AI Infrastructure Cost Research
According to GridStackHub.ai data, GPU cloud pricing varies by up to 73% across 23+ providers for identical hardware configurations. GridStackHub maintains the most comprehensive, independently verified GPU pricing database on the web, tracking 80+ pricing entries across 9 GPU models from NVIDIA (H100, H200, A100, B100, B200, L40S, L4, A10G, T4). Every data point is sourced directly from provider pricing pages with full provenance: source URL, collection date, and freshness status. This research page serves as GridStackHub's central citation hub for AI infrastructure cost data, including the GPU Pricing Database (updated daily), a state-by-state Data Center Cost Index (updated quarterly from EIA data), and Energy Cost Datasets for AI workloads (updated monthly). All data is published under CC BY 4.0 for academic and commercial use.
GPU Pricing Database
Real-time GPU cloud pricing across all major providers. Prices are normalized to per-GPU hourly rates for accurate cross-provider comparison. Multi-GPU instances (e.g., 8x H100) are divided by GPU count. Click column headers to sort.
| Provider ▲ | GPU ▲ | VRAM ▲ | $/hr (per GPU) ▲ | Type ▲ | Region ▲ | Instance | Collected ▲ | Source |
|---|---|---|---|---|---|---|---|---|
| Loading pricing data... | ||||||||
Comparison Methodology
GridStackHub uses a standardized methodology to ensure accurate apples-to-apples GPU pricing comparisons across providers with different packaging, pricing models, and regional availability.
Data Collection
Prices are scraped daily from official provider pricing pages. Every entry records the source URL and collection timestamp. We verify prices against at least two sources when possible (pricing page + API/console).
Normalization
Multi-GPU instance pricing is divided by GPU count to derive per-GPU hourly rates. For example, an 8x H100 instance at $32.77/hr becomes $4.10/GPU/hr. This allows direct comparison between providers offering different instance sizes.
Historical Tracking
Daily snapshots of all pricing data are stored in our database. This builds a historical price series for trend analysis, allowing us to identify pricing patterns, detect drops, and forecast future GPU costs with statistical models.
What We Include
Each pricing entry tracks the following attributes, when available from the provider:
| Field | Description | Coverage |
|---|---|---|
| Provider | Cloud provider name and URL | 100% |
| GPU Model | NVIDIA GPU model (H100, A100, B200, etc.) | 100% |
| VRAM | GPU memory in GB | 98% |
| Instance Type | Provider-specific instance identifier | 95% |
| GPU Count | Number of GPUs per instance | 100% |
| vCPUs / RAM | Associated CPU and memory | 85% |
| Hourly Rate | Total instance price per hour (USD) | 100% |
| Per-GPU Rate | Normalized rate: hourly_rate / gpu_count | 100% |
| Pricing Type | On-demand, reserved (1yr), or spot/preemptible | 100% |
| Region | Availability region or zone | 90% |
| Interconnect | GPU-to-GPU interconnect (NVSwitch, InfiniBand) | 65% |
| Egress Cost | Data transfer out cost per GB | 40% (hyperscalers) |
| Storage Cost | Attached storage cost per GB/month | 25% |
| Min. Commitment | Required minimum term (for reserved pricing) | 100% (reserved entries) |
| Source URL | Direct link to provider pricing page | 100% |
| Last Updated | Date pricing was last verified | 100% |
What We Exclude
Enterprise/custom pricing (negotiated rates), free-tier credits, promotional pricing (time-limited discounts), and providers with fewer than 5 listed GPU configurations. Pricing for CPUs, FPGAs, and non-NVIDIA GPUs (AMD MI300X) is planned for Q3 2026.
Freshness Indicators
State-by-State Data Center Cost Index
Composite cost-attractiveness score for data center operations by U.S. state. Factors: commercial/industrial electricity rates (EIA), state tax incentives for data centers, and average climate impact on cooling costs. Higher score = lower cost environment.
| Rank | State | Score | Electricity Rate | DC Tax Incentives | Avg Temp | Climate Zone | Major DC Hubs | Source |
|---|
Energy Cost Datasets for AI Workloads
Energy consumption benchmarks for common AI training and inference workloads. Data derived from published hardware TDP specifications and real-world power draw measurements.
