Takeaways
- For on-demand A100s without lock-in, Thunder Compute is the cheapest mainstream option at $1.09/GPU-hr.
- If you want serverless inference and a large community marketplace, Runpod is strong (A100 Serverless ~$2.17/GPU-hr Active).
- For large clusters or H100/H200 access, consider Lambda, CoreWeave, or Oracle Cloud (OCI lists H100 at $10/GPU-hr).
- Paperspace now lists A100 at $3.18/hr and H100 at $5.95/hr, so it sits above Thunder Compute on raw on-demand A100 and H100 cost.
Quick Comparison (On-Demand or List Price in the U.S., Where Available)
Prices move fast. Always check each provider's page before you launch.
| Provider | A100 Price | H100 Price | Contracts / Notes | Best For |
|---|---|---|---|---|
| Thunder Compute | $1.09/hr (A100 80GB) | $2.19/hr | - True pay-as-you-go. - Per-minute billing. |
- Lowest on-demand cost for experiments, fine-tuning, and dev. |
| Paperspace (DigitalOcean) | $3.18/hr (A100 80GB) | $5.95/hr | - Managed notebook workflow and Gradient ecosystem. | - Teams already on Gradient. - Managed notebook workflows. |
| Runpod | Community: A100 80GB often near $1.39/hr | H100 Serverless ~$3.35/hr Active. | - Serverless inference. - Fast scale-out. - Big GPU marketplace. |
- Scalable inference. - Marketplace variety. |
| Lambda | $2.79/hr (A100 80GB) | H100 $3.99/hr on-demand. | - Lower with reservations. - Clusters available. |
- Research teams that want simple, reliable deep-learning stacks. |
| Vast.ai | ~$1.94/hr (A100 80GB) | Varies by host | - Spot-style marketplace. - Reliability varies. |
- Lowest possible rates if you can tolerate variability. |
| Oracle Cloud (OCI) | A100 80GB in 8-GPU node | $10.00/GPU-hr (H100) list price. | - Multi-GPU bare-metal shapes. - Strong HPC networking. |
- Enterprises, fixed-price clusters, hybrid with free egress. |
| Google Cloud | $5.07/hr (A100 80GB) | $11.06/hr (H100 80GB) | - Per-GPU list prices. - VM cost extra. |
- Teams on GCP wanting Vertex AI + TPUs. |
| AWS | $3.43/hr (A100 80GB) | $6.88/hr (H100 80GB) | - Buy time windows ahead of use. - traditional on-demand varies by region. |
- Production scale with AWS integrations. |
| Azure | $4.41/hr (A100 80GB) | NC H100 v5 single-GPU $8.30/hr (guidance). | - Complex but broad enterprise coverage. | - Microsoft-centric orgs, compliance, global presence. |
| Nebius AI | — | H100 $3.85/hr; H200 $4.50/hr | - Also lists GB200 and B200. | - European teams. - Modern NVIDIA lineup. |
Why Teams Look for a Paperspace Alternative
- Pricing clarity. Paperspace's public pricing now shows A100 at $3.18/hr and H100 at $5.95/hr. That is still well above Thunder Compute on raw on-demand GPU cost, so it is worth deciding whether Gradient's managed notebook workflow is worth the premium for your team.
- Regions and availability. Paperspace operates three datacenter regions (NY2, CA1, AMS1). If your users or data are elsewhere, latency and quotas can bite. DigitalOcean confirms the region count and lists which GPUs are in each site. See the regional availability doc.
- Modern GPU access. If you need H100/H200 or multi-GPU NVLink clusters, several providers now publish lower per-GPU list prices or offer simpler scale-out paths.
How to Choose (Fast)
- Estimate hours. Training + fine-tuning + eval + retries.
- Map GPU need. A100 is still excellent; H100 is ~2-3× faster on many LLM workloads and may be cheaper on a time-to-result basis even at a higher $/hr.
- Decide control vs. convenience. VMs give full control. Serverless removes idle cost and handles autoscaling but limits customization.
- Check the extras. Storage, egress, snapshot pricing, per-minute billing, pause/hibernate, NVLink/InfiniBand, support SLAs.
- Avoid surprise commitments. Verify whether the price assumes reserved terms or a monthly plan.
The Best Paperspace Alternatives (Details)
1. Thunder Compute - Lowest On-Demand A100 Price, No Lock-In
- Price highlights: A100 80GB $1.09/hr, H100 $2.19/hr. True pay-as-you-go with per-minute billing to trim idle time. Thunder Compute pricing.
