for transformer workloads.

H100 SXM pricing →

Frequently Asked Questions

What is the cheapest GPU cloud in 2026?

It depends on the GPU model you need. Based on live pricing data from 11 providers, the cheapest option changes frequently as providers adjust prices. Check current prices on our compare page, which updates every 6 hours.

How much does a cloud GPU cost per hour?

Cloud GPU pricing ranges from under $0.20/hr for consumer GPUs (RTX 3090, RTX 4090) to $2-4/hr for enterprise GPUs (H100, H200). The average across all 108 listings is $1.71/hr. Monthly costs at 24/7 usage range from ~$150 to $3,000+.

What is the difference between spot and on-demand GPU pricing?

Spot (interruptible) instances can be 30-70% cheaper than on-demand, but the provider can reclaim them at any time. On-demand instances cost more but are guaranteed. Use spot for fault-tolerant batch jobs; on-demand for production serving and interactive work.

Which cloud GPU is best for deep learning?

For training large models: H100 SXM or A100 80GB. For inference and fine-tuning: RTX 4090 offers exceptional price-to-performance. For production serving: L40S or A10G. See our GPU model pages for detailed specs and pricing per model.

How often are GPUHawk prices updated?

Prices are scraped from all 11 providers every 6 hours. The data on this page is cached for 5 minutes to keep page loads fast while staying current.

Compare All 108 GPU Listings

Filter by GPU model, provider, pricing tier, and VRAM. Find the cheapest option for your specific workload.

Open Full Comparison →