GPU Pricing Intelligence
What are current GPU lease rates for AI training?
H100 80GB SXM averages $2.15/hr ($1.49-$3.50 range), A100 80GB at $1.45/hr in a maturing market, and AMD MI300X at $2.10/hr as the leading alternative. H200 is at limited availability around $4.50/hr. B200 (Blackwell) is expected late 2026 at $10-15/hr with 2x training performance vs H100. GPU residual values are declining 30-45% annually for previous-generation hardware.
Last updated: April 2026
GPU Models & Pricing
H100 80GB SXM
NVIDIA Hopper architecture. The current standard for AI training.
A100 80GB SXM
NVIDIA Ampere architecture. Widely available, declining premium.
H200 141GB
Hopper refresh with HBM3e. Rolling out to hyperscalers.
B200 192GB
Blackwell architecture. 2x training performance vs H100.
MI300X 192GB
AMD CDNA 3. Competitive on memory, growing ROCm support.
MI250X 128GB
AMD CDNA 2. Budget alternative for HPC and inference.
GPU Residual Values
Depreciation curves and collateral valuation
Site Selection & Infrastructure
GPU deployment requires careful site selection. Explore infrastructure data to find the optimal location.
Frequently Asked Questions
What is the cheapest GPU for AI training in 2026?
The A100 80GB SXM offers the lowest cost for production training at $1.45/hr average, down from $2.50/hr in early 2025 as supply matured. AMD MI250X is even cheaper at $1.05/hr but requires ROCm compatibility. For inference workloads, spot A100 instances can drop below $1/hr.
How fast do GPUs depreciate for loan collateral?
H100 residual values are declining 30-45% annually as H200 and B200 enter the market. A100s have already lost 60-70% of original value. Private credit underwriters typically use 24-month residual curves for DSCR modeling. Our GPU Residual Value Calculator provides model-specific depreciation estimates.
Should I lease or buy GPUs for AI workloads?
Lease for training bursts under 18 months or when the next-gen transition (H100 → B200) makes ownership risky. Buy when utilization exceeds 70% over 24+ months and you can absorb residual risk. Sale-leaseback structures let operators extract capital from owned hardware while retaining use.
What is the H100 vs H200 vs B200 performance difference?
H200 offers 1.4-1.9x inference throughput over H100 via HBM3e memory (141GB vs 80GB) at similar training performance. B200 (Blackwell) targets 2x training and 5x inference performance vs H100 with 192GB HBM3e. Price-performance ratios shift at scale — compare using our GPU comparison tools.