Takeaways
- Cost: Thunder Compute is ~4.4 times cheaper per GPU-hour.
- Billing: Thunder Compute bills per second; Azure rounds up to whole minutes.
- Storage: Thunder Compute persistent volumes start at $0.15/GB/mo; Azure Premium SSD is $0.18/GB/mo.
- Ease of use: One-click VS Code workspace, no quota tickets, swap CPU/RAM on the fly.
1. Raw GPU pricing
| Provider | SKU | GPUs | On-demand $/hr | Billing granularity | 100 GPU-hr cost |
|---|---|---|---|---|---|
| Azure | Standard ND24ads A100 v4 | 1× A100 80 GB | $3.67 | per-minute (rounded) - see Azure VM pricing FAQ | $367.00 |
| Thunder Compute | A100 80 GB | 1× A100 80 GB | $0.78 | true per-minute | $78.00 |
Azure Spot for the same SKU averages $1.145/hr (Vantage), but interruptions revoke your GPU with <30 s notice—unsuitable for fine-tuning runs that aren't checkpoint-safe.
2. Hidden line items
| Cost item | Azure | Thunder Compute |
|---|---|---|
| Persistent storage | Premium SSD $0.18/GB/mo | $0.15/GB/mo |
| Snapshots | Incremental snapshots billed separately | Snapshots built-in, same $0.15/GB rate |
| Data egress | from US regions $0.087/GB after first 100 GB | First 3 TB per month free, then same $0.087/GB¹ |
¹Thunder matches Azure's public egress rate only beyond the free tier.
3. Quotas and setup friction
- New Azure subscriptions start with 0 N-series vCPUs; you must file a GPU quota increase request.
- Availability is region-limited—East US 2 and West Europe frequently show capacity errors.
- Thunder Compute accounts spin up instantly with no GPU quota tickets.
4. Developer experience
| Feature | Azure NC A100 | Thunder Compute |
|---|---|---|
| IDE integration | Use VS Code Remote-SSH or Azure ML Studio | One-click VS Code in browser |
| Instance re-sizing | Must redeploy VM; storage detach/attach required | Live RAM, vCPU, and disk tweaks without reboot |
| API | Azure Resource Manager | Simple REST & Terraform module |
| Pricing transparency | Calculator + separate disk/network pricing | All-in dashboard; per-minute billing |
5. When is Azure still a fit?
- Multi-GPU scale (>8 A100s) in a single VM via NC96ads A100 v4.
- Co-locating with other Azure PaaS services for low-latency data pipelines.
For individual fine-tunes, model inference endpoints, and bursty hackathon workloads, Thunder Compute's pay-as-you-go model wins on both price and friction.
Conclusion
If you need a single A100 80 GB quickly, Thunder Compute costs $262 less per 100 GPU-hours and bills per minute of usage. Unless you're already locked into Azure's ecosystem or require massive multi-GPU nodes, switching cuts spend and setup time. To see how the legacy giants are reacting to the rise of GPU-only clouds, check out our deep dive: What is a Neocloud? The Rise of GPU-Only Clouds.