Petaflare

AWS GPU pricing: what every EC2 GPU instance costs per GPU, and a crypto alternative

AWS rents GPUs inside EC2 instances, from a slice of an L4 to eight B300s, billed to a card or an invoice, while Petaflare rents whole GPUs for crypto only. Divided per GPU, from AWS's own price file, here is what each instance costs, which one fits which job, and how it compares with renting a card by the hour.

Crypto-only payment, no card, no KYCUpdated 19 November 2026Prices checked 19 November 2026

AWS GPU pricing on demand in US East (N. Virginia), divided per GPU, runs from $0.53 an hour for an NVIDIA T4 (g4dn.xlarge) to $6.88 for an H100 (p5), $7.91 for an H200 (p5en), $14.24 for a B200 (p6-b200) and $17.80 for a B300 (p6-b300). An A100 80GB costs $3.43 on a p4de and an L40S $1.86 on a g6e.xlarge. These are Linux on-demand prices from AWS's price file published on 18 November 2026; Spot, Savings Plans and Capacity Blocks can lower them, and new accounts need a GPU quota increase first.

EC2 GPU instance prices per GPU-hour

US East (N. Virginia), Linux, on demand. Per-GPU price = instance price ÷ number of GPUs. Smaller sizes of the same family have fewer CPUs and less memory, so their per-GPU price differs.

Source: AWS's official EC2 on-demand price file for US East (N. Virginia), published 18 November 2026 and read on 19 November; see Amazon EC2 on-demand pricing. GPU counts from AWS's instance-type pages.
InstanceGPUsPer instance-hourPer GPU-hour
p6-b300.48xlarge8 × B300$142.42$17.80
p6-b200.48xlarge8 × B200$113.93$14.24
p5en.48xlarge8 × H200$63.30$7.91
p5.48xlarge8 × H100$55.04$6.88
p5.4xlarge1 × H100$6.88$6.88
p4de.24xlarge8 × A100 80GB$27.45$3.43
p4d.24xlarge8 × A100 40GB$21.96$2.74
g7e.48xlarge8 × RTX PRO 6000 Blackwell 96GB$33.14$4.14
g7e.2xlarge1 × RTX PRO 6000 Blackwell 96GB$3.36$3.36
g6e.xlarge1 × L40S$1.86$1.86
g5.xlarge1 × A10G$1.01$1.01
g6.xlarge1 × L4$0.80$0.80
g4dn.xlarge1 × T4$0.53$0.53
g6f.largeA fraction of an L4$0.20—

Source: AWS's official EC2 on-demand price file for US East (N. Virginia), published 18 November 2026 and read on 19 November; see Amazon EC2 on-demand pricing. GPU counts from AWS's instance-type pages.

Which AWS GPU instance for which job

  • P6 (B200, B300) and P5 (H100, H200): large-scale training and serving the biggest models, eight GPUs linked by NVLink and fast networking between instances. The p5.4xlarge is the only way to rent a single H100.
  • P4d and P4de (A100): the previous generation for training, still common for fine-tuning; the p4de has the 80 GB cards.
  • G7e (RTX PRO 6000 Blackwell, 96 GB): a lot of memory per card at a lower price than the P family, for inference of mid-size models and graphics work.
  • G6e (L40S, 48 GB): inference, image generation and rendering.
  • G6 (L4) and G5 (A10G), 24 GB: cost-efficient inference and video work.
  • G4dn (T4, 16 GB): the cheapest whole GPU on EC2, for small models and development.
  • G6f: fractions of an L4 for very light workloads.

AWS also offers its own AI chips, Trainium and Inferentia, in the Trn and Inf families. They are not GPUs and need AWS's Neuron software.

Did AWS change its GPU prices?

