Methodology &
Data Sources
Enterprise GPU decisions are high-stakes. This page documents exactly where every number on GPU Advisor comes from, how often it is updated, and where our data has known limitations.
Weekly
Pricing Updates
100%
Public Sources
0
Vendor-Paid Rankings
Jun 2026
Last Full Audit
Editorial Standards
How We Track Prices
- —Prices are collected directly from public provider pricing pages. We do not accept price data submitted by providers.
- —All rates are standardized to per-GPU-hour for comparison. Per-node prices are divided by the GPU count in that instance type.
- —Freshness is shown as "regularly tracked · last tracked [date]" on each pricing table. Rows where public pricing is unavailable are marked as estimated.
We Do Not Rate Providers
- —GPUAdvisor does not score, rank or rate providers. Tables show tracked prices, minimum commitments and published compliance status, with the date each was last tracked.
- —Pricing is provider-reported and regularly tracked. Where a row has no recorded verification date it is shown as unverified.
- —Compliance entries come from provider documentation and third-party certifications, and are shown as stated by the provider or certifier.
Commercial Disclosure
- —All listings are free and editorial. No provider can pay for placement, or inclusion in any table.
- —Paid options are limited to clearly disclosed sponsor slots and per-signup referral links. These never affect data, ordering, or editorial analysis.
How We Calculate Each Metric
Performance Benchmarks
- —All FP16/BF16 throughput figures use the decode phase of autoregressive LLM generation (output tokens per second), not prefill. This is the metric that dominates in production inference workloads.
- —NVIDIA and AMD figures are sourced from MLPerf Inference or reproducible community vLLM benchmarks at batch size 1 (latency-bound) and batch size 16–32 (throughput-bound). We display throughput-bound numbers.
- —Google TPU figures use JAX/JetStream benchmarks and are NOT directly comparable to NVIDIA/AMD CUDA/ROCm numbers due to different framework overhead, memory allocation strategies, and programming models. This is clearly noted wherever TPU data appears.
- —FP8 TFLOPS are peak theoretical and assume full hardware utilization. Real-world FP8 inference depends on model support and kernel availability.
Cloud Pricing Collection
- —All prices are public list prices in USD for the us-east-1 / us-central1 / East US regions unless otherwise noted. Prices in other regions vary by 5–25%.
- —On-demand pricing is verified manually each week against the official pricing page of each provider. We do not accept pricing data from providers directly.
- —Spot pricing is inherently volatile and represents a recent average, not a guarantee. Actual spot prices can be 30–70% of the listed on-demand rate.
- —CoreWeave and Lambda Labs prices are per-GPU. AWS, GCP, Azure prices are per-instance (8-GPU node) divided by 8 for per-GPU comparisons.
- —We do not receive commission or placement fees that influence pricing data. Provider links in the pricing table (labeled 'Get Started', 'Rent Now', or 'Get Access') may be referral links. These commercial relationships do not influence pricing data or rankings.
TCO & ROI Calculations
- —Training cost estimates use the Chinchilla scaling law (20× parameters = optimal token count) and 35% Model FLOPs Utilization (MFU), which is typical for production clusters without extreme optimization.
- —Inference cost per token assumes 70% GPU utilization, single-GPU throughput at batch size 16, and 730 hours/month (full calendar month). Real costs vary by batch size, model quantization, and actual utilization.
- —ROI comparisons assume equivalent GPU generations (e.g., H100 vs H100) across providers. Cross-generation comparisons (H100 vs B200) account for throughput differences using benchmark-derived correction factors.
- —On-premise TCO excludes power and cooling costs by default. Enable the 'Include OpEx' toggle in the TCO calculator to add estimated $0.10/kWh and 1.4 PUE.
What We Don't Cover
- —Spot pricing SLA — spot instances can be preempted without notice. We show historical averages only.
- —Private / negotiated enterprise pricing — hyperscaler contract pricing for large deployments can be 20–50% below list price. Our numbers are list price.
- —Multi-region latency, compliance, and data residency requirements — these are workload-specific and outside our scope.
- —Fine-tuned model performance — all benchmarks use standard pretrained model weights. Fine-tuned or quantized models will show different throughput.
