50+ standalone, citable statistics on the cloud GPU rental market — computed from 5,213 live listings across 54 providers and 24 daily snapshots. Every number on this page is sourced directly from public provider feeds. Free to cite, embed, and republish with attribution.
4 statistics
The GPU cloud market tracks 5,213 distinct rental listings as of April 2026.
Each listing represents a unique GPU model + provider + region + commitment-type combination. The largest publicly tracked GPU cloud dataset.
54 cloud providers publicly list GPU pricing.
From hyperscalers (AWS, GCP, Azure, Oracle) to specialty clouds (CoreWeave, Lambda Labs, Crusoe, Nebius) to marketplaces (RunPod, Vast.ai, Verda) to bare-metal providers.
75 distinct GPU models are available in the cloud.
From NVIDIA T4 (2018) through NVIDIA B200 Blackwell (2025), plus AMD MI300X/MI350X, and consumer cards like RTX 4090 and RTX 5090.
GPU cloud instances are available in 130 unique regions worldwide.
Including US East, US West, EU Central, EU North, EU West, APAC (Tokyo, Singapore, Sydney), Latin America, and emerging African markets.
4 statistics
Cloud H100 prices range from $0.80/hr to $97.44/hr — a 122x spread for the same physical chip.
Across 327 tracked H100 listings, the cheapest is on a spot/marketplace provider and the most expensive is a hyperscaler reserved-multi-GPU configuration.
The cheapest publicly listed cloud H100 is $0.80/hr on spot.
Spot/interruptible pricing on a specialty cloud provider. On-demand floor across all 54 tracked providers is roughly 2x that figure.
The median cloud H100 price across all tracked listings is $8.97/hr.
The median is pulled upward by hyperscaler multi-GPU bundle pricing. The 25th-percentile price (a more honest "real-world buyer" number) is much lower.
The 25th-percentile H100 price — what a careful buyer can realistically obtain — is $3.50/hr.
The 25th percentile filters out fleeting spot-market deals and hyperscaler list prices that no one actually pays. This is the most defensible "market price" for cloud H100.
1 statistics
Hyperscaler H100 prices are 99% higher than specialty cloud providers for the same chip.
AWS / GCP / Azure / Oracle median H100 is $13.96/hr. Specialty clouds (CoreWeave, Lambda, Crusoe, RunPod, Nebius, Vast.ai) median is $7.00/hr.
3 statistics
Google Cloud Platform alone accounts for 39.8% of all publicly tracked cloud GPU listings.
2074 GCP instances across all tracked GPU models, spanning A2, G2, A3, A3 Mega, A3 Ultra, and A3 High instance families.
The top 10 cloud providers control 93.7% of all tracked GPU cloud listings.
The remaining 44 providers — including emerging specialty clouds, regional providers, and marketplaces — split the bottom 6.3% of inventory.
Hyperscalers (AWS + GCP + Azure + OCI) hold 69.2% of publicly listed GPU cloud capacity.
Despite charging an average 99% premium for H100 vs specialty clouds, the big-four hyperscalers still control the majority of public listings — likely because their inventory is easier to scrape than private specialty-cloud contracts.
3 statistics
The NVIDIA L40S has a 1391x cloud price spread — the widest of any tracked GPU.
L40S ranges from $0.32/hr spot to $445.25/hr on hyperscaler multi-GPU bundles. The L40S spot market has collapsed faster than any other GPU as inference workloads migrate upstream.
Cloud H200 prices span a 143x range — from $1.19/hr to $169.60/hr.
133 H200 listings tracked across 7+ providers. The spread reflects the gap between specialty cloud spot pricing and hyperscaler reserved capacity.
Cloud B200 (Blackwell) prices span a 53x range — from $1.71/hr to $90.22/hr.
86 B200 listings tracked. Blackwell is the newest generation with the tightest supply; specialty cloud spot pricing has emerged 6 months ahead of hyperscaler general availability.
4 statistics
L40S spot instances cost 82% less than on-demand on average — the largest spot discount of any tracked GPU.
L40S spot averages $2.17/hr vs on-demand $12.05/hr. The L40S has become the spot-market darling for inference workloads.
A100 spot instances are 61% cheaper than on-demand on average.
A100 spot averages $6.56/hr vs on-demand $17.01/hr. The A100 spot market is the hidden gem of GPU cloud pricing for cost-sensitive teams.
H100 spot instances are 40% cheaper than on-demand on average.
H100 spot averages $10.95/hr vs on-demand $18.39/hr across all tracked providers.
40.5% of all publicly listed GPU cloud capacity is spot/interruptible.
2,109 spot instances vs 3,104 on-demand. The spot share has grown faster than on-demand as buyers get more comfortable with checkpointing and fault-tolerant training.
1 statistics
The cheapest cloud A100 is $0.08/hr on spot — the most cost-efficient datacenter GPU per token for LLM inference.
Spot/interruptible A100 80GB on a specialty cloud provider. At this rate, the A100 beats the H100 on cost-per-token for 7B–70B inference workloads under batch size 32.
2 statistics
The median cloud RTX 4090 rental price is $0.60/hr.
18 tracked RTX 4090 listings across specialty clouds. The RTX 4090 is the most rented consumer GPU because it delivers near-A100 inference performance on 7B-class models at 1/4 the cost.
The NVIDIA RTX 5090 is now available across multiple cloud providers, with prices starting at $0.65/hr.
4 RTX 5090 listings tracked — the newest consumer GPU to reach the cloud rental market. With 32 GB VRAM, the RTX 5090 handles Q8-quantized 30B-class models inference workloads previously requiring datacenter cards.
2 statistics
The GPU Tracker H100 Price Index sits at 100 (vs February 22, 2026 = 100) — H100 prices have moved +0.0% over the tracked window.