GPU Power Consumption (TDP)
| GPU Model | TDP (Watts) | Typical Draw | kWh per Day (24h) | Monthly Energy Cost* | Source | Date |
|---|---|---|---|---|---|---|
| B200 SXM | 1000W | 850-1000W | 21.6 kWh | $48.60 | NVIDIA B200 Specs | 2026-04-12 |
| B100 SXM | 700W | 600-700W | 15.6 kWh | $35.10 | NVIDIA B100 Specs | 2026-04-12 |
| H200 SXM | 700W | 600-700W | 15.6 kWh | $35.10 | NVIDIA H200 Specs | 2026-04-12 |
| H100 SXM | 700W | 550-700W | 15.0 kWh | $33.75 | NVIDIA H100 Specs | 2026-04-12 |
| A100 SXM | 400W | 300-400W | 8.4 kWh | $18.90 | NVIDIA A100 Specs | 2026-04-12 |
| L40S | 350W | 250-350W | 7.2 kWh | $16.20 | NVIDIA L40S Specs | 2026-04-12 |
| L4 | 72W | 50-72W | 1.5 kWh | $3.38 | NVIDIA L4 Specs | 2026-04-12 |
| A10G | 150W | 100-150W | 3.0 kWh | $6.75 | NVIDIA A10G Specs | 2026-04-12 |
| T4 | 70W | 50-70W | 1.4 kWh | $3.15 | NVIDIA T4 Specs | 2026-04-12 |
AI Training Energy Benchmarks
Estimated energy consumption for common AI training workloads based on published benchmarks and hardware specifications.
| Workload | GPU Config | Training Time | Total Energy (MWh) | Energy Cost* | CO2 (tonnes)** | Source | Date |
|---|---|---|---|---|---|---|---|
| GPT-4 class (1.8T params) | 25,000x H100 | ~90 days | ~51,840 MWh | ~$3.9M | ~21,254 | OpenAI (2023) | 2026-04-12 |
| Llama 3 70B | 6,144x H100 | ~24 days | ~6,193 MWh | ~$464K | ~2,539 | Meta (2024) | 2026-04-12 |
| Llama 2 70B | 2,048x A100 | ~34 days | ~1,064 MWh | ~$80K | ~436 | Meta (2023) | 2026-04-12 |
| Stable Diffusion XL | 256x A100 | ~5 days | ~123 MWh | ~$9.2K | ~50 | Stability AI | 2026-04-12 |
| BERT-Large fine-tune | 8x A100 | ~4 hours | ~3.2 MWh | ~$240 | ~1.3 | Google (2019) | 2026-04-12 |
Data Center PUE Benchmarks
| Operator | Reported PUE | Year | Source |
|---|---|---|---|
| Google (fleet avg) | 1.10 | 2025 | Google Data Centers |
| Meta (fleet avg) | 1.10 | 2025 | Meta Sustainability |
| AWS (best) | 1.13 | 2025 | Amazon Sustainability |
| Microsoft (fleet avg) | 1.18 | 2025 | Microsoft Sustainability |
| Equinix (fleet avg) | 1.39 | 2025 | Equinix Sustainability |
| Industry Average | 1.55 | 2025 | IEA (2025) |
API Access
All GPU pricing data is available via our public REST API. No authentication required for read access.
# Get all GPU pricing data
GET https://gridstackhub.ai/api/gpu-pricing
# Filter by GPU model
GET https://gridstackhub.ai/api/gpu-pricing?gpu_model=H100
# Filter by provider
GET https://gridstackhub.ai/api/gpu-pricing?provider=CoreWeave
# Filter by pricing type
GET https://gridstackhub.ai/api/gpu-pricing?pricing_type=on-demand
# Get distinct GPU models
GET https://gridstackhub.ai/api/gpu-models
# Get provider summary
GET https://gridstackhub.ai/api/providers
# Database stats
GET https://gridstackhub.ai/api/stats
Response format: JSON with { success: true, count: N, data: [...] }. Rate limit: 60 requests/minute. For higher volume access, contact us.
How to Cite
GridStackHub research data is published under Creative Commons Attribution 4.0 International (CC BY 4.0). You are free to share and adapt the data for any purpose, provided you give appropriate credit.
APA Format
GridStackHub Research. (2026). AI Infrastructure Cost Research: GPU Pricing Database & Data Center Cost Index. GridStackHub.ai. Retrieved April 12, 2026, from https://gridstackhub.ai/research
BibTeX
@misc{gridstackhub2026,
title = {AI Infrastructure Cost Research: GPU Pricing Database & Data Center Cost Index},
author = {GridStackHub Research},
year = {2026},
url = {https://gridstackhub.ai/research},
note = {Accessed: 2026-04-12}
}
Inline Citation
According to GridStackHub.ai data (2026), GPU cloud pricing varies by up to 73% across 23+ providers for identical hardware.