GPU Cloud vs AWS GPU Instances: Cost and Performance
A GPU cloud, or neocloud, is a provider built to rent out GPUs for AI work. AWS is a general-purpose cloud that sells GPU instances alongside storage, databases and other services.
Both rent GPUs by the hour or for a set term, usually at a lower rate than paying by the hour. Hourly pricing costs less for short or on-and-off jobs, and a term costs less for work that keeps the GPUs running most of the time.
You pay for unused hours on a GPU cloud term and on an AWS reservation that holds GPUs, so compare the options by your work's share of the term's hours.
Quick Verdict by GPU Workload
Pick the pricing model from the work's length and steadiness:
| GPU workload | Lower-cost option | Reason |
|---|---|---|
| Months of steady training on a fixed schedule | A reserved GPU cloud term, or an AWS Savings Plan (a discount for a one- or three-year commitment) with a capacity reservation | Near-continuous use fills the reserved hours |
| Inference with traffic that rises and falls | On-Demand capacity from AWS or a GPU cloud | Instance count can shrink in quiet hours |
| Small inference models | AWS instances that rent a slice of one GPU | The buyer pays for a fraction of a GPU |
| Checkpointed experiments | AWS Spot, discounted spare capacity | A saved checkpoint lets the job restart after an interruption |
| Multi-week training with a fixed start date | AWS Capacity Blocks, GPU reservations of 1 to 182 days, or a reserved GPU cloud term | Both hold the GPUs for a set window |
AWS's Five EC2 Pricing Models for GPU Instances
AWS sells GPU capacity on EC2, its rented virtual server service, under five pricing models:
- On-Demand: pay for the hours used, at the rate that sets the baseline for AWS's commitment discounts.
- Spot: rent spare EC2 capacity at a discount, with a two-minute interruption notice before AWS stops or terminates the instance.
- Compute Savings Plans: commit to an hourly spend for one or three years across instance families and regions, for up to 66% off On-Demand.
- EC2 Instance Savings Plans: commit to an hourly spend on one instance family in one region, for up to 72% off On-Demand.
- Capacity Blocks for ML: reserve GPU instances for 1 to 182 days, paid up front at a price set by supply and demand.
| Pricing model | Commitment length | How capacity is held |
|---|---|---|
| On-Demand | Pay as you go | A separate On-Demand Capacity Reservation |
| Spot | Pay as you go | Interruptible at two minutes' notice |
| Both types of Savings Plans | 1 or 3 years | A separate On-Demand Capacity Reservation |
| Capacity Blocks for ML | 1 to 182 days | Held for the block |
A commitment discount lowers the hourly rate, and an On-Demand Capacity Reservation holds the GPUs. AWS bills reserved hours at the On-Demand rate, used or idle, less the buyer's commitment discount.
Buyers can't cancel a Capacity Block after purchase. Commitment discounts don't apply to Capacity Blocks, per AWS's Capacity Blocks billing page.
AWS sends Spot interruption notices on a best-effort basis and recommends checking for one at 5-second intervals. A Spot job needs recent checkpoints to restart after an interruption.
AWS GPU Instance Families and Their Pricing Options
Choose the instance family before the pricing model, because an EC2 Instance Savings Plan locks its discount to one family in one region. Capacity Blocks cover the P families, per AWS's Capacity Blocks guide.
Four groups of NVIDIA GPU families on AWS cover most AI work:
- G4dn and G5 (NVIDIA T4 and A10G Tensor Core GPUs): AWS aims these at inference and small training jobs. Capacity Blocks leave them out, so capacity comes On-Demand, from Spot or from a reservation.
- G6 and G6f (NVIDIA L4): G6f instances rent a fraction of one GPU, down to 1/8 of a GPU with 3 GB of memory. A small inference model can run on that slice, so the buyer stops paying for GPU capacity the model leaves unused.
- P5, P5e and P5en (NVIDIA H100 and H200 Tensor Core GPUs): these run training across several servers. AWS cut P5 On-Demand prices by 44% in a June 2025 price change, so a rate fixed before a cut can end up above the market.
- P6 and P6e (NVIDIA Blackwell): a P6e-GB200 UltraServer, a group of instances on one shared interconnect, joins up to 72 GPUs through NVLink, NVIDIA's direct GPU-to-GPU link. A model too large for one 8-GPU server can train on that link.
What a Reserved GPU Cloud Term Includes
A reserved term with a GPU cloud can put three things in one contract:
- Capacity guarantee: the contract holds a set number of GPUs for the term.
- Fixed rate: one price per GPU-hour for the length of the term.
- Dedicated environment: networking, storage and scheduling set up for one customer's workload.
GPU rental contracts run from 1 month to 5 years, per a rental market report from SemiAnalysis, a semiconductor and AI research firm. A buyer can size a term to the project.
The contract's upgrade terms may exclude a GPU generation released mid-term.
Rising prices raise the value of a fixed rate. SemiAnalysis's H100 index shows one-year rental contracts up almost 40%, from about $1.70 per GPU-hour in October 2025 to about $2.35 in March 2026.
Ask for the capacity guarantee and its length in the signed contract.
Head-to-Head: AWS GPU Instances vs a Reserved GPU Cloud Term
Compare AWS and a reserved GPU cloud term, using AWS terms as published in October 2026:
| Dimension | AWS EC2 GPU instances | Reserved GPU cloud term |
|---|---|---|
| Shortest commitment | Pay as you go with On-Demand or Spot | Set by the contract, from 1 month in the rental market |
| Longest standard commitment | 3 years with a commitment discount | Up to 5 years in the rental market |
| Capacity held | A separate reservation, or a Capacity Block of up to 182 days | Stated in the contract |
| Rate | On-Demand rate, less up to 72% with a one-family commitment | One rate for the term |
| Idle reserved hours | On-Demand rate, less the buyer's commitment discount | Billed at the contract rate |
The Break-Even Test for a Reserved GPU Term
Compare two costs over the same term:
- Hourly cost: the On-Demand rate × term hours × the share of term hours in use.
- Reserved cost: the reserved rate × term hours, since a reserved term bills each hour.
Reserve the term when the hourly cost exceeds the reserved cost. A reserved rate 30% below On-Demand breaks even when the GPUs run 70% of the term's hours.
Take the share of hours from the job schedule. Model FLOPs and bandwidth utilization measure a different number, the useful work per GPU-hour.
A low reading raises the cost of that work by the same share under both pricing models, so it doesn't move the break-even point.
Price the project end to end with cost to completion, which adds restarts, queuing and engineering time to the hourly rate.
See our reserved vs Spot GPU pricing comparison for interruptible capacity.
We operate zCLOUD, a GPU cloud that sells On-Demand, reserved and private GPU capacity from more than 40 providers, and have a commercial interest in this comparison.