Market insights
AWS P5 vs Thunder Compute: Pricing, Specs, and the Cheapest Way to Train (December 2025)
September 16, 2025
12 mins read
TL;DR
- Need a massive, multi-node H100 cluster? AWS P5 gives you 8× H100s per node, 3,200 Gbps networking, and 30.7 TB NVMe. Price: ~$55–66/hr depending on region.
- Want the cheapest, no-commitment way to fine-tune or serve models? Thunder Compute offers A100 80 GB for $0.78/hr, billed per second with persistent storage and one-click VS Code.
- AWS adds separate costs for storage and data egress — up to 0.08/GB/month for EBS and 0.09/GB for egress after 100 GB.
- AWS requires quotas and has a 60-second billing minimum. Thunder Compute doesn’t.
What each platform is best at
- AWS P5: Multi-node H100 clusters, ultra-fast networking, and deep AWS integration.
- Thunder Compute: Low-cost A100 80 GB with persistent storage, per-second billing, snapshots, and fast setup — perfect for prototyping, fine-tuning, and inference.
Pricing Snapshot (December 2025)
| Provider | SKU | GPUs | Memory per GPU | Hourly price (node) | Approx per-GPU |
|---|---|---|---|---|---|
| AWS | P5 p5.48xlarge | 8 × H100 | 80 GB | ~$55–66 | ~$6.8–8.25 |
| AWS | P4de.24xlarge | 8 × A100 | 80 GB | $40.96 | $5.12 |
| AWS | P4d.24xlarge | 8 × A100 | 40 GB | $32.77 | $4.10 |
| Thunder Compute | A100 80 GB | 1 × A100 | 80 GB | $0.78 | $0.78 |
Note: AWS pricing varies by region and is subject to quotas. Thunder Compute pricing is flat and usage-based.
Specs Comparison
| AWS P5 | AWS P4de | Thunder Compute | |
|---|---|---|---|
| GPUs | 8 × H100 | 8 × A100 | 1 × A100 |
| GPU Memory | 640 GB | 640 GB | 80 GB |
| System RAM | 2 TiB | 1.125 TiB | 32 GB (customizable) |
| Storage | 30.7 TB NVMe (ephemeral) | 8 TB NVMe (ephemeral) | Persistent block storage |
| Billing | Per-second (60s min) | Per-second (60s min) | True per-second |
| Storage Cost | EBS $0.08/GB | Same | $0.15/GB/month |
| Snapshots | Standard $0.05/GB, Archive $0.0125/GB | Same | $0.075/GB (compressed) |
| Egress | First 100GB free, then $0.09/GB | Same | Custom / bring-your-own CDN |
Quick cost calculations
- 8× A100 80 GB Equivalent
- Thunder Compute: 8 × $0.78 = $6.24/hr
- AWS P4de: $40.96/hr
- Savings: Thunder is ~6.6x cheaper
- 8× H100
- AWS P5: ~$55–66/hr
- Thunder H100s (coming soon): $1.47/hr × 8 = $11.76/hr
If you don't need P5-level networking, Thunder is cheaper by a large margin for nearly all fine-tuning, eval, and model serving tasks.
Storage and Persistence
- AWS: Ephemeral storage. To persist data, you’ll need to add EBS volumes (0.08/GB/mo), pay for snapshots, and budget for data egress.
- Thunder Compute: Persistent storage by default. You can stop and resume without data loss. Snapshots and spec changes (RAM, disk, vCPU) are one click.
Billing and Availability
- AWS: Per-second billing, but with a 60-second minimum. Many GPU families (like P5) require a quota request before use. See AWS quotas.
- Thunder Compute: No quotas. Start instantly. Full per-second billing with no minimums and no lock-in.
When to Choose Each
Use AWS P5 if:
- You need UltraCluster networking
- You're running massively parallel H100 jobs
- You require tight AWS service integration
Use Thunder Compute if:
- You want the cheapest A100 80 GB
- You need persistent storage and snapshots
- You want no lock-in and a simple interface
Get Started
- Spin up an A100 80 GB at $0.78/hr with persistent storage and per-second billing on Thunder Compute.
- If needed, migrate to AWS P5 later for high-scale training.