- Best for: Experiments, fine-tuning, and small/medium inference where cash outlay and predictable on-demand costs matter.
- Why it beats Paperspace for budgets: Thunder Compute still delivers the lower raw on-demand A100 and H100 price, so teams that do not need Gradient-specific workflow features can usually stretch their budget further on Thunder Compute.
2. Runpod - Serverless Inference + Big Marketplace
- Price highlights: Serverless A100 ~$2.17/hr Active (or ~$2.72/hr Flex), H100 ~$3.35/hr Active. Community Pods frequently list A100 80GB near ~$1.39/hr, with Secure Cloud priced higher. Runpod pricing · A100 comparison.
- Best for: Fast deployments, autoscaling inference, and teams that don't want to manage VM lifecycles.
- Watch-outs: Marketplace hosts vary; read specs and ratings. Serverless is priced per-second and by worker type; model throughput, cold-start, and concurrency determine the true TCO.
3. Lambda - Research-Friendly, Simple Stack
- Price highlights: Public materials show A100 80GB $2.79/hr and H100 $3.99/hr on-demand; commitments may be lower. Lambda pricing.
- Best for: Teams that want reliable hardware, solid images, and optional large clusters without hyperscaler complexity.
4. Vast.ai - Lowest Prices if You Can Tolerate Variability
- Price highlights: Marketplace pricing often shows A100 near $1.94/hr and many consumer RTX cards at pennies per minute. Vast.ai.
- Best for: Lowest possible cost, preemptible/interruptible workloads, non-critical runs.
- Watch-outs: Reliability and bandwidth vary by host; plan for checkpoints and migration.
5. Oracle Cloud (OCI) - Straightforward List Pricing for H100/H200
- Price highlights: Official list price for H100 is $10.00 per GPU-hour; H200 is also listed at $10.00 per GPU-hour. Shapes are multi-GPU bare-metal with high-bandwidth RDMA. OCI price list · GPU pricing.
- Best for: Enterprises that value predictable list pricing, big clusters, and free egress.
6. Google Cloud - Broad Services + TPUs
- Price highlights: Transparent per-GPU pricing (for example, A100 80GB $5.07/hr and H100 $11.06/hr). VM cost is additional. GCP GPU pricing.
- Best for: Teams already invested in Vertex AI/BigQuery; TPU access.
7. AWS - Scale and Ecosystem
- Price highlights: p5.48xlarge (8× H100) listed at $31.464/hr in us-west-2 when purchased as Capacity Blocks for ML (≈ $3.93/GPU-hr). Traditional on-demand pricing varies by region and purchasing model; spot can be significantly lower. Capacity Blocks pricing.
- Best for: Enterprise production, security/compliance, global reach.
8. Azure - Single-GPU H100 VMs
- Price highlights: Public guidance around $8.30/hr for single-GPU NC H100 v5 VMs in U.S. regions. Check the regional calculator for exact rates. Azure pricing.
- Best for: Microsoft-centric orgs and Windows workflows.
9. Nebius AI - Modern NVIDIA Lineup, EU Focus
- Price highlights: H100 $3.85/hr, H200 $4.50/hr, B200 $7.15/hr (per GPU-hr). Nebius pricing.
- Best for: European teams wanting current-gen parts with clear per-GPU pricing.
11. Paperspace (DigitalOcean) - When It Still Makes Sense
- When to pick it: You're committed to the Gradient ecosystem, want managed notebooks/workflows, or can fully utilize reserved pricing.
- Key numbers to know: Current listed pricing is $3.18/hr for A100 and $5.95/hr for H100. Paperspace still makes the most sense if you want the managed notebook workflow and Gradient tooling, rather than the absolute lowest raw GPU price. Regions: NY2, CA1, AMS1. See the official Paperspace pricing and DigitalOcean docs plus regional availability.
Methodology & Update Cadence
- Prices were checked against public pricing pages and official documentation for this (July 2026) snapshot. Some providers publish only ranges; marketplace rates fluctuate. When providers only sell multi-GPU nodes, we note total node price and/or normalize to per-GPU where reasonable.
- Always confirm prices in your target region and add storage, data egress, and VM costs where applicable.
Thinking About Switching?
Spin up an A100 on Thunder Compute in minutes and benchmark your workload. If it finishes in half the cost, keep it. If not, you've validated your path with real numbers.