AITOT

Calculator

GPU Pricing & Rental Calculator

Compare hourly and monthly GPU rental costs across cloud providers, including spot vs on-demand and power cost factors.

Pricing data refreshed:

The AITOT GPU Pricing Calculator compares hourly and monthly rental rates for NVIDIA H100, H200, A100, B200, RTX 5090, and consumer GPUs across 12 cloud providers — AWS, GCP, Azure, RunPod, Lambda Labs, Vast.ai, CoreWeave, Together, Replicate, Modal, Banana, and Paperspace.

H100 80GB ranges from $1.85/hour on spot (Vast.ai) to $7.20/hour on Azure on-demand — a 4× spread. RunPod community at $2.49/hour and Lambda Labs at $2.99/hour are the typical sweet-spot for training and stateless inference. Hyperscalers win for regulated workloads needing VPC, IAM, and S3 integration.

Toggle "include power cost" to add electricity at your $/kWh rate × GPU TDP. An H100 at 700W with $0.12/kWh costs about $61/month in power running 24/7 — often forgotten in self-host break-even math. Spot pricing saves 50–70% but instances can be preempted.

Cheapest · On-demand

Hyperbolic · H100-SXM

$1,072.80

/month · 720 h

ProviderPer hourCompute / moTotal / mo
HyperbolicSXM5 (spot-style)· Spot-style; community reliability$1.49$1,072.80$1,072.80
Vast.aiSXM5· 24h median$2.40$1,728.00$1,728.00
CrusoeSXM5 HGX$2.65$1,908.00$1,908.00
Lambda LabsSXM5$2.99$2,152.80$2,152.80
RunPodSXM5$3.29$2,368.80$2,368.80
CoreWeaveSXM5 HGX$3.30$2,376.00$2,376.00
PaperspaceSXM5$5.95$4,284.00$4,284.00
GCPSXM5 (A3)· A3 ÷ 8$11.06$7,963.20$7,963.20
AWSSXM5 HGX (8-GPU)· p5.48xlarge ÷ 8 GPUs$12.29$8,848.80$8,848.80
AzureSXM5 (NDH100v5)· ND H100 v5 ÷ 8$12.96$9,331.20$9,331.20

Spot prices are 30-day medians and can spike. Community-tier reliability is lower than secure-spot — don't use for production training without checkpoints.

What this calculator does

12 GPU clouds compared

AWS, GCP, Azure, RunPod, Lambda, Vast.ai, CoreWeave, Together, Replicate, Modal, Banana, Paperspace.

Spot vs on-demand

See both pricing tiers side-by-side. Spot saves 50–70% on H100; great for fault-tolerant training.

Power cost toggle

Add electricity at your local $/kWh × TDP. Critical for self-host break-even math.

VRAM-compatible filter

Tool flags GPUs that can't fit your model — Llama 4 70B FP16 needs 140GB, single H100 (80GB) won't.

Hourly + monthly

Plug in expected utilization to see realistic monthly cost, not just headline hourly rate.

Reserved discounts

AWS, GCP, Azure 1-year reserved pricing included. Typical 30–40% savings vs on-demand.

Quick comparison

H100 80GB hourly rate across major providers

ProviderSpot / CommunityOn-DemandNotes
Vast.ai$1.85$2.20Community marketplace; variable reliability
RunPod Community$2.49—Bare containers; great for training
Lambda Labs$2.49$2.99Reliable; good DX
RunPod Secure—$3.99SOC 2; production-grade
CoreWeave—$3.50High-end DC; reserved discount
GCP$4.40$5.50A3 instance; full GCP stack
AWS$4.91$5.45p5.48xlarge per-GPU; full AWS
Azure—$7.20NC H100 v5; enterprise focus

How to use this calculator

Find the cheapest GPU rental for your AI workload across 12 cloud providers.

  1. 1

    Pick your GPU and model

    Choose H100, H200, A100, or RTX 5090. The tool flags GPU/model VRAM mismatches automatically.

  2. 2

    Set utilization hours

    Enter expected monthly utilization (160h for part-time, 730h for 24/7).

  3. 3

    Choose pricing tier

    Toggle spot vs on-demand. Spot saves 50–70% if your workload tolerates preemption.

  4. 4

    Add power cost (optional)

    For self-host comparisons, enable power-cost modeling at your local kWh rate.

Why use this calculator

  • ✓12 GPU clouds, refreshed monthly
  • ✓Spot + on-demand + reserved pricing
  • ✓Power cost modeling included
  • ✓VRAM compatibility checks built in
  • ✓Open-source community providers covered
  • ✓No login required

Frequently Asked Questions

How much does an H100 cost per hour in 2026?+
H100 80GB pricing ranges from $1.85/hr on spot (Vast.ai, community providers) to $2.49/hr on RunPod community, $3.99/hr on RunPod Secure Cloud, and $4.50–$7.20/hr on AWS, GCP, Azure on-demand. Reserved 1-year contracts on hyperscalers cut to roughly $3.10–$3.50/hr.
Is RunPod cheaper than AWS for GPU rentals?+
Yes — typically 40–60% cheaper for H100 and A100. RunPod community runs $2.49/hr H100; AWS p5.48xlarge runs $4.91/hr per H100 equivalent. Tradeoff: AWS includes VPC, IAM, S3 integration; RunPod is barebones containers. For training or stateless inference, RunPod wins. For regulated workloads, hyperscalers still win.
What about spot vs on-demand GPU pricing?+
Spot saves 50–70% but instances can be preempted at any time. Good for fault-tolerant training (checkpoint every 10 min) and batch inference. Avoid for production serving or critical training runs without a checkpointing strategy.
Does this include electricity cost for self-hosted GPUs?+
Yes — toggle "include power cost" to add electricity at your $/kWh rate × the GPU TDP. An H100 at 700W with $0.12/kWh costs about $61/month in power running 24/7. Often forgotten in self-host break-even math.
When does buying a GPU beat renting one?+
For an H100 at $30,000 capex and $2.50/hr cloud rate, break-even is roughly 12,000 hours of utilization — about 16 months of 24/7 use. Add electricity, cooling, datacenter colo. Most teams underutilize purchased GPUs and lose. Rent unless you have steady-state 80%+ utilization.
Can I run an open-weight 70B model on a single GPU?+
A 70B model in FP16 needs roughly 140GB VRAM — a single H100 (80GB) cannot. Use 2× H100, 1× H200 (141GB), or quantize to int4 to fit in one 80GB GPU. The calculator flags VRAM-incompatible GPU + model pairs.