Calculated from the 25th-percentile cloud H100 price across all tracked providers. Index methodology and full history available at /gpu-price-index.
GPU Tracker captures full price snapshots of the cloud GPU market every 24 hours.
Each snapshot records min, average, 25th percentile, and 75th percentile price for every GPU model with at least one valid listing on that day.
3 statistics
RunPod is the largest specialty GPU cloud by listing count, with 679 tracked instances — more than Lambda Labs, CoreWeave, and Vast.ai combined.
RunPod's marketplace model aggregates third-party compute hosts. Its scale reflects the long-tail of independent GPU operators that don't appear as standalone providers in this dataset.
Vast.ai lists 128 GPU rental instances — the largest peer-to-peer GPU marketplace tracked.
Vast.ai is consistently the cheapest provider for budget-tier and consumer GPU rentals (RTX 3090, RTX 4090, RTX 5090) due to its marketplace structure connecting compute buyers with individual operators.
Lambda Labs maintains 120 publicly listed GPU instances, focused on managed H100/H200/B200 access.
Lambda Labs combines specialty cloud pricing with pre-installed ML stacks (PyTorch, JAX, TensorFlow) and InfiniBand networking on multi-GPU instances. Popular for research and managed training workloads.
Data source. Live cloud GPU pricing is scraped every 6 hours from each provider's public pricing page or API. The dataset covers single-GPU instances by default; multi-GPU instances are normalized to per-GPU pricing for comparability. The full live dataset is published at /gpu-data.json.
Daily snapshots. Once per day at the same UTC time, the live dataset is collapsed into a snapshot recording min, average, 25th percentile (p25), and 75th percentile (p75) for every GPU model with at least one valid listing on that day. The full snapshot history is published at /price-history.json.
Why the 25th percentile? The minimum price often reflects rare spot or marketplace deals that vanish in minutes. The average is dragged upward by hyperscaler list prices that few buyers actually pay. The 25th percentile captures the price a careful buyer can realistically obtain — a more honest signal of where the market is.
Inclusion criteria. Any cloud provider with publicly listed GPU pricing and at least one valid in-stock listing is included. Private enterprise contracts (CoreWeave reserved, AWS Savings Plans, multi-year reserved capacity) are not visible to us and are excluded.
Independence. GPU Tracker takes no payment, sponsorship, or affiliate consideration from any provider in exchange for placement, scoring, or visibility. All statistics on this page reflect the live data, not commercial relationships.
Every statistic on this page is licensed under Creative Commons BY 4.0. You're free to quote, screenshot, or republish any number on this page in articles, blog posts, research papers, presentations, or AI training datasets. The only request: include attribution to GPU Tracker with a link back to this page.
GPU Tracker (2026). GPU Cloud Statistics 2026. Retrieved from https://gputracker.dev/gpu-cloud-statistics-2026Want the raw data instead? Download /gpu-data.json (live snapshot) or /price-history.json (full archive).
GPU Tracker monitors 54 cloud GPU providers as of April 2026, spanning hyperscalers (AWS, Google Cloud, Microsoft Azure, Oracle Cloud), specialty AI clouds (CoreWeave, Lambda Labs, Crusoe, Nebius, Hyperstack, Together AI), marketplaces (RunPod, Vast.ai, Verda, Shadeform), bare-metal providers (Latitude.sh, OVHcloud, Scaleway), and regional clouds across North America, Europe, and Asia-Pacific.
Live pricing across all 54 tracked providers is scraped every 6 hours. A full daily snapshot — capturing min, average, 25th percentile, and 75th percentile prices for every GPU model with at least one valid listing — is recorded once per day at the same UTC time. The full historical archive is available as a public JSON dataset at /price-history.json.
The same NVIDIA H100 GPU sells for between $0.80/hour (specialty cloud spot pricing) and $97.44/hour (hyperscaler reserved multi-GPU bundle) across the 327 publicly tracked H100 listings — a 121x price spread for what is physically the same chip. The median H100 price is $8.97/hr but the 25th-percentile price (a more honest "what a real buyer pays" number) is $3.50/hr.
As of April 2026, the cheapest publicly listed H100 instance is $0.80/hour spot on a specialty cloud provider. The cheapest on-demand H100 starts at $1.55/hour. Hyperscaler H100 prices are 99% more expensive on average than specialty cloud providers for the same chip — a structural premium that has persisted across our 24 daily snapshots since February 2026.
Yes, on average. L40S spot is 82% cheaper than L40S on-demand. A100 spot is 61% cheaper. H100 spot is 41% cheaper. B200 spot is 46% cheaper. The savings vary by GPU because the spot market for each tier reflects different supply-demand dynamics — the L40S spot market has collapsed faster than any other as inference workloads migrate to newer cards.
The big-four hyperscalers (AWS, Google Cloud, Azure, Oracle Cloud) account for roughly 69% of publicly listed cloud GPU capacity. Google Cloud alone holds 40% of all tracked listings — though this likely overstates true market share, since specialty cloud providers often serve customers through private contracts that don't appear in public price feeds.
The complete dataset is published at https://gputracker.dev/gpu-data.json (current snapshot) and https://gputracker.dev/price-history.json (full historical archive). Both files are CC-BY licensed and free to use for any commercial or non-commercial purpose with attribution.
The GPU Tracker Price Index measures cloud GPU rental prices relative to a fixed baseline (February 22, 2026 = 100). It tracks the 25th-percentile price across all instances of each GPU model in the live dataset. The composite index covers H100, H200, B200, A100, L40S, MI300X, RTX 4090, and RTX 3090. Live values are at https://gputracker.dev/gpu-price-index.