GPU Spot Pricing Analysis: 44 Days of Data Across 7 Cloud Providers

The Dataset

Starting in late March 2026, RoofRun began polling GPU spot and on-demand prices across all major cloud providers every 30 minutes. We don't sample — we collect every visible listing for every GPU type across every region we can reach. The result is a dense time-series dataset that lets us answer questions that require sustained observation rather than spot-checks.

Providers covered: AWS, GCP, Azure, Lambda Cloud, CoreWeave, RunPod, and Vast.ai. GPU models range from the NVIDIA H200 (141GB HBM3e) down to the T4 (16GB). Each snapshot records whether a GPU is available, its spot price, on-demand price, region, and instance type — giving us full context for every data point.

GPU Price Ranges: 89 Days of Min/Max/Avg

Price volatility is the defining characteristic of the GPU spot market. The spread between a GPU's minimum recorded price and its maximum isn't a rounding error — it's frequently 40-70%. Understanding these ranges is the foundation of any cost optimization strategy.

The H100 SXM consistently shows the widest price spread — the most expensive provider can charge 3x what the cheapest charges for equivalent compute. The T4 and A10G show tighter ranges, partly because they're cheaper commodities with more competition on the lower end of the market.

Provider Comparison: Who Has the Cheapest GPUs

Not all GPU clouds are created equal — and price is only part of the story. Availability, consistency, and what "availability" even means varies significantly across providers. Here's what our data shows:

Provider Profiles

• AWS: Highest prices, most consistent availability. Best for regulated workloads or teams already embedded in the AWS ecosystem. Spot interruption rates are real — use checkpointing.
• GCP: Strong H100 availability via TPU alternatives and A100 pools. Preemptible VMs have 24-hour max runtimes. Good for batch ML training.
• Azure: ND-series H100 instances are priced competitively when available. Availability is spottier than AWS/GCP in our dataset.
• Lambda Cloud: No preemption — reserved rates look like "spot" but aren't interruptible. Excellent for long-running training runs where you can't afford interruption.
• CoreWeave: Premium H100 and A100 supply with custom networking. Higher prices justified for inference workloads with low-latency requirements.
• RunPod: Best price-per-FLOP for A100 and L40S in our dataset. Community cloud — heterogeneous hardware, suitable for most ML training workloads.
• Vast.ai: Cheapest absolute prices on RTX 4090 and consumer-grade GPUs. Peer-to-peer marketplace model means availability is highly variable. Ideal for small experiments and inference.

For a live price comparison, see the GPU comparison table or individual GPU pages like H100 pricing.

Weekend Pricing Patterns

One of the most actionable patterns in our dataset is the consistent weekend price dip. Corporate AI workloads follow business hours — when enterprise teams shut down on Friday afternoon, spot availability increases and prices drop.

The pattern is consistent across providers but varies in magnitude. AWS shows the most pronounced weekend dips — reflecting its corporate customer base. Lambda and CoreWeave show less weekend variation, partly because their customers are research labs with continuous workloads.

Time-of-Day Patterns

Beyond weekend dips, we see a predictable intraday cycle. UTC 02:00–08:00 (overnight US time) consistently shows lower H100 prices on US-region instances. European and Asian business hours add demand that pushes prices up. For teams with workload flexibility, overnight US launches are a reliable second-order savings lever.

See real-time and historical pricing trends on the Trends page.

Cost Optimization: What the Data Actually Supports

A lot of GPU cost advice is hypothetical. Here's what our 44-day dataset actually supports:

• Checkpoint every hour, minimum. The expected cost of an interruption is (checkpoint_frequency × hourly_rate × recovery_overhead). At $2/hr H100 pricing, losing 4 hours of work to interruption costs the same as 4 extra hours of compute. Checkpointing hourly caps your loss at 1 hour of work.
• Don't benchmark on peak-demand windows. If you benchmark cost at 2pm Monday UTC and then schedule your production run for the same window, you'll consistently overpay. Our data shows 15-30% price variation within the same week for the same GPU on the same provider.
• Multi-provider arbitrage is real but operationally expensive. The price spread between cheapest and most expensive H100 access in our dataset frequently exceeds 50%. But moving workloads across providers requires engineering work — different APIs, different checkpoint formats, different networking. Calculate your engineering cost before assuming arbitrage savings are free money.
• Smaller GPUs are underrated for inference. A100 and H100 get all the attention, but L40S and A10G in our dataset show excellent price-per-token performance for inference workloads that don't require 80GB VRAM. The L40S is consistently 40-60% cheaper than H100 with 80% of the throughput for many transformer inference workloads.
• Set price alerts, don't poll manually. Spot prices move fast — a window below your target price can open and close in under an hour. RoofRun alerts notify you when prices hit your threshold so you can launch immediately rather than checking a dashboard every few hours.
• Vast.ai for experiments, Lambda/CoreWeave for production. The cheapest absolute prices in our dataset consistently come from Vast.ai's peer-to-peer marketplace. For exploratory work where you can tolerate variable hardware, it's unbeatable. For production inference or critical training runs, the reliability premium from Lambda or CoreWeave is worth paying.