AWS P5 instances are Amazon's flagship GPU offering for large-scale AI training, built around 8-GPU H100 clusters with 3,200 Gbps EFA networking. Thunder Compute is built for individual developers and small teams who need affordable on-demand H100 and A100 access without the quota overhead and per-GPU premium that comes with AWS.
This comparison covers AWS P5 pricing across instance types, the full cost picture including storage and egress, and when each provider is the right choice.
Key takeaways:
- AWS P5 p5.48xlarge: $55.04/hr for 8 H100s ($6.88/GPU-hr). Single-GPU p5.4xlarge: $6.88/hr.
- Thunder Compute H100 PCIe: $2.19/hr. A100 80GB: $1.09/hr. Both with per-minute billing.
- AWS adds EBS storage at $0.08/GB/month and egress at $0.09/GB after 100GB. Thunder Compute includes 100GB per GPU free and charges no egress fees.
- AWS P5 requires quota approval. Thunder Compute offers instant access with no approval process.
AWS P5 Instance Types
AWS P5 instances come in three variants. The P5 series uses H100 GPUs; the P5e and P5en series use H200s.
| Instance | GPU | Count | GPU Memory | On-demand $/hr | $/GPU-hr |
|---|---|---|---|---|---|
| p5.4xlarge | H100 | 1 | 80GB | $6.88 | $6.88 |
| p5.48xlarge | H100 | 8 | 80GB each | $55.04–$68.80 | $6.88–$8.60 |
| p5en.48xlarge | H200 | 8 | 141GB each | $63.30–$79.12 | $7.91–$9.89 |
| Thunder Compute H100 | H100 PCIe | 1–8 | 80GB | $2.19–$17.52 | $2.19 |
| Thunder Compute A100 | A100 | 1–8 | 80GB | $1.09–$8.72 | $1.09 |
| Last update: July 15, 2026. | |||||
AWS P5 Specs vs Thunder Compute
| Feature | AWS P5 (p5.48xlarge) | AWS P5en (p5en.48xlarge) | Thunder Compute H100 |
|---|---|---|---|
| GPUs | 8 × H100 80GB | 8 × H200 141GB | 1–8 × H100 PCIe 80GB |
| System RAM | 2 TiB | 2 TiB | 32–960 GB |
| Storage | 30.7 TB NVMe (ephemeral) | 30.7 TB NVMe (ephemeral) | 100-500GB persistent disk per GPU |
| Networking | 3,200 Gbps EFA | 3,200 Gbps EFA | 7 Gbps |
| Billing | Per second (60s minimum) | Per second (60s minimum) | Per minute |
| Persistent storage | EBS at $0.08/GB/month (separate) | Same | 100GB included; $0.05/GB/month for snapshots |
| Egress | $0.09/GB after 100GB | Same | None |
| Quota required | Yes | Yes | No |
| IDE integration | SSH, SageMaker Studio | Same | VS Code, Cursor, Windsurf (one-click) |
| Last update: July 15, 2026. | |||
AWS P5 Cost Calculations
At $2.19/hr versus $6.88/hr, the per-GPU price difference between Thunder Compute and AWS P5 compounds quickly on any meaningful training run.
| Runtime | Thunder Compute H100 ($2.19/hr) | AWS P5 p5.4xlarge ($6.88/hr) | Lower Cost |
|---|---|---|---|
| 30 seconds | $0.04 | $0.11 | Thunder Compute |
| 1 minute | $0.04 | $0.11 | Thunder Compute |
| 5 minutes | $0.18 | $0.57 | Thunder Compute |
| 1 hour | $2.19 | $6.88 | Thunder Compute |
| 10 hours | $21.90 | $68.80 | Thunder Compute |
| 24 hours | $52.56 | $165.12 | Thunder Compute |
AWS Spot pricing for P5 instances can reduce these costs by 60–70%, though availability is not guaranteed and instances can be interrupted with minimal notice. For training runs that cannot be checkpointed frequently, Spot is a meaningful risk.