Yes, downwards. In June 2025 AWS announced on-demand price cuts of up to 45% for P5 instances, up to 26% for P5en and up to 33% for P4d and P4de, effective from 1 June 2025 for Linux, with slightly smaller cuts for other operating systems. Prices in the table above already include those cuts. Capacity Blocks, which reserve GPUs for future dates, are priced separately and change more often.

Ways to pay for GPUs on AWS

  • On-Demand: the prices above, per second with a one-minute minimum for Linux, no commitment.
  • Spot Instances: spare capacity at a discount that AWS says can reach 90%, reclaimed with a two-minute warning. GPU Spot prices change constantly; check them with the command below before relying on them.
  • Savings Plans: a commitment to spend a fixed amount per hour for one or three years, in exchange for lower rates.
  • Capacity Blocks for ML: GPU instances reserved for a future window, priced per block, for jobs with a fixed start date.

All of them sit on top of the instance price; storage on EBS and data sent out to the internet are billed separately.

Quotas and the aws CLI

EC2 limits the vCPUs you can run per instance family, with separate quotas for P instances and for G and VT instances. New accounts often start at zero for GPU families, so request an increase in Service Quotas before you plan a job. These commands list GPU instance types, show your GPU quotas and print recent Spot prices for a single H100.

# GPU model, count and memory for some instance families
aws ec2 describe-instance-types --filters "Name=instance-type,Values=g6*,g6e*,p5*" \
  --query "InstanceTypes[].[InstanceType,GpuInfo.Gpus[0].Name,GpuInfo.Gpus[0].Count,GpuInfo.TotalGpuMemoryInMiB]" --output table

# Your on-demand quotas for P and G instances (in vCPUs)
aws service-quotas list-service-quotas --service-code ec2 \
  --query "Quotas[?contains(QuotaName,'On-Demand P') || contains(QuotaName,'On-Demand G')].[QuotaName,Value]" --output table

# Recent Spot prices for one H100
aws ec2 describe-spot-price-history --instance-types p5.4xlarge \
  --product-descriptions "Linux/UNIX" --max-items 5
Quotas are counted in vCPUs, not GPUs: a p5.48xlarge needs 192 vCPUs of P-instance quota.

AWS or a GPU cloud paid in crypto

The same cards, per GPU-hour on demand, on 19 November 2026.

AWS prices include each instance's CPUs, memory, local storage and networking; ours are per GPU with the machine included. Every card and block is on what a GPU-hour costs.
GPUAWSPetaflare, fixed
B200$14.24$4.89
H200$7.91$3.29
H100$6.88$2.39
A100 80GB$3.43$0.99
L40S$1.86$0.87

AWS prices include each instance's CPUs, memory, local storage and networking; ours are per GPU with the machine included. Every card and block is on what a GPU-hour costs.

When AWS is worth it

AWS earns its price when your data is in S3, your pipelines run on SageMaker or EKS, or you need multi-node clusters with EFA networking under an enterprise agreement. For a GPU on its own, without a commitment, the gap is wide: you can rent an H100 here for $2.39 an hour against $6.88 on a p5. The other big clouds are compared on Compute Engine GPU costs and Azure H100 prices, and every provider on cloud GPU providers compared.

AWS GPU questions

How much does a GPU cost on AWS?

From $0.53 an hour for a T4 to $6.88 per H100 and $14.24 per B200, on demand in US East, divided per GPU. Spot and Savings Plans can lower that.

How much does an A100 cost on AWS?

$3.43 per GPU-hour on a p4de.24xlarge with eight A100 80GB cards ($27.45 an hour), or $2.74 per GPU on a p4d.24xlarge with eight A100 40GB cards.

What is the cheapest AWS GPU instance?

The g6f.large, with a fraction of an L4, at $0.20 an hour. For a whole GPU, the g4dn.xlarge with one T4 at $0.53 an hour.

Does AWS use NVIDIA GPUs?

Yes: every P and G instance in the table uses NVIDIA GPUs. AWS also designs its own Trainium and Inferentia chips, sold as Trn and Inf instances.

Sources