GPU Price History Charts & Trends — H100, A100 & RTX Daily Rates Across 91+ Providers
H100 Price Trends
The H100 SXM5 80GB remains the dominant training GPU. On-demand rates have stabilized in the $2.49–$3.50/hr range across specialist clouds, with hyperscalers (AWS p5, Google A3) running 2–3× higher. Spot availability is limited — most H100 capacity is reserved or committed.
A100 Price Trends
A100 80GB on-demand rates have declined as H100 supply expanded. Prices now range from $1.29–$2.49/hr, making the A100 the value choice for inference workloads that don't require H100 bandwidth. Spot A100s are more available than H100s.
RTX 4090 Price Trends
RTX 4090 cloud pricing has remained relatively stable at $0.34–$0.79/hr since mid-2024, making it the most cost-effective option for 7B–13B model inference and fine-tuning. Unlike data center GPUs, RTX 4090 supply is constrained by consumer GPU production cycles rather than hyperscaler procurement, so prices have not fallen as sharply as H100 or A100. The 24GB GDDR6X VRAM fits most 7B models at FP16 and 13B models at INT8, and the 1,008 GB/s bandwidth is competitive with older A100 40GB configurations. RunPod community cloud and Vast.ai marketplace consistently offer the lowest RTX 4090 rates — often $0.34–$0.44/hr for spot instances.
What Drives GPU Price Changes
GPU cloud prices are driven by supply (new data center capacity, chip availability), demand (AI training cycles, inference scaling), and competition between providers. Specialist clouds respond faster to market conditions than hyperscalers, whose prices change quarterly at most.
GPU price forecast: H2 2026 and 2027
H100 on-demand prices are projected to continue declining through H2 2026 and into 2027. As NVIDIA Blackwell (B200) supply increases and H100 transitions to second-generation status, specialist providers are expected to offer H100 80GB SXM at $1.99–$2.49/hr by end of 2026 and $1.49–$1.99/hr by mid-2027. Spot rates may fall below $1.00/hr on marketplace platforms as H100 inventory grows relative to demand. A100 80GB is projected to reach $0.99–$1.29/hr on-demand by end of 2026 as it is further displaced by H100. These projections assume continued supply expansion and no major demand shock — a large-scale AI training run or enterprise adoption wave could temporarily reverse the trend. Use the price history charts above to track actual rates against these projections.
How to use GPU price history for procurement decisions
The 90-day price history charts are most useful for three procurement decisions. First, timing on-demand purchases: if a GPU's price has been declining steadily over 30 days, waiting 2–4 weeks before committing to a long-running job may save 5–15%. Second, evaluating reserved instance commitments: if a provider's on-demand rate has been stable for 60+ days, a 3-month or 6-month reserved commitment at a 20–30% discount is likely to pay off. Third, identifying spot price windows: spot prices often dip during weekends and off-peak hours — the history charts reveal these patterns, letting you schedule batch training jobs during low-cost windows. Filter by GPU model and provider to isolate the specific price series relevant to your workload.
Comparing hyperscalers vs specialist providers over time
The price history data reveals a persistent and widening gap between hyperscaler (AWS, Google Cloud, Azure) and specialist provider (Lambda Labs, CoreWeave, RunPod, Fluidstack) GPU rates. In early 2024, AWS p5 H100 on-demand was roughly 2.5× more expensive than Lambda Labs. By mid-2026, that gap has grown to 2.4× ($5.89/hr vs $2.49/hr). Hyperscalers compete on ecosystem integration, compliance certifications, and SLA guarantees rather than raw GPU price — their rates have declined more slowly because enterprise customers pay a premium for managed networking, IAM, and support. For pure compute workloads without hyperscaler ecosystem dependencies, specialist providers consistently offer 40–60% lower rates for equivalent hardware. The history charts make this gap visible over time, helping teams quantify the cost of staying on a hyperscaler versus migrating to a specialist provider.
Export GPU price history
Download the full time-series dataset as CSV or JSON. Each row is one (provider, GPU model, date, price) observation. Includes an observation_class column — points are either observed (live API price) or not_observed (estimate). Add ?class=observed to the API URL for measured data only.
History retained permanently · Updated daily · CC BY 4.0 license · Methodology