✦ PROVIDER COMPARISON

Which Cloud GPU Provider Is Cheapest in 2026? Real Data From 7 Providers

GPU cloud costs vary wildly — the same H100 can cost 2-3× more depending on which provider you choose. We've been collecting spot prices every 30 minutes from seven providers since March 2026. This is what real 30-day data shows about who's actually cheapest.

Provider Pricing Comparison: H100, A100, L40S

The table below shows 30-day average spot prices per GPU model across all seven providers. ★ marks the cheapest provider for each GPU. Stability rating reflects price volatility (coefficient of variation — lower means more predictable pricing).

Source: RoofRun price_snapshots, 30-day window ending 2026-06-03. ★ = cheapest for that GPU model. Raw data available via API.

Key Findings

On-demand pricing tells a different story — spot prices can be 40-70% cheaper than on-demand rates but come with interruption risk. For fault-tolerant workloads (training with checkpointing, batch inference, data preprocessing), spot instances offer substantial savings. For latency-sensitive or stateful workloads, on-demand is safer.

Provider-by-Provider Breakdown

Specialist Providers vs. Hyperscalers

The GPU-specialist providers — RunPod, Vast.ai, and Lambda Cloud — consistently undercut the hyperscalers on spot pricing. They run leaner operations focused exclusively on GPU compute, which shows up in the numbers.

The hyperscalers (AWS, GCP, Azure) charge a premium but offer integrations, compliance, and SLA guarantees that matter for enterprise workloads. CoreWeave sits between these categories — GPU-native like the specialists, but with enterprise-grade infrastructure and network.

Spot vs. On-Demand: When Each Makes Sense

Spot pricing introduces interruption risk — your instance can be terminated with little notice when the provider needs capacity. The economics only make sense when your workload can handle it:

• Good for spot: ML training with regular checkpointing, batch inference jobs, data pipeline processing, hyperparameter sweeps
• Avoid spot: Real-time inference APIs, stateful applications, workloads without restart logic, jobs shorter than 30 minutes
• Hybrid approach: Run on-demand for the primary model endpoint, use spot for background retraining and eval jobs

Price Stability Matters as Much as Price Level

A provider with highly variable spot pricing creates planning challenges — budgets blow out when prices spike, and automation built around a "typical" price breaks. Our stability rating (coefficient of variation) measures this: a score under 10% means pricing is predictable; above 25% means expect surprises.

Check the 30-day trends page for per-provider, per-GPU volatility charts and stability ratings across all GPU models we track.

How to Minimize GPU Spot Costs

1. Compare before you launch. Use RoofRun's live comparison table to check current spot prices across all 7 providers. Prices update every 30 minutes.
2. Set price alerts. Configure threshold alerts to get notified when a GPU hits your target price — don't manually refresh dashboards.
3. Consider the full cost. Egress fees, storage, and networking can offset compute savings. Providers with cheaper GPUs sometimes charge more for data transfer.
4. Use multi-provider arbitrage. For large training runs, start on the cheapest provider, checkpoint frequently, and be ready to resume on a different provider if interrupted.
5. Watch weekend patterns. Our data consistently shows lower spot prices on weekends when enterprise demand drops. For time-flexible workloads, scheduling for Saturday/Sunday saves 10-20% on average.

GPU Model Selection

Not every workload needs an H100. The A100 and L40S often deliver 80-90% of H100 performance for inference and fine-tuning at significantly lower cost. The A10G and T4 are worth considering for smaller models or high-throughput inference where batch sizes fit.

See our H100 pricing page and A100 pricing page for current live prices and historical trends per GPU model.

Methodology

All pricing data in this article comes from RoofRun's own polling infrastructure. We query provider APIs and pricing pages every 30 minutes and store results in price_snapshots. Averages reflect the trailing 30 days ending 2026-06-03. Providers with no spot pricing for a GPU model show — in the table. Access raw data via the public JSON API.