GPU Spot Price Trends —
30-Day Historical Data Across 7 Providers

What GPU Pricing Trends Tell Us

GPU spot pricing isn't static — it moves with supply, demand, and the operational rhythms of each cloud provider. After 30 days of continuous collection, several patterns emerge that fundamentally change how you should be sourcing compute.

Weekends and off-peak hours are cheaper. Enterprise AI training runs on business schedules. When US-based teams wrap on Friday afternoon, spot inventory opens up — and prices drop. If your workload is asynchronous (batch inference, fine-tuning, data pipelines), scheduling for Saturday mornings can cut your compute bill by 15–30% with zero compromise on capability.

Provider floors matter more than averages. The 30-day low is your actual achievable price if you're patient and can tolerate interruptions. The gap between 30d low and 30d high on H100s and A100s regularly spans $1–3/hr — a meaningful delta when you're running hundreds of GPU-hours per week.

Smaller providers are more volatile, but cheaper on the floor. RunPod and Vast.ai show wider price swings than AWS or Azure — but their 30-day lows are often 40–60% below hyperscaler spot prices. If your workload can tolerate interruption, chasing floor prices on smaller providers is the highest-leverage cost optimization available today.

H100 and A100 trend differently. H100 availability is still constrained globally, which keeps its price floor elevated relative to older architectures. A100s — despite being the previous generation — often show more stable pricing because the install base is larger and inventory is better distributed. For long-running training jobs where you need price certainty, A100s frequently win on total cost.

Seasonal patterns are real. Q4 sees tighter GPU supply as cloud providers lock capacity for holiday-season inference demand. Q1 often sees prices soften as hyperscalers provision new capacity from prior-year procurement. The best time to lock long-term contracts is January through March.