Takeaways
<ul><li><a href="https://aws.amazon.com/ec2/instance-types/p5/" rel="noopener nofollow" target="_blank">AWS P5</a> gives you multi-node clusters with 8× H100s per node, 3,200 Gbps networking, and 30.7 TB NVMe priced at ~$55-66/hr (varies by region).</li><li>Want the cheapest, no-commitment way to fine-tune or serve models? <a href="https://www.thundercompute.com/pricing">Thunder Compute</a> offers H100 80 GB for $1.38/hr, billed per second with persistent storage and one-click instance creation in VS Code.</li><li>AWS adds separate costs for storage and data egress, up to <a href="https://aws.amazon.com/ebs/pricing/" rel="noopener nofollow" target="_blank">$0.08/GB/month for EBS</a> and <a href="https://aws.amazon.com/ec2/pricing/on-demand/" rel="noopener nofollow" target="_blank">$0.09/GB for egress</a> after 100 GB.</li><li>AWS requires quotas and has a 60-second billing minimum. Thunder Compute doesn't.</li></ul>
AWS P5 Instances
There are different types of AWS P5 instances, each with its dedicated GPU accelerator:
| Instance | GPU | Variants |
|---|---|---|
| P5 | NVIDIA H100 | p5.4xlarge (1 GPU), p5.48xlarge (8 GPUs) |
| P5e | NVIDIA H200 | — |
| P5en | NVIDIA H200 | p5en.48xlarge (8 GPUs) |
AWS P5 Pricing (May 2026)
| Provider | SKU | GPUs | Memory per GPU | Hourly price | Price per-GPU |
|---|---|---|---|---|---|
| AWS | P5 p5en.48xlarge | 8 × H200 | 80 GB | $63.30-$79.12** | $7.91-$9.89 |
| AWS | P5 p5.48xlarge | 8 × H100 | 80 GB | $55.04-$68.80** | $6.88-$8.60 |
| AWS | P5 p5.4xlarge | 1 × H100 | 80 GB | $6.88 | $6.88 |
| AWS | P4de.24xlarge* | 8 × A100 | 80 GB | $27.45 | $3.43 |
| AWS | P4d.24xlarge* | 8 × A100 | 40 GB | $21.96 | $2.75 |
| Thunder Compute | H100 80 GB - Prototyping | 1-2 × H100 | 80 GB | $1.38-$2.76 | $1.38 |
| Thunder Compute | H100 80 GB - Production | 1-8 × H100 | 80 GB | $2.49-$19.92 | $2.49 |
| *AWS P4 instances added for reference. Full A100 Pricing Comparison. **AWS pricing varies by region and is subject to quotas. Last update: May 12. |
|||||
Specs Comparison
| Feature | AWS P5 | AWS P5en | Thunder Compute |
|---|---|---|---|
| GPUs | 8 × H100 | 8 × H200 | 1 × H100 |
| GPU Memory | 640GB | 1128GB | 80GB |
| System RAM | 2TiB | 2TiB | 32-128GB |
| Storage | 30.7 TB NVMe (ephemeral) | 57.6 TB NVMe (ephemeral) | Persistent block storage |
| Billing | Per-second (60s minimum) | Per-second (60s minimum) | Per minute |
| Storage Cost | EBS $0.08/GB | Same | $0.15/GB/month |
| Snapshots | Standard $0.05/GB | Same | $0.05/GB (compressed) |
| Egress | $0.09/GB (after 100GB) | Same | N/A |
Quick cost calculations
At $1.38 and $6.88 there is a wide pricing difference between Thunder Compute and AWS p5.4xlarge instances.
Thunder Compute is ~5x cheaper
If you don't need P5-level networking, Thunder Compute is cheaper by a large margin for nearly all fine-tuning, eval, and model serving tasks.
Storage and Persistence
AWS uses ephemeral storage; persistence requires EBS volumes ($0.08/GB/mo), snapshots, and egress costs.
Thunder Compute is persistent by default. You can stop and resume without data loss, with one-click snapshots and spec changes.
Billing and Availability
AWS bills per-second with a 60-second minimum, and many GPU families (like P5) require a quota request (see AWS quotas).
Thunder Compute bills per minute and has no minimums, no quotas, and no lock-in.
As shown on this chart, Thunder Compute is cheaper across the board given AWS' 60 second minimum and its premium price.
When to Choose Each
Use AWS P5 if:
<ul><li>You need UltraCluster networking</li><li>You're running massively parallel H100 jobs</li><li>You require tight AWS service integration</li></ul>
Use Thunder Compute if:
<ul><li>You want the cheapest A100 80 GB</li><li>You need persistent storage and snapshots</li><li>You want no lock-in and a simple interface</li></ul>
Get Started
<ul><li>Spin up an H100 for $1.38/hr with persistent storage and per-second billing on <a href="https://www.thundercompute.com/pricing" rel="noopener" target="_blank">Thunder Compute</a>.</li><li>If needed, you can migrate to <a href="https://aws.amazon.com/ec2/instance-types/p5/" rel="noopener nofollow" target="_blank">AWS P5</a> later for high-scale training.</li></ul>
Last Thoughts on AWS P5 Pricing
AWS P5 instances are built for large-scale training. That makes them hard to beat if you need 8-GPU H100 or H200 clusters with ultra-fast networking. But that power comes with complexity: quota requests, ephemeral storage, and extra costs for EBS, snapshots, and egress add up fast.
For most fine-tuning, eval, and inference workloads, Thunder Compute covers the same GPU hardware at a fraction of the cost, with none of the setup friction. Start there and scale to AWS if you outgrow it.