Spot GPU Pricing: How to Get H100 at $0.99/hr (2026 Guide)

If you're training or inferencing on NVIDIA H100 GPUs in 2026, the difference between spot and on-demand pricing is the single biggest lever you have to control costs. We're talking $0.99/hr vs $8.50/hr for the same hardware — an 88% savings that can mean the difference between a viable startup and a burned-through runway.

This guide breaks down exactly what spot GPU pricing is, which providers offer the best rates right now, and how to build a workload strategy that maximises savings without sacrificing reliability.

What Is Spot GPU Pricing?

Spot (or "interruptible") instances let you bid on unused GPU capacity that cloud providers would otherwise leave idle. The trade-off: the provider can reclaim the instance with little notice (typically 30 seconds to 2 minutes) if someone pays the on-demand rate and capacity gets tight.

This makes spot ideal for:

• Training jobs with checkpointing — you resume from the last save
• Batch inference — no strict latency requirements
• Experimentation and hyperparameter sweeps — fault-tolerant by nature
• CI/CD pipelines — build and test runners that can restart

On-demand, by contrast, guarantees the instance until you terminate it. You pay a premium for that certainty — often 2-5x the spot rate.

H100 Spot vs On-Demand: Full Pricing Table (July 2026)

We pulled real-time pricing from Parallel API and provider dashboards. All prices are per hour for a single NVIDIA H100 (80GB).

Prices captured July 29, 2026 via Parallel API and direct provider listings. Spot prices fluctuate — check live data before provisioning.

Provider Deep Dive

🥇 Vast.ai & Spheron — $0.99/hr (Best Value)

Both marketplace-style platforms aggregate individual GPU owners alongside datacenter capacity. This creates a deep pool of spot inventory that keeps prices competitive. Vast.ai has the broadest selection (H100, A100, A6000, RTX 4090), while Spheron focuses on a curated set of verified hosts. At $0.99/hr for H100 spot, they're the cheapest options by a wide margin. The catch: interrupt frequency is higher, so you must implement checkpointing. Use their Python SDKs or Docker-based workflows for seamless resumption.

🥈 Nebius — $1.80/hr (Best Balance)

Formerly known as Selectel's AI cloud, Nebius offers strong reliability at a reasonable spot price. Their $1.80/hr spot with only 44% savings over on-demand ($3.20/hr) means less upside — but also lower interrupt risk. If you're running production-adjacent workloads that can't tolerate frequent restarts but don't need full on-demand guarantees, Nebius is the sweet spot.

🥉 CoreWeave — $2.10/hr (Enterprise Ready)

CoreWeave is built on Kubernetes-native infrastructure and offers direct NVIDIA networking (InfiniBand) for multi-node training. Their spot at $2.10/hr and on-demand at $3.30/hr both include this high-performance networking — something the marketplace providers can't always guarantee. If you need multi-GPU training across nodes, factor in the networking advantage; the effective price-per-performance may beat the cheaper options.

When Should You Pay On-Demand?

Spot pricing is compelling, but on-demand has its place. Here's when paying full price makes sense:

• Production inference serving — your users can't wait for a spot reclaim + cold start
• Long-running fine-tunes without checkpoint infrastructure
• Demo or client-facing environments where reliability trumps cost
• Burst capacity on AWS/GCP/Azure when you need guaranteed hours during a launch

The smartest strategy? Hybrid provisioning. Run training on spot (Vast.ai or Spheron at $0.99/hr), keep production inference on on-demand (CoreWeave at $3.30/hr), and overflow to AWS on-demand ($8.50/hr) only when all other capacity is exhausted. This keeps your blended rate well under $2/hr while maintaining reliability for customer-facing workloads.

How to Get Started with Spot GPUs

If you've been on AWS on-demand and want to switch to spot, here's a practical migration path:

1. Set up checkpointing. Use PyTorch Lightning, Hugging Face Trainer, or Weights & Biases to save checkpoints every 5-10 minutes. This one change unlocks 60-88% savings.
2. Start with Vast.ai or Spheron. Sign up, add funds ($10 minimum on most platforms), and launch your first spot instance. Their UIs show real-time interrupt rates so you can pick stable hosts.
3. Use our comparison tool to check live pricing before each provisioning decision. Spot prices move hourly based on supply and demand.
4. Scale to Nebius or CoreWeave for workloads that need more reliability. Their spot interrupt rates are significantly lower.
5. Monitor your blend. Track effective hourly cost across all providers. A healthy blend is under $2/hr for H100 compute.

The 2026 Spot Market Outlook

Three trends are shaping spot GPU pricing for the rest of 2026:

1. More supply, but more demand. NVIDIA shipped more H100s in Q2 2026 than any previous quarter, but Blackwell B200 allocations are eating into H100 supply. Expect spot prices to stay flat through Q3 and rise 10-15% in Q4 as the Blackwell transition accelerates.

2. Marketplace providers are gaining share. Vast.ai and Spheron have grown their H100 fleets by 3x since January. More supply means more competitive pricing — the $0.99 floor may hold through year-end.

3. Hyperscaler spot is becoming less attractive. AWS spot on H100 went from $3.50 to $5.00 in 2026 as the big cloud providers reprioritise on-demand margin. The gap between hyperscaler spot and marketplace spot has never been wider.

Frequently Asked Questions

Is spot GPU pricing reliable enough for ML training?

Yes — if you checkpoint. Most modern ML frameworks (PyTorch Lightning, HF Trainer, MosaicML Composer) support automatic checkpointing out of the box. With 5-minute checkpoint intervals, a spot reclaim costs you at most 5 minutes of compute. At $0.99/hr vs $8.50/hr, you can afford a few restarts.

Which provider has the lowest interrupt rate?

CoreWeave and Nebius have the lowest spot interrupt rates among providers we track. AWS spot on H100 has the highest interrupt rate due to competition from on-demand customers. Vast.ai and Spheron are in between — their interrupt rate depends on the specific host you choose (datacenter hosts are more reliable than individual GPU owners).

Can I use spot for multi-GPU training?

Yes, but it's more complex. CoreWeave offers spot with InfiniBand, making it the best option for multi-node training on spot. For Vast.ai and Spheron, single-node multi-GPU (up to 8x H100) is reliable; multi-node requires careful orchestration with a framework like Ray or Horovod.

Prices captured July 29, 2026 via Parallel API, provider dashboards, and marketplace listings. All prices shown are for NVIDIA H100 (80GB) per GPU per hour. Spot prices are dynamic — check providers for current rates. GPUIndia earns commissions through affiliate links at no extra cost to you.