Storage and Egress: The Hidden AWS P5 Costs
AWS P5 storage is ephemeral. The 30.7 TB of NVMe included in a p5.48xlarge is lost when the instance stops. Teams that need persistent storage for model checkpoints, datasets, or development environments must attach EBS volumes separately at $0.08/GB/month, plus snapshot storage at $0.05/GB/month. Data transfer out of AWS costs $0.09/GB after the first 100GB per month.
Thunder Compute includes 100GB of disk per GPU at no extra charge. Your home directory and installed packages survive instance stops automatically. Snapshots for long-term storage cost $0.05/GB/month, and there are no egress fees.
For a training workflow that moves 500GB of checkpoints out of the cloud monthly, AWS adds approximately $45/month in egress alone. At Thunder Compute, that cost is zero.
Billing, Quotas, and Setup Time
AWS P5 instances require a quota increase request before provisioning. New accounts start with zero P5 vCPU quota, and approval can take hours to days depending on the region and GPU family. Once approved, P5 instances bill per second with a 60-second minimum, meaning if a job completes in under 60 seconds it's still billed for the full minute.
Thunder Compute requires no quota approval. Create an account, add a payment method, and launch an H100 instance in under a minute. Billing is per minute from first launch with no minimum job duration.
When to Choose Each Provider
AWS P5 is the right choice for large-scale distributed training on 8 or more H100s where 3,200 Gbps EFA networking meaningfully improves throughput. Teams already deep in the AWS ecosystem, those with SageMaker pipelines or S3-based data lakes, will also benefit from P5 over a standalone GPU cloud.
Thunder Compute suits most other H100 workloads: fine-tuning, inference, experimentation, and training that runs well on 1-8x GPUs in a single node. Lower per-GPU rates, persistent storage included, per-minute billing, and VS Code, Cursor, and Windsurf integration mean less time on infrastructure and more on model work.
Last Thoughts on AWS P5 Pricing
AWS P5 is built for one thing: massive parallel training at scale. For teams that genuinely need multi-node H100 clusters with 3,200 Gbps interconnects, it is hard to beat. For most fine-tuning, evaluation, and inference workloads, the $6.88/GPU-hr price tag, quota overhead, and storage make Thunder Compute the more practical starting point.
How much does AWS P5 cost per hour?
AWS P5 p5.48xlarge costs $55.04/hr for 8 H100 80GB GPUs. The single-GPU p5.4xlarge costs $6.88/hr. Prices vary by region. Additional costs apply for EBS storage ($0.08/GB/month) and data egress ($0.09/GB after 100GB).
What are the hidden costs of AWS P5?
EBS storage at $0.08/GB/month, data egress at $0.09/GB (free for the first 100GB every month), and a 60-second billing minimum. AWS also requires quota approval for P5 instances.
Should I use AWS P5 or Thunder Compute?
Use AWS P5 for multi-node H100 clusters with 3,200 Gbps networking and deep AWS integration. Use Thunder Compute for affordable and flexible H100 or A100 access at $2.19/hr and $1.09/hr respectively, with per-minute billing and no quota approval.
Does Thunder Compute require minimum commitments or quotas?
No. Thunder Compute offers instant access with per-minute billing and no quota approval process. AWS P5 requires a quota request and has a 60-second billing minimum.
What is the per-GPU cost of AWS P5 H100 instances?
AWS P5 p5.48xlarge is $55.04/hr for 8 H100s, or $6.88/GPU-hr. The single-GPU p5.4xlarge is also $6.88/hr. Thunder Compute H100 PCIe is $2.19/hr per GPU.
Does AWS P5 offer spot pricing?
Yes. AWS Spot pricing for P5 instances can be 60-70% lower than on-demand, though availability is limited and instances can be interrupted. Thunder Compute does not offer spot pricing; all instances are on-demand and non-interruptible.
What storage is included with AWS P5 instances?
AWS P5 includes 30.7TB of ephemeral NVMe storage that is lost when the instance stops. Persistent storage requires a separate EBS volume at $0.08/GB/month. Thunder Compute includes 100GB per GPU at no extra charge; snapshots cost $0.05/GB/month.