Update Schedule
Cloud GPU on-demand pricing
Weekly (every Monday)
Jun 2, 2026
Manual spot-check of AWS, GCP, Azure, Lambda, CoreWeave, RunPod pricing pages + automated diff alerts. Rows marked † are estimated from comparable instance types where public pricing is unavailable.
Spot / preemptible pricing
Weekly
Jun 2, 2026
Provider APIs where available; manual check otherwise. Spot prices reflect 7-day average, not real-time
Reserved / committed pricing
Monthly
Jun 1, 2026
Major provider pricing pages; 1yr and 3yr reserved rates rarely change mid-month
GPU performance benchmarks
On new MLPerf release
Jun 2026 (MLPerf v4.1)
MLCommons publishes ~2× per year; incorporated within 2 weeks of release. vLLM community benchmarks reviewed monthly
Hardware specifications
On product announcement
Jun 2026 (B300 Ultra)
Updated within 24–48 hrs of official manufacturer announcement. Cross-checked against Hot Chips and IEEE publications
TCO model assumptions
Quarterly
Jun 2026
Power costs ($0.07–$0.12/kWh), rack/colocation rates, PUE (1.2–1.5), and staffing ratios reviewed each quarter
vLLM throughput figures
Monthly
Jun 2026
FP16/FP8 tok/s benchmarked from community reproducible vLLM runs; single-chip, decode phase, batch size 16
Primary Data Sources
GPU Performance Benchmarks
Primary throughput source — H100, A100, L40S, MI300X offline/server scenarios
Training throughput for BERT large, ResNet-50, GPT-3 175B
FP8 (989 TFLOPS H100), FP16, INT8 peak TFLOPS; NVLink 4.0 / NVLink 5.0 bandwidth
MI300X 192GB HBM3 5.2 TB/s; MI355X 288GB HBM3e specs
TPU v5e, v5p, v6e (Trillium), v7 (Ironwood) BF16 and INT8 figures — JAX/XLA only
Reproducible FP16/FP8 tok/s for Llama-2/3 7B, 13B, 70B — decode phase, batch 16
Cloud GPU Pricing
p5.48xlarge (8× H100 SXM5) $55.04/hr · p4d.24xlarge (8× A100 40GB) $32.77/hr — us-east-1
a3-highgpu-8g (8× H100 80GB) $32.77/hr · TPU v5p/v6e per-chip rates — us-central1
ND96isr H100 v5 (8× H100 SXM5) ~$98/hr · ND96asr v4 (8× A100 40GB) $27.20/hr — East US
H100 SXM5 $2.49/hr per GPU · A100 80GB $1.79/hr · A10 $0.75/hr — on-demand
H100 SXM5 $2.65/hr · H200 $5.20/hr · B200 $6.50/hr · MI300X $4.10/hr per GPU
H100 SXM5 Secure Cloud $2.49/hr · A100 80GB $1.64/hr · spot rates ~30–50% lower
Hardware Specifications
H100: 80GB HBM3 3.35 TB/s · H200: 141GB HBM3e 4.8 TB/s · B200: 192GB HBM3e 8.0 TB/s
MI300X: 192GB HBM3 5.2 TB/s · MI325X: 288GB HBM3e · MI355X: 288GB HBM3e CDNA4
TPU v6e (Trillium): 32GB HBM · TPU v7 (Ironwood): 192GB HBM — 9,216-chip pods
HL-325L: 128GB HBM2e · 3.7 TB/s bandwidth · 1,835 TFLOPS BF16
Die-level architecture for Blackwell GB200, CDNA4 MI355X, Gaudi 3
Industry Research
Supply chain analysis, H100/B200 pricing trends, hyperscaler CapEx models
Training compute scaling laws, hardware efficiency trends, model FLOP estimates
Real-world LLM API latency, throughput, and quality scores across providers
vLLM vs TensorRT-LLM throughput, quantization impact on inference cost
Commercial Disclosure
GPU Advisor earns referral commissions through affiliate links on "Deploy" buttons and provider sign-up links. These commercial relationships do NOT influence our benchmark rankings, pricing data, or editorial analysis. Affiliate links are clearly marked. If you believe any data has been influenced by commercial interests, please contact us.